Mobile Networks and Applications 10, 78, 2005

2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Guest Editorial
Algorithmic Solutions for Wireless, Mobile, Ad Hoc and Sensor Networks
AMOTZ BAR-NOY, ALAN A. BERTOSSI, CRISTINA M. PINOTTI and CAULIGI S. RAGHAVENDRA

The field of wireless and mobile computing is an important and research challenging area of computing today. This has been
made possible due to the tremendous and continued growth of wireless technology, creating the need of ubiquitous distributed
services: anywhere and anytime. In addition to wireless networks based on a pre-existing infrastructure, where wireless
communications take place only between the end-nodes and the access points, mobile ad hoc wireless networks and sensor
networks are emerging rapidly. Such networks do not need any insfrastructure to work, but they comprise mobile clients
as well as mobile servers, and they pose challenges in diverse areas such as network topology control, routing and security,
resource placement, allocation and discovery, energy consumption, and media access. In this scenario, there is a great need
of algorithmic solutions to realize and maintain high-speed, high-performance, cost-effective, energy-efficient and reliable
wireless networks.
This special issue brings together contributions in discrete algorithms, optimization techniques, and performance evaluation
methods in the context of wireless, ad-hoc, and sensor networks.
More than 60 papers were submitted, out of which only the 16 papers below have been selected for publication. These
papers do not cover all aspects that belong to the scope of this special issue. However, they represent interesting research
efforts and subjects that definitely belong to the core research on algorithmic solutions for wireless, mobile, ad hoc, and sensor
networks.
The first three papers deal with the network topology control problem.
The paper Dynamic Coverage in Ad-Hoc Sensor Networks, by H. Huang, A.W. Richa and M. Segal, dynamically main-
tains measures on the quality of the coverage of a sensor network.
The paper Algorithmic Aspects of Topology Control Problems for Ad Hoc Networks, by E.L. Lloyd, R. Liu,
M.V. Marathe, R. Ramanathan and S.S. Ravi, shows how to assign power values in ad hoc networks to obtain a graph topology
satisfying some specified important properties.
The paper Wireless ATM Layouts for Chain Networks, by M. Flammini, G. Gambosi and A. Navarra, integrates the
benefits of the ATM technology with the wireless communication, and studies the existence of optimal layouts for special
network topologies.
The next two papers consider the problem of routing in ad hoc networks.
The paper Ad Hoc Multicast Routing Algorithm with Swarm Intelligence, by C.-C. Shen and C. Jaikaeo, proposes a novel
idea for multicast routing based on Swarm Intelligence that refers to complex behaviors that arise from very simple individual
behaviors and interactions.
The paper Regional Gossip Routing for Wireless Ad Hoc Networks, by X.-Y. Li, K. Moaveninejad and O. Frieder,
develops a location based routing protocol and presents a detailed analysis of this routing protocol.
The next three papers study the issues of placement, allocation and discovery of resources in cellular and ad hoc networks.
The paper Comparison and Evaluation of Multiple Objective Genetic Algorithms for the Antenna Placement Problem, by
L. Raisanen and R.M. Whitaker, evaluates the performance of a greedy algorithm to select and configure base station locations
using genetic algorithms methods.
The paper A Characterisation of Optimal Channel Assignments for Cellular and Square Grids Wireless Networks, by
M.V.S. Shashanka, A. Pati and A.M. Shende, proposes optimal channel assignment algorithms in wireless networks whose
topology can be represented by square and cellular grids.
The paper CARD: A Contact-Based Architecture for Resource Discovery in Ad Hoc Networks, by A. Helmy, S. Garg,
P. Pamu and N. Nahata, proposes a resource discovery mechanism based on distributed directories which is suitable for large
ad hoc networks.
The next group of papers involves the energy consumption problem in sensor and ad hoc networks.
The paper Energy-Balanced Task Allocation for Collaborative Processing in Wireless Sensor Networks by Y. Yu and
V.K. Prasanna, considers the problem of scheduling a real-time application onto a single-hop wireless sensor network taking
into account energy requirements for both computation and communication.
8 GUEST EDITORIAL

The paper Efficient and Robust Protocols for Local Detection and Propagation in Smart Dust Networks, by I. Chatzigian-
nakis, S. Nikoletseas and P. Spirakis, presents various protocols for smart dust based sensor networks for local event detection
and propagation of reports.
The paper Training a Wireless Sensor Network, by A. Wadaa, S. Olariu, L. Wilson, M. Eltoweissy and K. Jones, proposes
a protocol for training nodes in a sensor network. The protocol, partitioning nodes into clusters, obtains a scalable and energy-
efficient routing from cluster to the sink.
The paper Quorum-Based Asynchronous Power-Saving Protocols for IEEE 802.11 Ad Hoc Networks, by J.-R. Jiang,
Y.-C. Tseng, C.-S. Hsu and T.-H. Lai, addresses the asynchronous power management problem for an IEEE 802.11-based
Multi-Hop MANET, correlating it to the concept of quorum system.
The last group of four papers deals with the media access and transmission scheduling problems.
The paper CROMA An Enhanced Slotted MAC Protocol for MANETs, by M. Coupechoux, B. Baynat, C. Bonnet and
V. Kumar, presents a TDMA based MAC protocol with high utilization in synchronized mobile ad hoc networks. CROMA
handles both the hidden terminal and exposed terminal problems to achieve a high throughput.
The paper Dynamic Bandwidth Management in Single-Hop Ad Hoc Wireless Networks, by S.H. Shah, K. Chen and
K. Nahrstedt, presents dynamic bandwidth management and call admission control in a single hop ad hoc network at the
application level. This paper shows an elegant solution to this problem with simulation and experimental results.
The paper High Speed Networking Security: Design and Implementation of Two New DDP-Based Ciphers, by
N. Sklavos, N.A. Moldovyan and O. Koufopavlou, proposes two new fast ciphers suitable for wireless communications,
which set hard specifications in security implementations.
Finally, the paper Media Synchronization and Qos Packet Scheduling Algorithms for Wireless Systems, by A. Boukerche
and H. Owens II, considers the QoS requirements and the scheduling transmission problems arising when multiple streams of
text, images, audio and video are sent to mobile clients through a combined wired and wireless network.
The guest editors wish to thank all the referees for their valuable comments and suggestions, and all the authors for their
high quality submissions. Special thanks go to the Editor-in-Chief of MONET for hosting this special issue.
 Mobile Networks and Applications 10, 917, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Dynamic Coverage in Ad-Hoc Sensor Networks

HAI HUANG and ANDRA W. RICHA ,
Department of Computer Science and Engineering, Arizona State University, Tempe, AZ 85287-8809, USA

MICHAEL SEGAL
Communication Systems Engineering Department, Ben-Gurion University of the Negev, Beer-Sheva 84105, Israel

Abstract. Ad-hoc networks of sensor nodes are in general semi-permanently deployed. However, the topology of such networks con-
tinuously changes over time, due to the power of some sensors wearing out, to new sensors being inserted into the network, or even due
to designers moving sensors around during a network re-design phase (for example, in response to a change in the requirements of the
network). In this paper, we address the problem of how to dynamically maintain two important measures on the quality of the coverage
of a sensor network: the best-case coverage and worst-case coverage distances. We assume that the ratio between upper and lower trans-
mission power of sensors is bounded by a polynomial of n, where n is the number of sensors, and that the motion of mobile sensors can
described as a low-degree polynomial function of time. We maintain a (1 + )-approximation on the best-case coverage distance and
be
a ( 2 + )-approximation on the worst-case coverage distance of the network, for any fixed > 0. Our algorithms have amortized or
worst-case poly-logarithmic update costs. We are able to efficiently maintain the connectivity of the regions on the plane with respect to the
sensor network, by extending the concatenable queue data structure to also serve as a priority queue. In addition, we present an algorithm
that finds the shortest maximum support path in time O(n log n).
Keywords: coverage, ad hoc sensor network, kinetic data structure

1. Introduction
Ad-hoc sensor networks are emerging as a new sensing par-
adigm and have thus received massive research interest re-
cently. Usually sensor nodes are semi-permanently deployed,
since the sensors themselves barely have any moving capac-
ity. However, the topology of such networks continuously
changes over time due to a variety of reasons: For example,
a sensor node may wear out due to its very limited battery
power; a new sensor node may be inserted into the network;
or the layout of a sensor network may need to be changed
in order to improve the quality of the network coverage in
response to a change in the network requirements, which is
accomplished by changing the placement of current (or in-
serting, deleting) sensors in network.
In this paper, we address the problem of how to dynami-
cally maintain two important measures on the quality of the
coverage of a sensor network: the best-case coverage distance
and the worst-case coverage distance of the network. We also
address a closely related problem, namely that of finding a
shortest maximum support path.
In a sensor network, each sensor bears the ability to detect
objects around it. The coverage of a sensor is limited by its
energy level. Assuming that a sensors detecting ability is
omnidirectional, we can model the coverage of a sensor as a
disk (under 2-norm on the Euclidean plane1 ) centered at the
sensor. The radii of such disks are determined by the energy
This work was supported in part by NSF CAREER Award CCR-9985284.
Corresponding author.
1 A disk of radius r centered at (x, y) under 2-norm in R2 is the set of points


(p, q) such that (p x)2 + (q y)2
r.
level of the sensors. The coverage area (or simply coverage)
of the sensor network is the union of all such disks.
A sensor network is often used to detect intruders. An in-
truder may start at a point S, follow an arbitrary trajectory
(path) on the plane, and stop at some other point T on the
plane. In some applications, a sensor network may need to
keep track of the intruder at all times, as the intruder follows
its trajectory; in some other applications, the networks func-
tion may be simply to detect the presence of an intruder, in
which case the network only needs to cover some part of the
trajectory. Thus, given two points S and T , two relevant types
of trajectories on the plane are proposed [10]: the maximum
breach path and the maximum support path. (In [10], these
paths are called maximal breach path and maximal support
path, respectively.)
The maximum breach path measures the vulnerability of a
sensor network by, as the name suggests, completely avoid-
ing the coverage area of the sensor network: It is a trajectory
between the start point S and the stop point T that stays as
far away from the sensors as possible. On the other hand,
the maximum support path measures the efficiency of the net-
work coverage: This path is a trajectory between S and T
which stays as close to the sensors as possible. The distance
of a point P to the sensor network is defined as the smallest
Euclidean distance from P to one of the sensor nodes. A max-
imum breach path from S to T is a path from S to T such that
the minimum distance from a point P in the path to the sen-
sor network is maximized: this distance is called the worst-
case coverage distance of the network. Similarly, a maximum
support path from S to T is a path such that the maximum
distance of a point P in the path to the sensor network is min-
10
imized: this distance is called the best-case coverage distance
of the network.
When the topology of a sensor network changes, the qual-
ity of its coverage most probably will be affected. We would
like to maintain an assessment on the quality of the network
coverage which, as explained above, can be done by main-
taining the worst-case and best-case coverage distances ef-
ficiently at all times. This would give a clear indication on
how effective the network coverage is at any given point in
time, possibly calling for the insertion of new nodes in the
network (e.g., when the coverage deteriorates due to node
failures) or to a network re-design phase. Whenever neces-
sary, the actual paths which give the best-case and worst-case
coverage distances can be retrieved. As we will see later, in
sections 4 and 5, our algorithms for maintaining the worst-
case and best-case coverage distances have poly-logarithmic
update and query costs, as defined later. To the best of our
knowledge, this is the first work which formalizes and ad-
dresses this problem in a dynamic scenario.
For a moment, let us assume that all sensors have the same
energy power and thus all disks have the same radius r. We
call such a sensor network a uniform sensor network with cov-
erage radius r. In a uniform sensor network, all of the paths
whose minimum distance of a point in the path to a sensor
is larger than the coverage radius are equivalent, in the sense
that the sensors in the network will not be able to detect an in-
truder using any such path. Similarly, all of the paths whose
maximum distance of a point in the path to a sensor is smaller
than the coverage radius are equivalent, in the sense that any
such path is entirely contained in the coverage area of the net-
work. The worst coverage radius (see [10]) is defined to be
the maximum coverage radius r such that there exists a trajec-
tory P between given points S and T which does not intersect
the interior region of the area covered by the uniform sensor
network (i.e., P may touch the coverage area, intersecting
it at a discrete number of points only). We can think of the
worst-coverage radius as being the maximum energy that can
be assigned to the sensor nodes which still would not prevent
an intruder from escaping from S to T without being detected
(for simplicity, we assume that a sensor will not be able to
detect an intruder who only touches its coverage area). Cor-
respondingly, the best coverage radius (see [10]) is defined
to be the minimum coverage radius r such that there exists
a trajectory between S and T that is totally covered by the
uniform sensor network.
We introduce uniform sensor networks as a merely concep-
tual tool in order to facilitate the presentation of our approxi-
mation algorithms and their analyses, following a similar ap-
proach as Li et al. [9]. (The actual sensor network in consid-
eration has nodes with arbitrary energy levels and therefore
is not assumed to be uniform.) In fact, if we think of a uni-
form sensor network built on top of the placement of the sen-
sor nodes currently deployed in the general sensor network in
consideration, the worst-coverage radius of the uniform net-
work is indeed equal to the worst-case coverage distance of
the general sensor network, and the best-coverage radius is
indeed equal to the best-case coverage distance.
HUANG ET AL.
In order to dynamically maintain the best- and worst-case
coverage distance efficiently, we need to maintain some infor-
mation on the current topology of the sensor network; when
the network topology changes, we need to update this infor-
mation. We also perform queries for the current best-case
and worst-case coverage distances, based on the information
maintained. Hence, the cost (or running time) of our algo-
rithms are measured in terms of their respective update cost
i.e., the cost to update the topology information, which is
charged per relevant topology change in the network and
the query cost, which is the cost incurred when answering
a query for the current best-case or worst-case coverage dis-
tance.
In sections 4 and 5, we formally define a relevant topol-
ogy change which will henceforth be called an event for
the problems of maintaining the best-case and worst-case cov-
erage distances, respectively.
The remainder of the paper is organized as follows. Sec-
tion 1.1 states our results. In section 2, we present some re-
lated work in the literature. Section 3 covers some prelim-
inaries and sketches the basic framework of our solutions.
We present the low constant approximation algorithms for the
best- and worst-case coverage distance in sections 4 and 5 re-
spectively. In section 6 we address the closely related problem
of efficiently finding a shortest maximum support path. Sec-
tion 7 concludes the paper with some possible lines for future
work.
1.1. Our results
In this section, we summarize the main results of this paper.
One of the main contributions of this work is to take into ac-
count the dynamic nature of sensor networks, and to propose
a framework which can be used to continuously monitor the
quality of the network coverage. Let n denote the current
number of sensors in the network.
In the following sections, we present two algorithms to
maintain low constant approximations on the best-case and
worst-case coverage distances. Both algorithms have low
update and query costs. Namely, our algorithms achieve a
(1 + )-approximation
on the best-case coverage distance,
and a ( 2 + )-approximation on the worst-case cover-
age distance, for any fixed > 0. The amortized up-
date cost per event of the best-case coverage distance algo-
rithm is O(log3 n), and the respective query cost is worst-case
O(log n). For the worst-case coverage algorithm, the update
cost per event is worst-case O(log2 n) and the query cost is
worst-case O(1). A formal definition of an event for each of
the problems considered follows in sections 4 and 5, respec-
tively.
As a byproduct of our algorithm for maintaining the worst-
case coverage distance, we extend the concatenable queue
data structure to also serve as a priority queue. All the opera-
tions on this extended data structure have worst-case O(log n)
running time.
We also present an O(n log n) algorithm for computing
an exact shortest maximum support path between two given
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS
points S and T , improving on the best-known previous results
by Li et al. [9]. A shortest maximum support path from S
to T is a maximum support path from S to T such that the
Euclidean length of the trajectory followed in this path is min-
imum. In [9], two algorithms are presented for computing the
maximum support path: One algorithm computes an exact
shortest maximum support path in O(n2 log n) time; the other
algorithm provides a 2.5-approximation on the shortest max-
imum support path in O(n log n) time. One should note that
the algorithms presented by Li et al. can be implemented in
a distributed fashion (we use the communication complexity
as the time bound for the sequential versions of their algo-
rithms), whereas the algorithms presented in this paper are all
centralized.
The update costs of our algorithms for approximately
maintaining the best- and worst-case coverage distances are
much cheaper than maintaining the best- or worst-case cov-
erage distances using the best-known algorithms in the liter-
ature prior to this work. In fact, the best previously known
algorithm for maintaining the best-case (resp., worst-case)
coverage distance maintains the exact distance by repeat-
edly re-computing the maximum support path (resp., max-
imum breach path) using the O(n log n) algorithm by Li
et al. [9] (resp., the O(n2 log n) algorithm by Meguerdichian
et al. [10]) each time an event occurs. To the best of our
knowledge, this is the first work that explicitly addresses
the problems of dynamically maintaining (approximations of)
these two distances.
2. Related work
Meguerdichian et al. [10] considered the problems of finding
the maximum breach path and the maximum support path on
a sensor network. They [10] present an O(n2 log ) runtime
algorithm for the maximum breach path problem, where n is
the number of sensors in the sensor network, and  is the dif-
ference between the highest and the lowest weight of an edge
in the Voronoi Diagram of the sensor network. Their algo-
rithm for computing the maximum support path has the same
running time as their maximum breach path algorithm. The
O(log ) factor can be easily converted into O(log n) in the
algorithm that solves the maximum breach path problem if we
perform a binary search over a sorted list of the radii of sen-
sors instead of using a linear search as in [10]. The algorithms
presented in [10] heavily rely on geometric structures such as
the Voronoi Diagram and Delaunay triangulation of the net-
work, which cannot be constructed efficiently in a distributed
manner.
Li et al. [9] prove the correctness of the algorithms given
in [10]. They also show how to find a maximum support
path in O(n log n) time using a centralized algorithm, or with
O(n log n) communication complexity bits in a distributed
fashion. In addition, Li et al. [9] present two algorithms for
computing a shortest (with respect to the Euclidean length of
the trajectory followed in this path) maximum support path:
an algorithm that computes an exact shortest maximum sup-
11
port path with O(n2 log n) worst-case communication com-
plexity, and an algorithm that computes a 2.5-approximation
of a shortest maximum support path (i.e. the total length of
the obtained path is at most 2.5 times the length of a short-
est maximum support path) with O(n log n) communication
complexity.
Meguerdichian et al. [11] proposed an exposure-based for-
mulation for analyzing the coverage of paths taken by polyg-
onal objects: they define a path-dependent integral, which
consists of the trajectories of all the points of a polygonal ob-
ject (the polygonal object is able to rotate), and not only of
the trajectory of the objects center point.
Recently, Zhang and Hou [13] proved that if the commu-
nication rage of a sensor is at least twice its sensing range,
a complete coverage of a convex area implies connectivity
among the working set of nodes and derive optimality con-
ditions under which a subset of working sensor nodes can be
chosen for full coverage. Wang et al. [12] designed a Cov-
erage Configuration Protocol (CCP) that can provide differ-
ent degrees of connected coverage and present a geometric
analysis of the relationship between coverage and connectiv-
ity. Huang and Tseng [8] present an algorithm with runtime
of O(n2 log n) that decides whether every point in a given ser-
vice area is covered by at least one sensor.
3. Preliminaries
Before heading into the technical details of our algorithms,
we introduce some basic concepts which will be used in both
sections 4 and 5. The first concept we introduce is that of
growing disks, which will help us translate our problems into
graph connectivity problems.
The growing disks concept was previously proposed in [9].
We restate it in terms of the coverage radius of a uniform sen-
sor network as defined in section 1. (In section 1, we saw how
the coverage radius of a virtual uniform overlay sensor net-
work directly relates to the worst-case and best-case coverage
distances of the actual network.) Assume we have a uniform
sensor network with coverage disks centered at the sensors.
Define U (r) to be the region on the plane composed of the
union of all of the coverage disks when the coverage radius
is r. Let U (r) be the complement of the region U (r). At the
very beginning, we set the coverage radius to be equal to 0.
Then U (r) is the union of discrete singletons. As the cov-
erage radius grows, the disks centered at the sensors become
larger and might get connected into larger regions. Therefore,
U (r) might get disconnected into separate regions. For any
two given points S and T , the best coverage radius is the min-
imum r such that S and T are in the same connected region of
U (r), while the worst coverage radius is the minimum r such
that S and T belong to two disconnected regions in U (r).
Hence, the best and worst coverage radius problems translate
to connectivity problems on U (r) and U (r), respectively. Fig-
ure 1 illustrates these ideas. We will further translate the best
and worst coverage radius problems into graph connectivity
problems.
12
(a)
(b)
Figure 1. Best and worst coverage radii. (a) Best-coverage radius: mini-
mum r such that S and T are connected in U (r). (b) Worst-coverage radius:
minimum r such that S and T are disconnected in U (r).
We first show how to translate the best coverage radius
problem into a graph connectivity problem. A uniform disk
graph is the intersection graph of disks with uniform radius r
(see [4]). In this graph, disks are vertices and there is an edge
between two vertices if and only if the corresponding disks
intersect.2 The connectivity of U (r) is naturally modeled by
that of a uniform disk graph of radius r, denoted by G(U (r)).
The best coverage radius is the minimum r such that the ver-
tex corresponding to the disk containing S is connected to that
corresponding to the disk containing T in G(U (r)).
We also translate the worst coverage radius problem into a
graph connectivity problem. However this case is rather more
involved and we delay its presentation to section 5.
When r is fixed, suppose that we have a poly-logarithmic
running time query to check whether the region in either U (r)
or U (r) containing S is connected to that containing T . Then
we can build an -approximation algorithm, > 1, for either
the best or the worst coverage radius problem, as we show in
the next paragraph.
For the best coverage radius, consider the sequence of
U (ri ), such that ri = ri1 . Let i be such that S and T are
connected in U (ri ) but not in U (ri1 ). Since the best cover-
age radius falls in the interval [ri1 , ri ] and since ri is at most
ri1 , we know that ri is an -approximation on the best cov-
erage radius. A similar argument on the sequence of U (ri )s
gives an -approximation of the worst coverage radius.
Assume sensors occupy some space and cannot overlap.
Then there is a constant lower bound on the coverage radius,
denoted by rmin . Due to the limited battery power, we assume
that there is a constant upper bound on the coverage radius,
2 If we rescale one unit to be 2r, then a uniform disk graph is a unit-disk
graph.
HUANG ET AL.
denoted by rmax . Let R = rmax /rmin . We need to main-
tain log (R) copies of U (ri ) or U (ri ). If updating the rele-
vant connectivity information for each U (ri ) or U (ri ) takes
time g(n), the overall update time is log (R) g(n). The up-
date time is poly-logarithmic on n provided that g(n) is poly-
logarithmic on n, and that R is bounded by a polynomial on n.
4. Dynamic best-case coverage distance
In this section, we present our (1 + )-approximation algo-
rithm to maintain the best-case coverage distance following
the framework presented in section 3. Recall that, as shown
in section 3, finding the best-case coverage distance for given
points S and T is equivalent to finding the minimum r such
that S and T are connected in G(U (r)). Thus our main goal
is to devise an approach to maintain the connectivity of the
uniform disk graph G(U (r)) such that both the update cost
and the query cost are poly-logarithmic on n, where n is the
number of sensors in the network.
Holm et al. [7] showed that the connectivity of a graph
can be maintained in amortized poly-logarithmic update cost,
whereas each query takes worst-case O(log n/ log log n) time.
Guibas et al. [5] used Holm et al.s algorithm to maintain con-
nectivity on a unit-disk graph. The update cost in [5,7] is
charged per edge insertion or deletion. In order to be able
to detect when uniform disks meet or separate on the plane
(corresponding to an edge insertion or deletion on a unit-disk
graph, respectively), Guibas et al. [5] introduced a kinetic data
structure specially tailored to handle this scenario.
The kinetic data structure framework was first proposed by
Basch et al. [2,3] to deal with dynamics. Their main contri-
bution is a method to maintain an invariant of a set of moving
objects in a discrete manner. They introduce the idea of keep-
ing certificates as triggers for updates. When an object moves
and a certificate fails, the consistency of the kinetic data struc-
ture is invalidated and an update is mandatory. Each failure of
a certificate incurs a setup of up to a constant number of new
certificates. Hence we are allowed to monitor the dynamics
of a set of objects discretely and efficiently. The kinetic data
structure requires that we know the flight plan (a specification
of the future motion) [2,5] of all disks, and that the trajectory
of each disk can be described by some low-degree algebraic
curve. We have the freedom to change the flight plan of a disk
at any time. Basch [2] shows that kinetic data structures can
efficiently support the dynamic operations of inserting and
deleting objects into the system, provided those operations
do not occur too often. The details of kinetic data structures
are beyond the scope of this paper. Please refer to [2,3,5] for
more information.
The kinetic data structure utilized in [5] can be viewed as a
discrete event monitor. The events we need to monitor in or-
der to maintain accurate connectivity information on G(U (r))
are when two disks meet or separate. In [5], two types of
certificates are set up and the data structure allows us to de-
termine a priori the time when an event will occur. When an
event occurs, the topology of the uniform disk graph G(U (r))
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS
changes and an update on the connectivity information is trig-
gered. Hence the update cost is the cost to update the connec-
tivity information of G(U (r)) per event. When a certificate
fails and an event occurs, it takes constant time for the kinetic
data structure to process the failure (due to the setup of at most
a constant number of new certificates). We do not explicitly
take this cost into account when computing the update cost of
the maintenance of the connectivity information of G(U (r)),
since it would not change the asymptotic bound on the update
cost.
We adapt the main theorem in [5, theorem 5.4], to better
serve our purposes. The uniform disk graph G(U (r)) corre-
sponds to a unit-disk graph if we rescale one unit to be equal
to 2r.
Lemma 1 (Adapted from [5, theorem 5.4]). In [5], an algo-
rithm to dynamically maintain the connectivity of G(U (r)) is
presented. The update cost is amortized O(log2 n) per event.
The query cost is worst-case O(log n/ log log n).
We still need to show how to determine which disks con-
tain the given points S and T , at any given time. We sort all
sensors according to their distances to the fixed point S. We
maintain a binary heap on this ordering. Once the ordering
changes, we update the heap in O(log n) time. This intro-
duces a new type of event namely, when a sensor changes
its location and needs to be re-inserted in this ordering be-
sides the other two events defined earlier. The update cost
for this event is O(log n). To check which disk contains S,
we find the closest sensor p to S. We check if the distance
from p to S is larger than the coverage radius. If so, then S is
not contained in any disk. Otherwise, we know that the disk
centered at p contains the point S. This query takes constant
time. We maintain the ordering of the sensors with respect
to T in a similar way.
Combining the result in this section with the algorithmic
framework presented in section 3, we have our (1 + )-
approximation algorithm (for any > 0) for the best-
case coverage distance by maintaining log1+ R copies of
G(U (r)), for r = 1, (1+), (1+)2 , . . . . We perform a query
operation by doing a binary search on the log1+ R copies of
G(U (r)).
Theorem 1. Our algorithm dynamically maintains a (1 + )-
approximation, for any > 0, of the best-case coverage
distance. The update cost of this algorithm is amortized
O(log2 n log1+ R) per event and the query cost is worst-
case O((log n/ log log n) log log1+ R).
Corollary 1. If > 0 is fixed, then our algorithm has amor-
tized O(log3 n) update cost per event, and worst-case O(log n)
query cost.
5. Dynamic worst-case coverage distance
In this section, we present our ( 2 + )-approximation
algorithm, for any > 0, to dynamically maintain the
13
worst-case coverage distance. We first present a (1 + )-
approximation algorithm (for any > 0) for a simplified sen-
sor network model, where the coverage disks are considered
under infinity-norm. Since there is only a 2 gap between
infinity-norm and 2-norm, a (1 + /2)-approximation fac-
tor for infinity-norm dilates into a ( 2 + )-approximation
factor when applied to the 2-norm scenario, for any fixed
> 0. The infinity-norm of a vector v = (x1 , . . . , xd ) in a
d-dimensional space is defined as v = max(|x1 |,
. . . , |xd |). Under infinity-norm, the distance between two
points on the plane is the maximum of the difference of
their x coordinates and the difference of their y coordinates.
Hence the coverage region of a sensor is square shaped and
its boundary is composed of four line segments. As we will
see later, this simple boundary shape allows for an efficient
maintenance scheme.
Recall the solution framework presented in section 3. The
core of our algorithm is to check, for any two given points S
and T , whether the region in U (r) containing S is connected
to that containing T . If we can maintain some information
such that each query on connectivity of regions takes only
poly-logarithmic time, the cost of update against mobility is
also poly-logarithmic.
In our algorithm, regions in U (r) are represented by their
boundaries. Only one region in U (r) may be infinite in area.
We call such an unbounded region the outer face. All of the
other (bounded) regions are called inner faces. Since we con-
sider the infinity-norm, each disk is represented by a square
on the plane. Thus the boundary of any inner face is a sim-
ple cycle composed of a sequence of line segments, while the
boundary of the outer face comprises several simple cycles.
To differentiate these cycles, we call a cycle that is the bound-
(a)
(b)
Figure 2. Representation in G(U (r)). (a) A square is represented by 8 ver-
tices. (b) Dynamics of vertices and edges when squares overlap. Vertices C,
C  , and E and E  are relocated. Edges (C, C  ) and (E, E  ) are removed, and
edges (C, E  ) and (E, C  ) are inserted.
14
Figure 3. Outer- and inner-cycles. When the outer cycle in (a) breaks into
two cycles, it can either break into an outer and an inner cycle, as shown
in (b); or it can break into two outer cycles, as shown in (c).
ary of an inner face an inner cycle, and a cycle on the bound-
ary of the outer face an outer cycle. Figure 3 illustrates some
of these concepts. The shaded areas in the figure define U (r),
and the unshaded areas define U (r). In (b), U (r) is divided
into two regions, the unbounded region is the outer face, the
bounded region is the inner face. The boundary of the inner
face is an inner cycle and that of the outer face is an outer
cycle. In (c), the boundary of the outer face consists of two
disjoint outer cycles.
Below we describe a method which translates the connec-
tivity of regions in U (r) into a graph connectivity problem.
The first step is to represent outer cycles and inner cycles by
a graph. There are only vertical line segments and horizon-
tal line segments in both outer and inner cycles, and those
line segments only meet at their endpoints. Hence we can
draw a graph such that the vertices are the endpoints and the
edges are the line segments. We call this graph the connectiv-
ity graph G(U (r)). (For convenience, the connectivity graph
will actually be implemented in a slightly different way, as we
explain in section 5.1.)
Every outer or inner cycle is a cycle in the graph and any
two of them are disjoint, i.e., disconnected in the graph. This
coincides with the fact that any two distinct inner faces are
disconnected, and that any inner face is disconnected from
the outer face.
The connectivity of G(U (r)) is thus analogous to that of
U (r): Two regions are connected in U (r) if and only if their
boundary cycles are connected in the graph, or they are both
part of the outer face boundary. Thus we could apply the algo-
rithm proposed by Holm et al. [7], which dynamically main-
tains graph connectivity, to maintain the connectivity of the
regions in U (r). The update cost per edge insertion or dele-
HUANG ET AL.
tion for each G(U (r)) in [7] is amortized O(log2 n) and the
query cost is O(log n/ log log n), implying an overall amor-
tized update cost of O(log3 n) and worst-case query cost of
O(log n), with an approximation factor of (1 + ), for any
fixed > 0. However, G(U (r)) is the union of simple dis-
joint cycles, each uniquely defining a region in U (r) and thus
it allows for a more efficient update and query scheme, at the
expense of a small degradation in the approximation factor.
As we will show later, we can maintain the connectivity of all
G(U (r)) in overall worst-case update cost of O(log2n), with
worst-case query cost of O(1), while maintaining a ( 2 + )-
approximation on the worst-case coverage distance, for any
fixed > 0.
In the remainder of this section, we first describe the dy-
namics of the connectivity graph. Then we define three types
of events which mandate updates. Our update cost is charged
per event. Following that, we present a data structure, which
is an extension of concatenable queues [1], to maintain the
connectivity of the graph efficiently. Finally, we present our
major result on the worst-case coverage distance.
5.1. Dynamics of cycles
In this section, we first formally define the representation we
use for the connectivity graph G(U (r)). Second, we ad-
dress the dynamics of the connectivity graph. And finally,
we present an algorithm for maintaining the connectivity in-
formation on the regions of U (r).
The boundary of a standalone square is the simplest cycle
in G(U (r)). We represent a square by eight vertices and eight
edges as shown in figure 2. For every corner X of a square,
we introduce two vertices X and X . Hence we have O(n)
vertices and edges in G(U (r)), where n always denotes the
current number of sensors in the network. The extra vertices
help us to efficiently maintain the graph when squares start to
move and overlap on the plane (including when sensors are
added or removed from the network). In the following, we
will show that the dynamics of sensors will not change the
O(n) bound on the number of vertices and edges.
When two squares meet, at most two pairs of line seg-
ments of their boundaries intersect. Without loss of gener-
ality, suppose a vertical edge B  C intersects with a horizon-
tal edge E  F at a point Z, and the new boundary comprises
edges B  Z and ZF . Then we simply relocate vertices C and
E  to Z, insert an edge CE  and remove edges CC  and EE 
from G(U (r)). Figure 2 illustrates this operation. Note that
we do not introduce any new vertex or remove any old ver-
tex. In fact, since G(U (r)) contains no information of the
vertexs location, we do not need to perform any relocation
of a vertex when we operate on G(U (r)). The cases of a verti-
cal edge intersecting with a vertical edge, and of a horizontal
edge intersecting with a horizontal edge are analogous, and
can thus be also handled by at most two edge insertions and
at most two edge deletions. Since we never change the num-
ber of vertices in the graph, and since each vertex has degree
at most 2, the O(n) upper bound on number of vertices and
edges in G(U (r)) always hold. The following fact follows:
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS
Fact 1. When two squares meet or separate, up to four edge
insertions and deletions are needed to update the connectivity
graph G(U (r)).
When the topology of the network changes, cycles in
G(U (r)) may also undergo changes. A cycle may break into
two smaller cycles; or two cycles may merge into a longer
cycle. Both these operations impose changes on the connec-
tivity of G(U (r)). Cycles break or merge only when two sen-
sors coverage disks meet or separate. Hence we need to de-
tect the time when those happen in order to trigger an update.
When a cycle breaks, it could break into an outer cycle and
an inner cycle (as shown in figure 3). We need to differentiate
outer cycles from inner cycles since all outer cycles define
the same region, namely the outer face. In order to determine
whether a cycle is an outer cycle, one only needs to identify
the topmost edge of the cycle: If the topmost edge of the cycle
is the top boundary of a square, then the cycle is an outer
cycle; otherwise, the topmost edge of a cycle is the bottom
boundary of a square, and the cycle is an inner cycle. Hence
we need to maintain the topmost edge of each cycle as sensors
move. The topmost edge of a cycle may change only when
two horizontal line segments swap their y position. Therefore
we also need to monitor these line segment swaps.
Recall that the original problem we aim to solve is to check
whether the region containing a given point S is connected to
that containing T . We need to determine which region con-
tains a given point and also to update this information as sen-
sors move. As described in section 4, we sort all sensors ac-
cording to the distance from the fixed point S and maintain a
binary heap on this ordering, with update cost O(log n) on the
heap. In order to check which region S belongs to, we need
to find the cycle representing the region. Again we find the
closest sensor p to S and check if the distance is smaller than
the radius of the coverage disk of p. If so, then the point S
does not belong to any region of U (r). Otherwise, we check
the eight vertices of the square representing the closest sensor
to S, find the closest one of these vertices to S, and the cycle
containing this closest vertex represents the region contain-
ing S. This query takes constant time. We maintain a similar
data structure for T . Thus we also need to monitor and de-
tect the time when two sensors swap their relative position in
these orderings.
We summarize all of the above in the following three
types of events, which we need to monitor in order to trig-
ger mandatory updates, as sensors move on the plane:
(I) Two vertical line segments swap their x position,
(II) Two horizontal line segments swap their y position, and
(III) Two sensor swap their position in the orderings of the
sensors distance to the given points S and T .
When events (I) or (II) occurs, we can check in constant
time whether two coverage disks meet or separate. If they do,
we check whether the event leads to a cycle break or merge,
and update the data structure accordingly. When event (II) oc-
curs, we can check whether the two horizontal line segments
15
belong to the same cycle. If so, we may also need to update
the topmost edge of the cycle. When event (III) occurs, we
update the orderings with respect to distances to S and T .
We use the kinetic data structure as defined in [3] as our
event monitor (unlike G(U (r)), G(U (r)) is not a unit-disk
graph and therefore the results in [5] do not apply). Each
event can be detected and processed in constant time.
In the following, we present our update scheme. We will
also show that the update cost per event is O(log n). We store
a cycle as a sequence of consecutive edges. In section 5.2
we introduce a data structure which supports the following
operations on sequences of edges:
INSERT insert an edge into a sequence
DELETE delete an edge from a sequence
CONCATENATE concatenate a sequence to the end of an-
other sequence
SPLIT split a sequence into two sequences
SWAP swap the y position of two edges
MAX return the topmost edge of a sequence
MEMBER return the representative edge of a sequence
Each of these operations can be executed in worst-case
running time O(log n), as stated in lemma 4.
The update per type (I) or (II) event is as follows. When
squares move and the shape of a cycle changes, up to a con-
stant number of INSERT and DELETE operations are needed
to update the cycle per event. When two edges in a cycle ex-
changes their y position, we execute SWAP to update the y
position per event. We can execute MAX to know whether a
cycle is an outer cycle or not. Recall that a cycle is an outer
cycle if and only if the topmost edge of the cycle is the top
boundary line segment of a square. Cycle merges or breaks
can be carried out by a constant number of CONCATENATE
and SPLIT operations. Since only a constant number of IN-
SERT, DELETE, CONCATENATE, SPLIT, SWAP and MAX
operations are executed per event, the update cost per event is
worst-case O(log n). As we have explained earlier, the update
cost per type (III) event is also O(log n).
A data structure that supports the operations above can also
be used for efficiently performing a connectivity check. As-
sume that S and T are not covered by any sensor in U (r)
and therefore both belong to U (r). We can find the closest
vertices u and v to points S and T , respectively, in constant
time. Then we check if u and v belong to the same cycle by
performing two MEMBER operations. If so, then S and T
belong to the same region in U (r). Otherwise, we need to
check whether the closest cycles to S and T are both outer
cycles by two executions of the MAX operation. If both of
them are outer cycles, then both S and T belong to the outer
face, and hence are in the same region. Otherwise, S and T
belong to two disconnected regions. This procedure can be
implemented in O(log n) time.
We summarize all of the above in lemma 2.
Lemma 2. For any two given points S and T , we maintain a
data structure with O(log n) update cost per event such that
16
the query to check whether the region in U (r) containing S is
connected to that containing T takes O(log n) time.
Combining lemma 2 with the algorithmic framework pre-
sented in section 3, we have our (1 + )-approximation al-
gorithm, for any > 0, for the worst-case coverage distance
under infinity-norm, as stated in the lemma below. If every
time we perform an update operation, we keep track of the
smallest ri such that S and T are disconnected in G(U (ri )),
then each query operation can be performed in O(1) time.
Lemma 3. Under infinity-norm, our algorithm dynamically
maintains a (1 + )-approximation of the worst-case cover-
age distance for any > 0. The update cost is worst-case
O(log n log1+ R) per event, and the query cost is worst-
case O(1).
Hence the ( 2+)-approximation algorithm for the worst-
case coverage distance under 2-norm follows:
Theorem 2. Our algorithm dynamically maintains a ( 2+)
-approximation of the worst-case coverage distance, for any
> 0. The update cost is worst-case O(log n log1+/2 R) an O(n log n) runtime solution, thus improving on the best-
per event, and the query cost is worst-case O(1).
Corollary 2. If > 0 is fixed, then our algorithm has worst-
case O(log2 n) update cost per event, and worst-case O(1)
query cost.
5.2. Extended concatenable queue
In this subsection we introduce a data structure that sup-
ports the operations INSERT, DELETE, CONCATENATE,
SPLIT, SWAP, MAX and MEMBER efficiently. The data
structure is an extension of the concatenable queue data struc-
ture [1]. In [1], a concatenable queue is implemented by a
23 tree (a RedBlack tree would also work, for example),
and all the data is stored at the leaf nodes. A concaten-
able queue supports the operations INSERT, DELETE, CON-
CATENATE, SPLIT and MEMBER, and each operation takes
time O(log n) in the worst case. In the following paragraphs,
we will show how to also implement the SWAP and MAX
operations on a concatenable queue in O(log n) time.
We associate each edges y coordinate to the correspond-
ing leaf node in the 23 tree. To each internal node t, we asso-
ciate the maximum y coordinate of a leaf node in the subtree
rooted at t. This is done by comparing all the y coordinates
associated to ts children in the tree, taking constant time per
internal node. When the y coordinate of an edge changes, and
a SWAP operation is invoked, it takes at most O(log n) time to
climb up the tree and update all the internal nodes on the way
up. Starting from any given edge on a cycle, it takes O(log n)
time to reach the root of the 23 tree where we can find the
topmost edge of the cycle. Hence the O(log n) running time
of MAX follows.
We need also to justify that the above modification does
not increase the running time of all other operations. Per each
HUANG ET AL.
INSERT or DELETE, it takes an additional O(log n) time to
update the y coordinate of all internal nodes due to the edge
insertion or deletion. Both CONCATENATE and SPLIT are
implemented by up to O(log n) joins or breaks of trees at the
root node. Since updating the y coordinate at the root node
takes constant time (by comparing all the children of the root),
we incur at most an additional O(log n) time per CONCATE-
NATE or SPLIT. Thus the asymptotic running time of IN-
SERT, DELETE, CONCATENATE, and SPLIT remains un-
changed. The running time of MEMBER is not affected by
SWAP or MAX operations.
Lemma 4. The extension of the concatenable queue data
structure supports the operations of INSERT, DELETE,
CONCATENATE, SPLIT, SWAP, MAX and MEMBER.
Each operation has worst-case running time of O(log n).
6. Exact shortest maximum support path
We consider the problem of finding a maximum support path
between S and T such that the Euclidean length of the tra-
jectory followed by this path is minimum. Below we present
known previous results by Li et al. [9]. One should note that
the algorithms presented in [9] can be implemented in a dis-
tributed fashion, whereas the algorithm we present in this sec-
tion is intrinsically centralized.
We proceed as follows. First we compute the best coverage
radius rbest using the algorithm of Li et al. [9] in O(n log n)
time. Next, we obtain a collection of uniform disks by setting
the radius of each sensor to be rbest . Let U denote the union
of all these uniform disks. Define the complement region of
the union C = R2 \ U .
The problem of finding a shortest maximum support path
is equivalent to the problem of finding a shortest S, T -path
in R2 avoiding C, since we are seeking for a maximum sup-
port path and rbest is the best coverage radius. (Since rbest is
the best coverage radius, any maximum support path is con-
tained in U ; in fact any path from S to T in U is a maximum
support path.) A shortest maximum support path can only
contain straight line segments as edges, otherwise the path
would not be shortest. Therefore, we can replace each arc in
C by a straight line segment. In such fashion we obtain a new
set of obstacles C  as a collection of polygonal objects with
possible holes that have a total O(n) number of vertices.
We can remove these holes and obtain slightly larger num-
ber of disjoint polygonal objects by cutting the existing ob-
jects with segments that connect the vertices of the holes and
the external boundary in an arbitrary fashion. Note that the
number of total vertices of the disjoint polygonal objects has
not changed, i.e., it is still O(n). Thus, our problem translates
to that of finding a shortest path in R2 that avoids the polyg-
onal obstacles in C  . The idea now is to use an algorithm by
Hershberger and Suri [6], which finds a shortest path between
S and T on the Euclidean plane that avoids polygonal obsta-
cles in O(n log n) time. Hence the total running time of our
algorithm is O(n log n).
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS
As described in section 2, two algorithms are presented
for computing the maximum support path by Li et al. in [9]:
One algorithm computes an exact shortest maximum sup-
port path in O(n2 log n) time; the other algorithm provides a
2.5-approximation on the shortest maximum support path in
O(n log n) time. Our algorithm improves on the running time
of the former and on the approximation factor of the latter al-
gorithm. One should note, however, that the algorithms pre-
sented by Li et al. can be implemented in a distributed fashion
(we use the communication complexity as the time bound for
the sequential versions of their algorithms), whereas our al-
gorithm is centralized in nature.
7. Future work
In this paper, we present poly-logarithmic dynamic algo-
rithms to maintain approximations of two relevant measures
namely, the best- and worst-case coverage distances of the
quality of the network coverage in wireless sensor networks.
An interesting open question is whether we can maintain ex-
act best-case and worst-case coverage distances for Euclidean
metric with poly-logarithmic update time.
Acknowledgement
We express our thanks to Micha Sharir for his comments.
References
[1] A.V. Aho, J.E. Hopcroft and J.D. Ullman, The Design and Analysis of
Computer Algorithms (Addison-Wesley, Reading, MA, 1974).
[2] J. Basch, Kinetic data structures, Ph.D. dissertation, Stanford Univer-
sity (1999).
[3] J. Basch, L.J. Guibas and J. Hershberger, Data structures for mobile
data, in: Proc. of 8th ACMSIAM Symposium on Discrete Algorithms
(1997) pp. 747756.
[4] B.N. Clark and C.J. Colbourn, Unit disk graphs, Discrete Math. 86
(1990) 165177.
[5] L.J. Guibas, J. Hershberger, S. Suri and L. Zhang, Kinetic connectivity
for unit disks, Discrete Comput. Geom. 25 (2001) 591610.
[6] J. Hershberger and S. Suri, An optimal algorithm for Euclidean shortest
paths in the plane, SIAM J. Comput. 28(6) (1999) 22152256.
[7] J. Holm, K. de Lichtenberg and M. Thorup, Poly-logarithmic deter-
ministic fully-dynamic graph algorithms I: Connectivity and minimum
spanning tree, Technical Report DIKU-TR-97/17, Department of Com-
puter Science, University of Copenhagen (1997).
[8] C.-F. Huang and Y.-C. Tseng, The coverage problem in a wireless sen-
sor networks, in: Proc. of the 2nd ACM Internat. Conf. on Wireless
Sensor Networks and Applications (2003) pp. 115121.
[9] X.-Y. Li, P.-J. Wan and O. Frieder, Coverage in wireless ad-hoc sensor
networks, IEEE Trans. Comput. 52 (2003) 111.
[10] S. Meguerdichian, F. Koushanfar, M. Potkonjak and M.B. Srivastava,
Coverage problems in wireless ad-hoc sensor networks, in: Proc. of the
20th IEEE INFOCOM (2001) pp. 13801387.
[11] S. Meguerdichian, F. Koushanfar, G. Qu and M. Potkonjak, Exposure in
wireless ad-hoc sensor networks, in: Proc. of the 7th ACM MOBICOM
(2001) pp. 139150.
17
[12] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless and C.D. Gill, Integrated
coverage and connectivity configuration in wireless sensor networks,
in: Proc. of the 1st ACM Conf. on Embedded Networked Sensor Systems
(2003).
[13] H. Zhang and J.C. Hou, Maintaining sensing coverage and connectivity
in large sensor networks, Technical Report UIUCDCS-R-2003-2351,
UIUC (2003).
Hai Huang is a Ph.D. student in the Department of
Computer Science and Engineering at Arizona State
University. He has received a M.S. in the same
department under the supervision of Prof. Andrea
W. Richa in 2003. He received a B.A. and a M.A.
in the Mathematics Department at Tsinghua Univer-
sity, P. R. China, in 1996 and 1999, respectively. His
current research work focus on clustering and rout-
ing problems in mobile ad-hoc networks, and on the
chordal graph completion problem with applications
to scientific computing.
E-mail: hai@asu.edu
Andra W. Richa joined the Department of Com-
puter Science and Engineering at Arizona State Uni-
versity in 1998, where she is now an Associate Pro-
fessor. She received her M.S. and Ph.D. degrees
from the School of Computer Science at Carnegie
Mellon University, in 1995 and 1998, respectively.
She also earned an M.S. degree in computer systems
from the Graduate School in Engineering (COPPE),
and a B.S. degree in computer science, both at the
Federal University of Rio de Janeiro, Brazil, in 1992
and 1990, respectively. Prof. Richas main area of research is in network
algorithms. Some of the topics Dr. Richa has worked on include packet
scheduling, distributed load balancing, packet routing, mobile network clus-
tering and routing protocols, and distributed data tracking. Prof. Richas data
tracking (or lookup) algorithm has been widely recognized as the first bench-
mark algorithm for the development of distributed databases in peer-to-peer
networking, having received over 55 academic journal or conference publi-
cations, and being implemented as part of two of the current leading pojects
in peer-to-peer networking. Dr. Richas was the recipient of an NSF CA-
REER Award in 1999. For a selected list of her publications, CV, and current
research projects, please visit http://www.public.asu.edu/aricha.
E-mail: aricha@asu.edu
Michael Segal was born at October 12, 1972 in
USSR. In 1991 he immigrated to Israel and started
to study computer science in Ben-Gurion University
of the Negev. He finished his B.Sc., M.Sc. and Ph.D.
degrees in 1994, 1997, and 1999, respectively. Dur-
ing a period of 19992000 Dr. Michael Segal held a
MITACS National Centre of Excellence Postdoctoral
Fellow position in University of British Columbia,
Canada. Dr. Segal joined the Department of Com-
munication Systems Engineering, Ben-Gurion Uni-
versity, Israel in 2002 where he holds now a position of senior lecturer and
serves as departments Deputy Chairman. His primary research is algorithms
(sequential and distributed), data structures with applications to optimization
problems, mobile wireless networks, communications and security.
E-mail: segal@cse.bgu.ac.il
 Mobile Networks and Applications 10, 1934, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Algorithmic Aspects of Topology Control Problems for Ad Hoc

Networks
ERROL L. LLOYD and RUI LIU ,
Department of Computer and Information Sciences, University of Delaware, Newark, DE 19716, USA

MADHAV V. MARATHE
Los Alamos National Laboratory, MS M997, P.O. Box 1663, Los Alamos, NM 87545, USA

RAM RAMANATHAN
Internetwork Research Department, BBN Technologies, Cambridge, MA, USA

S.S. RAVI
Department of Computer Science, University at Albany SUNY, Albany, NY 12222, USA

Abstract. Topology control problems are concerned with the assignment of power values to the nodes of an ad hoc network so that the power
assignment leads to a graph topology satisfying some specified properties. This paper considers such problems under several optimization
objectives, including minimizing the maximum power and minimizing the total power. A general approach leading to a polynomial algorithm
is presented for minimizing maximum power for a class of graph properties called monotone properties. The difficulty of generalizing the
approach to properties that are not monotone is discussed. Problems involving the minimization of total power are known to be NP-complete
even for simple graph properties. A general approach that leads to an approximation algorithm for minimizing the total power for some
monotone properties is presented. Using this approach, a new approximation algorithm for the problem of minimizing the total power for
obtaining a 2-node-connected graph is developed. It is shown that this algorithm provides a constant performance guarantee. Experimental
results from an implementation of the approximation algorithm are also presented.
Keywords: power control, approximation algorithms, topology

1. Introduction Given the transmission powers of the transceivers, an

1.1. Motivation
An ad hoc network consists of a collection of transceivers.
All communication among these transceivers is based on ra-
dio propagation. For each ordered pair (u, v) of transceivers,
there is a transmission power threshold, denoted by p(u, v),
with the following significance: A signal transmitted by the
transceiver u can be received by v only when the transmis-
sion power of u is at least p(u, v). The transmission power
threshold for a pair of transceivers depends on a number of
factors including the distance between the transceivers, an-
tenna gains at the sender and receiver, interference, noise,
etc. [23].
A preliminary version of this paper appeared in Proc. of the Third ACM
International Symposium on Mobile Ad Hoc Networking and Comput-
ing (MobiHoc 2002), Lusanne, Switzerland, June 2002, pp. 123134.
Prepared through collaborative participation in the Communications
and Networks Consortium sponsored by the U.S. Army Research Lab-
oratory under the Collaborative Technology Alliance Program, Coop-
erative Agreement DAAD19-01-2-0011. The U.S. Government is au-
thorized to reproduce and distribute reprints for Government purposes
not withstanding any copyright notation thereon.
Research supported by the Department of Energy under Contract
W-7405-ENG-36.
Supported by NSF Grant CCR-97-34936. cussion of the power control issues in ad hoc networks.
ad hoc network can be represented by a directed graph. The
nodes of this directed graph are in one-to-one correspondence
with the transceivers. A directed edge (u, v) is in this graph
if and only if the transmission power of u is at least the trans-
mission power threshold p(u, v).
The main goal of topology control is to assign transmission
powers to transceivers so that the resulting directed graph sat-
isfies some specified properties. Since the battery power of
each transceiver is an expensive resource, it is important to
achieve the goal while minimizing a given function of the
transmission powers assigned to the transceivers. Examples
of desirable graph properties are connectivity, small diameter,
etc. Examples of minimization objectives considered in the
literature are the maximum power assigned to a transceiver
and the total power of all transceivers (the latter objective
is equivalent to minimizing the average power assigned to a
transceiver).
As stated above, the primary motivation for studying topol-
ogy control problems is to make efficient use of available
power at each node. In addition, using a minimum amount
of power at each node to achieve a given task is also likely to
decrease the MAC layer interference between adjacent radios.
We refer the reader to [20,22,23,26,28,31] for a thorough dis-
20
1.2. Formulation of topology control problems
Topology control problems have been studied under two
graph models. The discussion above corresponds to the di-
rected graph model studied in [23]. The undirected graph
model proposed in [16] represents the ad hoc network as an
undirected graph in the following manner. First, the directed
graph model for the network is constructed. Then, for any
pair of nodes u and v, whenever both the directed edges (u, v)
and (v, u) are present, this pair of directed edges is replaced
by a single undirected edge {u, v}. All of the remaining di-
rected edges are deleted. Under this model, the goal of a
topology control problem is to assign transmission powers to
nodes such that the resulting undirected graph has a specified
property and a specified function of the powers assigned to
nodes is minimized. Note that the directed graph model al-
lows two-way communication between some pairs of nodes
and one-way communication between other pairs of nodes.
In contrast, every edge in the undirected graph model corre-
sponds to a two-way communication.
In general, a topology control problem can be specified
by a triple of the form M, P, O. In such a specification,
M {D IR , U NDIR } represents the graph model, P repre-
sents the desired graph property and O represents the mini-
mization objective. For the problems considered in this pa-
per O {M AX P, T OTAL P} (abbreviations of Max Power and
Total Power). For example, consider the D IR , S TRONGLY
C ONNECTED, M AX P problem. Here, powers must be as-
signed to transceivers so that the resulting directed graph is
strongly connected and the maximum power assigned to a
transceiver is minimized. Similarly, the U NDIR , 2-N ODE
C ONNECTED, T OTAL P problem seeks to assign powers to
the transceivers so that the resulting undirected graph has a
node connectivity (see below for definition) of (at least) 2 and
the sum of the powers assigned to all transceivers is mini-
mized.
2. Additional definitions
This section collects together the definitions of some graph
theoretic and algorithmic terms used throughout this paper.
Given an undirected graph G(V , E), an edge subgraph
G (V , E  ) of G has all of the nodes of G and the edge set E 
is a subset of E. Further, if G is an edge weighted graph, then
the weight of each edge in G is the same as it is in G.
The node (edge) connectivity of an undirected graph is the
smallest number of nodes (edges) that must be deleted from
the graph so that the resulting graph is disconnected. For
example, a tree has node and edge connectivities equal to 1
while a simple cycle has node and edge connectivities equal
to 2. When the node (edge) connectivity of a graph is greater
than or equal to k, the graph is said to be k-node connected
(k-edge connected). Given an undirected graph, polynomial
algorithms are known for finding its node and edge connec-
tivities [30].
The main results of this paper use the following definition.
LLOYD ET AL.
Definition 2.1. A property P of the (directed or undirected)
graph associated with an ad hoc network is monotone if the
property continues to hold even when the powers assigned to
some nodes are increased while the powers assigned to the
other nodes remain unchanged.
Example. For any k
1, the property k-N ODE C ONNECTED
for undirected graphs is monotone since increasing the pow-
ers of some nodes while keeping the powers of other nodes
unchanged may only add edges to the graph. However, prop-
erties such as ACYCLIC or B IPARTITE are not monotone.
Some of the topology control problems considered in this
paper are NP-complete. For such problems, we study approx-
imation algorithms. In this context, an approximation algo-
rithm provides a performance guarantee of if for every in-
stance of the problem, the solution produced by the approxi-
mation algorithm is within the multiplicative factor of of the
optimal solution. A polynomial time approximation scheme
(PTAS) is an approximation algorithm that, given a problem
instance and an accuracy requirement , produces a solution
that is within a factor 1 + of the optimal solution.
3. Previous work and summary of results
3.1. Previous work
The form of topology control problems considered in this
paper was proposed by Ramanathan and Rosales-Hain [23].
They presented efficient algorithms for two topology control
problems, namely U NDIR , 1-N ODE C ONNECTED, M AX P
and U NDIR , 2-N ODE C ONNECTED, M AX P. After deter-
mining the minimum value for the objective, their algorithms
also reduce the power assigned to each transceiver such that
each power level is minimal while maintaining the desired
graph property. In addition, they presented efficient distrib-
uted heuristics for these problems.
Several groups of researchers have studied the D IR ,
S TRONGLY C ONNECTED, T OTAL P problem [5,7,8,16].
However, it is not difficult to see that their NP-hardness re-
sults as well as approximation algorithms also hold for the
U NDIR , 1-N ODE C ONNECTED, T OTAL P problem. The pa-
per [5] proves that the problem is NP-hard and presents
an approximation algorithm with a performance guarantee
of 2. The other references consider a geometric version of
the problem along with a symmetry assumption concerning
transmission power thresholds. More precisely, these refer-
ences assume the following: (a) Each transceiver is located
at some point of d-dimensional Euclidean space. (b) For
any pair of transceivers u and v, p(u, v) = p(v, u) =
the Euclidean distance between the locations of u and v.
For a justification of this model, see [16]. They show
that the D IR , S TRONGLY C ONNECTED, T OTAL P problem
is NP-hard when transceivers are located in 3-dimensional
space. They also present an approximation algorithm with
a performance guarantee of 2 for the problem in any met-
ric space. In addition, they provide some results for the
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 21

1-dimensional version of the D IR , S TRONGLY C ONNECTED,
T OTAL P problem where there is an additional constraint on
the diameter of the resulting undirected graph. Clementi
et al. [7] show that the 2-dimensional version of the
D IR , S TRONGLY C ONNECTED , T OTAL P problem remains
NP-hard. They also show that the 2-dimensional version
with a diameter constraint can be efficiently approximated to
within some constant factor and that the 3-dimensional ver-
sion does not have a polynomial time approximation scheme.
Under a slightly different model, where there is an explicit re-
lationship between the transmission power and distance, ref-
erences [2,3] study topology control problems for connectiv-
ity properties. The complexity of several problems under this
model is established in [2]. A (1 + ln 2)-approximation algo-
5. Finally, we present experimental results obtained from
rithm for the problem is presented in [3]. The approximation
ratio is improved to 5/3 in a journal submission based on [3].
Additional related work may be found in [13,1820,
22,31].
general approach. As an illustration of our general ap-
proach, we present a constant factor approximation algo-
rithm for the U NDIR , 2-N ODE C ONNECTED, T OTAL P
problem. No approximation algorithm was previously
known for this problem. In analyzing this approxi-
mation algorithm, we use some properties of critically
2-node connected graphs [10,21,32]. By a minor modi-
fication to this approximation algorithm, we also obtain
a constant factor approximation algorithm for producing
2-edge-connected graphs. As in the case of minimizing
maximum power, our general heuristic for approximating
total power is also applicable to graph properties specified
by proper functions.
an implementation of the above approximation algorithm
and compare its performance with an algorithm discussed
in [23].

3.2. Summary of main results 4. Results for minimizing maximum power

Throughout this paper, it is assumed that the power threshold In this section, we present our results for the M AX P OWER
values are symmetric. The main results of this paper are the objective. We begin with a general algorithm for the topology
following. control problem where the graph property is both monotone
1. We show that for any monotone graph property P that and polynomial time testable. For a problem with n trans-
can be tested in polynomial time for undirected (directed) ceivers, the algorithm uses O(log n) invocations of the algo-
graphs, the problem U NDIR , P, M AX P (D IR , P, M AX P) rithm to test the graph property. We also present a polynomial
can be solved in polynomial time. This generalizes time approximation scheme which can, under certain circum-
some of the results in [23] where efficient algorithms stances, substantially reduce the number of invocations of the
were presented for two monotone properties, namely, property testing algorithm. Next, we give an example of a
1-N ODE C ONNECTED and 2-N ODE C ONNECTED. Our nonmonotone property for which the problem of minimizing
polynomial time algorithm can also be extended to graph the maximum power is NP-hard. Finally, we show that the ad-
properties specified by proper functions1 [12]. ditional requirement of minimizing the number of nodes that
use the maximum power also renders the problem NP-hard,
2. We establish that there are nonmonotone and efficiently even for certain monotone properties. Note that both of the
testable properties (e.g., G RAPH IS A TREE) for which NP-hardness results utilize arbitrary power thresholds. The
even determining whether there is a power assignment complexity of the problems in the geometric model (i.e., the
that can induce a graph with the specified property is power threshold is a function of the Euclidean distance) re-
NP-complete. This result shows that, in general, if the mains open.
monotonicity condition is eliminated, then obtaining an
efficient algorithm for minimizing maximum power may 4.1. An algorithm for monotone and efficiently testable
not be possible. properties
3. As mentioned above, for any monotone and efficiently
We begin with a simple lemma that points out the usefulness
testable property P, a solution that minimizes the maxi-
of monotonicity.
mum power can be obtained in polynomial time. How-
ever, if we introduce the additional requirement that the
Lemma 4.1. For any instance of U NDIR , P, M AX P and
number of nodes that use the maximum power must also
D IR , P, M AX P where the graph property P is monotone,
be minimized, we show that there are monotone properties
there is an optimal solution in which all of the nodes are as-
for which the resulting problem is NP-complete.
signed the same power value.
4. We present a general approach for developing approxima-
tion algorithms for NP-hard topology control problems un- Proof. Consider an optimal solution to the given instance
der the T OTAL P OWER minimization objective. The ap- where the nodes do not necessarily have the same power val-
proximation results of [5,6,16] are special cases of this ues. Let Q denote the maximum power assigned to any node.
1 Given a graph G(V , E), a function is f : 2V {0, 1} is proper if it satisfies Since the graph property is monotone, for any node whose
the following two conditions: (1) f (S) = f (V S) for all S V ; and
power value is less than Q, we can increase it to Q without
(2) If A B = , then f (A) = f (B) = 0 implies f (A B) = 0. destroying the property. 
22
Theorem 4.1. For any monotone and polynomial time tes-
table graph property P, the problems U NDIR , P, M AX P and
D IR , P, M AX P can be solved in polynomial time.
Proof. We will present the proof for D IR , P, M AX P. (The
proof for U NDIR , P, M AX P is virtually identical.)
Consider an instance of D IR , P, M AX P. By lemma 4.1,
there is an optimal solution in which every transceiver is as-
signed the same power value. We can estimate the number of
candidate optimal power values as follows. Let T denote the
set of all transceivers in the system and let |T | = n. Consider
any transceiver u T . The number of different power values
that need to be considered for u is at most n 1, since at most
one new power value is needed for each transceiver in T {u}.
Therefore, for all of the n transceivers, the total number of
candidate power values to be considered is n(n 1) = O(n2 ).
For each candidate power value, the corresponding di-
rected graph can be constructed in O(n2 ) time. Let FP (n)
denote the time needed to test whether property P holds for
a directed graph with n nodes. Thus, the time needed to
test whether property P holds for each candidate solution
value is O(n2 + FP (n)). An optimal solution can be ob-
tained by sorting the O(n2 ) candidate solution values and us-
ing binary search to determine the smallest value for which
property P holds. Since the number of candidate solu-
tion values is O(n2 ), the time taken by the sorting step is
O(n2 log n). The binary search would try O(log n) candidate
solution values and the time spent for testing each candidate
is O(n2 + FP (n)). Thus, the total running time of this algo-
rithm is O((n2 + FP (n)) log n). Since FP (n) is a polynomial,
the algorithm runs in polynomial time.  have the property P. The binary search uses O(log k) =
As an illustration of the above theorem, let P denote the
property 2-N ODE C ONNECTED for undirected graphs. It is
known that this property can be tested in O(n2 ) time for a
graph with n nodes [30]. For this property, the general al-
gorithm outlined in the proof of theorem 4.1 yields an algo-
rithm with a running time of O(n2 log n). This running time
matches the time of the algorithm given in [23]. However, it
should be noted that the algorithm in [23] not only finds an op-
timal solution but also reduces the power of each transceiver
so that the power levels are minimal. There is no increase in
their asymptotic running time.
Instead of requiring the entire graph to be connected, one
may require connectivity only for a specified subset of the
nodes. Such a requirement arises in the context of multicas-
ting (see, for example, [27]), where the subset of nodes in-
cludes the sender and all the intended receivers. Connect-
edness of a specified subset of nodes can be seen to be a
monotone property. Thus, the general approach presented
above leads to a polynomial time algorithm for this prop-
erty as well. In fact, the result extends to large class of
network design problems that can be specified using proper
functions [1,12]. As noted in [12], the Steiner tree problem
and the Steiner forest problem can be specified using this for-
malism. Given a network and a proper function specifica-
tion, it is easy to test in polynomial time if the network sat-
LLOYD ET AL.
isfies the given proper function. Moreover, it is easy to see
that any graph property specified using a proper function is
a monotone property. Thus, our results apply to this class of
network design problems as well.
We now present a polynomial time approximation scheme
for U NDIR , P, M AX P and D IR , P, M AX P problems. As a
compensation for the slight deviation from the optimal value,
this approach has the potential to reduce the running time sub-
stantially.
Theorem 4.2. Let P be a monotone graph property that can
be tested for an n-node graph in time FP (n). For any fixed
> 0, the problems U NDIR , P, M AX P and D IR , P, M AX P
can be approximated to within the factor 1 + in O((n2 +
FP (n)) log log(max / min)) time, where max and min are re-
spectively the maximum and minimum power threshold val-
ues in the given problem instance.
Proof. We will present the proof for D IR , P, M AX P.
Since the number of power threshold values is O(n2 ), the val-
ues of min and max can be found in O(n2 ) time. Note that
for any candidate power value (which is assigned to all the
nodes), testing whether P holds for the induced graph can be
done in O(n2 + P(n)) time.
Let k be the smallest integer such that (1 +)k min
max.
Thus, k = O(log(max / min)). Consider the following set of
k + 1 power values: {min, (1 + ) min, (1 + )2 min, . . . ,
(1 + )(k1) min, max}. By doing a binary search on this
set, we can determine the smallest integer j such that
the power value (1 + )j min causes the induced graph to
O(log log(max / min)) calls to the algorithm for testing P.
Thus, the running time of the algorithm is O((n2 + FP (n)) log
log(max / min)).
Further, since j is the smallest value for which the power
value (1 + )j min causes the induced graph to have the prop-
erty P, the optimal value must be at least (1 + )(j 1) min.
Thus, the solution found by the algorithm is within a factor
(1 + ) of the optimal value. 
When the ratio max / min is substantially smaller than 2n ,
the above approximation scheme reduces the number of calls
to the property testing algorithm to a value that is asymptoti-
cally smaller than O(log n).
4.2. Difficulty of generalizing to nonmonotone properties
We now show that there is a natural nonmonotone graph
property for which the problem of minimizing the maximum
power is NP-hard. As mentioned earlier, this result points out
that if the monotonicity requirement is omitted, then an effi-
cient algorithm for minimizing maximum power may not be
possible.
The property that we use for this purpose is G IS A
T REE. Surprisingly, we show that this property makes the
topology control problem NP-complete even without any
minimization objective. The proof of lemma 4.2 utilizes a
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS
reduction from Exact Cover by 3-Sets (X3C), which is known
to be NP-complete [11].
Lemma 4.2. To determine whether there is a power assign-
ment such that the resulting undirected graph G is a tree is
NP-complete.
Proof. See appendix A.  A slight modification of this approximation algorithm yields
Theorem 4.3. There is a nonmonotone property P for which
U NDIR , P, M AX P is NP-hard.
Proof. Let P denote the property G IS A T REE.
The NP-hardness of U NDIR , P, M AX P follows from
lemma 4.2.  is, for any pair of transceivers u and v, the power thresholds
4.3. Difficulty of minimizing the number of nodes of
maximum power
An extension of U NDIR , P, M AX P for monotone graph
properties is explored in this section. While such problems
can be solved efficiently, our algorithm in section 4.1 assigns
the maximum power value to all of the nodes. From a prac-
tical point of view, it is important to reduce the number of
nodes with maximum power without affecting the required
property. In this section, we show that this additional require-
ment renders the problem NP-hard even for certain monotone
graph properties. A formal statement of the decision version
of the problem is as follows.
Max-power users.
Instance: A positive integer M, a positive number P (max-
imum allowable power value), a node set V , a power
threshold value p(u, v) for each pair (u, v) of transceivers
and a graph property P.
Question: Is there a power assignment where the power
assigned to each node is at most P and the number of the
nodes that are assigned power P is at most M, such that
the resulting undirected graph G satisfies P?
Theorem 4.4. There is a monotone and polynomial time
testable property P for which the problem Max-power users
is NP-complete.
Proof. See appendix B.  graph G (V , E  ) of G with total edge weight W (G ).
5. A general approach for minimizing total power
5.1. Approximating minimum total power
Topology control problems in which the minimization objec-
tive is total power tend to be computationally intractable. For
example, the problem is NP-hard even for the (simple) prop-
erty 1-N ODE -C ONNECTED [16]. A common way of coping
with such problems is to develop polynomial time approxima-
tion algorithms for them. In this section, we present a general
23
outline for such an approximation algorithm for topology con-
trol problems of the form U NDIR , P, T OTAL P. We observe
that this general outline encompasses the approximation al-
gorithm for U NDIR , 1-N ODE C ONNECTED, T OTAL P pre-
sented in [16]. Based on the general outline, we also develop
an approximation algorithm with a constant performance
guarantee for U NDIR , 2-N ODE C ONNECTED , T OTAL P.
an approximation algorithm for the problem of obtaining a
2-edge-connected graph while minimizing total power.
In presenting our general scheme, we assume (as done in
section 4.1) that the property P to be satisfied by the graph is
monotone and that it can be tested in polynomial time. We
also assume symmetric power thresholds as in [7,8,16]; that
p(u, v) and p(v, u) are equal.
An outline for our general approximation algorithm (called
Heuristic G EN -T OTAL -P OWER) is shown in figure 1. Note
that steps 1 and 3 of the outline can be implemented in poly-
nomial time. The time complexity of step 2 depends cru-
cially on the property P. For some properties such as 1-
N ODE C ONNECTED, step 2 can be done in polynomial time.
For other properties such as 2-N ODE C ONNECTED, step 2
cannot be done in polynomial time, unless P = NP [11].
In such cases, an efficient algorithm that produces an ap-
proximately minimum solution can be used in step 2. The
following theorem proves the correctness of the general ap-
proach and establishes its performance guarantee as a func-
tion of some parameters that depend on the property P and
the approximation algorithm used in step 2 of the general out-
line.
Theorem 5.1. Let I be an instance of U NDIR , P, T OTAL P
where P is a monotone property. Let OPT(I ) and GTP(I )
denote respectively the total power assigned to the nodes in
an optimal solution and in a solution produced by Heuristic
G EN -T OTAL -P OWER for the instance I .
(i) The graph G resulting from the power assignment pro-
duced by the heuristic (i.e. step 3) satisfies property P.
(ii) Consider the complete graph Gc (V , Ec ) constructed in
step 1 of the heuristic. Let H (V , EH ) be an edge sub-
graph of Gc with minimum total edge weight satisfying
property P and let W (H ) denote the total edge weight
of H . Let step 2 of the heuristic produce an edge sub-
Suppose there are quantities > 0 and > 0 such that
(a) W (H )  OPT(I ) and (b) W (G )  W (H ). Then,
GTP(I )  2OPT(I ). That is, Heuristic G EN -T OTAL -
P OWER provides a performance guarantee of 2.
Before proceeding to the proof of this result, we illus-
trate its use by discussing how the 2-approximation algorithm
presented in [16] for the U NDIR , 1-N ODE C ONNECTED,
T OTAL P problem can be derived from the above general out-
line. In step 2 they use an efficient algorithm for constructing
a minimum spanning tree of Gc . They also show that the total
24 LLOYD ET AL.

power assigned by any optimal solution is at least the weight
of a minimum spanning tree of Gc . Thus, using the notation
of theorem 5.1, = = 1 for their approximation algo-
rithm. Since 1-N ODE -C ONNECTED is a monotone property,
it follows from theorem 5.1 that the performance guarantee of
their algorithm is 2.
Proof of theorem 5.1. Part (i). The edge subgraph G (V , E  )
constructed in step 2 of the heuristic satisfies property P. We
show that every edge in E  is also in the subgraph G induced
by the power assignment produced in step 3. Then, even if G
has other edges, the monotonicity of P allows us to conclude
that G satisfies P.
Consider an edge {u, v} with weight p(u, v) in E  . Recall
that p(u, v) is the minimum power threshold for the existence
of edge {u, v} and that the power thresholds are symmetric.
Since step 3 assigns to each node the maximum of the weights
of edges incident on that node, we have (u)
p(u, v) and
(v)
p(u, v). Therefore, the graph G induced by the
power assignment also contains the edge {u, v} and this com-
pletes the proof of part (i).
Part (ii). By conditions (a) and (b) in the statement of the
theorem, we have W (G )  OPT(I ). We observe that
GTP(I )  2 W (G ). This is because in step 3 of the heuris-
tic, the weight of any edge is assigned to at most two nodes
(namely, the endpoints of the edge). Combining the two in-
equalities, we get GTP(I )  2OPT(I ), and this completes
the proof of theorem 5.1.  sulting approximation algorithm for the U NDIR , 2-N ODE
5.2. New approximation algorithms
This section presents two new approximation algorithms de-
rived from the general approach outlined in figure 1. These
algorithms are for the two monotone properties 2-N ODE
C ONNECTED and 2-E DGE C ONNECTED respectively. The
corresponding problems are denoted by U NDIR , 2-N ODE
C ONNECTED, T OTAL P and U NDIR , 2-E DGE C ONNECTED ,
T OTAL P.
5.2.1. An approximation algorithm for U NDIR , 2-N ODE
C ONNECTED, T OTAL P
This section presents an approximation algorithm for the
U NDIR , 2-N ODE C ONNECTED, T OTAL P problem. The
NP-hardness of this problem is established in [4]. Our
algorithm is derived from the general approach outlined in
figure 1. The following notation is used throughout this
section. I denotes the given instance of U NDIR , 2-N ODE
C ONNECTED, T OTAL P with n transceivers. For each trans-
ceiver u, (u) denotes the power assigned to u in an optimal
solution. Further, OPT(I ) denotes the sum of the powers as-
signed to the nodes in an optimal solution.
We obtain an approximation algorithm for the U NDIR ,
2-N ODE C ONNECTED, T OTAL P problem from the outline
of figure 1 by using an approximation algorithm from [14] for
the minimum weight 2-N ODE -C ONNECTED subgraph prob-
lem in step 2 of the outline. This approximation algorithm
provides a performance guarantee of (2 + 1/n). Using the
notation of theorem 5.1, we have  (2 + 1/n).
We also show (see lemma 5.1 below) that for the com-
plete edge weighted graph Gc (V , Ec ) constructed from the
instance I in step 1 of the outline, there is an edge sub-
graph G1 (V , E1 ) such that G1 is 2-N ODE -C ONNECTED and
the total weight W (G1 ) of the edges in G1 is at most (2
2/n)OPT(I ). Using the notation of theorem 5.1, this result
implies that  (2 2/n).
Thus, once we establish lemma 5.1, it would follow
from theorem 5.1 that the performance guarantee of the re-
C ONNECTED, T OTAL P problem is 2(2 2/n) (2 + 1/n),
which approaches 8 asymptotically from below. The remain-
der of this section is devoted to the formal statement and proof
of lemma 5.1.
Lemma 5.1. Let I denote an instance of the U NDIR ,
2-N ODE C ONNECTED, T OTAL P problem with n transceiv-
ers. Let OPT(I ) denote the total power assigned to the trans-
ceivers in an optimal solution to I . Let Gc (V , Ec ) denote
the complete graph constructed in step 1 of Heuristic G EN -
T OTAL -P OWER. There is an edge subgraph G1 (V , E1 ) of Gc

Input: An instance I of U NDIR , P, T OTAL P where the property P is monotone and polynomial time testable.
Output: A power value (u) for each transceiver u such that the graph induced by the power assignment satisfies
property P and the total power assigned to all nodes is as small as possible.
Steps:
1. From the given problem instance, construct the following undirected complete edge weighted graph Gc (V , Ec ).
The node set V is in one-to-one correspondence with the set of transceivers. The weight of every edge {u, v} in
Ec is equal to the power threshold value p(u, v) (which is also equal to p(v, u) by the symmetry assumption).
2. Construct an edge subgraph G (V , E  ) of Gc such that G satisfies property P and the total weight of the edges
in E  is minimum among all edge subgraphs of Gc satisfying property P.
3. For each node (transceiver) u, assign a power value (u) equal to the weight of the largest edge incident on u.

Figure 1. Outline of heuristic G EN -T OTAL -P OWER for approximating total power.

ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS
Figure 2. A simple cycle v1 , v2 , v3 , v4 , v5 , v6 , v1  with two chords {v1 , v5 }
and {v3 , v6 }.
such that G1 is 2-N ODE -C ONNECTED and the total weight
W (G1 ) of the edges in G1 is at most (2 2/n)OPT(I ).
Our proof of lemma 5.1 begins with an optimal power
assignment to instance I and constructs a graph G1 satisfy-
ing the properties mentioned in the above statement. This
construction relies on several definitions and known results
from graph theory. We begin with the necessary defini-
tions.
Definition 5.1. Let G(V , E) be an undirected graph. Sup-
pose the node sequence v1 , v2 , v3 , . . . , vk , v1  forms a sim-
ple cycle C of length at least 4 in G. Any edge {vi , vj } of G
(1  i = j  k) which is not in C is a chord.
Figure 2 shows a simple cycle of length 6 with two chords.
Definition 5.2. An undirected graph G(V , E) is critically 2-
N ODE -C ONNECTED if it satisfies both of the following con-
ditions: (i) G is 2-N ODE -C ONNECTED. (ii) For every edge
e E, the subgraph of G obtained by deleting the edge e is
not 2-N ODE -C ONNECTED.
For example, a simple cycle on three or more nodes is crit-
ically 2-N ODE -C ONNECTED. This is because such a cycle
is 2-N ODE -C ONNECTED, and deleting any edge of the cycle
yields a simple path which is not 2-N ODE -C ONNECTED.
A number of properties of critically 2-N ODE -C ONNECTED
graphs have been established in the literature (see, for exam-
ple, [10,21,32]). We use the following property in proving
lemma 5.1.
Theorem 5.2. If a graph G is critically 2-N ODE -C ON -
NECTED then no cycle of G has a chord.
25
For a proof of the above2 theorem, see [10,21]. We also
use some terminology associated with Depth-First-Search
(DFS) [9]. When DFS is carried out on a connected undi-
rected graph G(V , E), a spanning tree T (V , ET ) is produced.
Each edge in T , called a tree edge, joins a child to its parent.
An ancestor of a node u in T is a node which is not the parent
of u but which is encountered in the path from u to the root
of T . Each edge in E ET , called a back edge, joins a node u
to an ancestor of u. The following lemma establishes a simple
property of back edges that arise when DFS is carried out on
a critically 2-N ODE -C ONNECTED graph.
Lemma 5.2. Let G(V , E) be a critically 2-N ODE -C ON -
NECTED graph and let T (V , ET ) be a spanning tree for G
produced using DFS. For any node u, there is at most one
back edge from u to an ancestor of u in T .
Proof. The proof is by contradiction. Suppose a node u has
two or more back edges. Let v and w be two ancestors of u
in T such that both {u, v} and {u, w} are back edges. Note that
these two edges are in G. Without loss of generality, let w be
encountered before v in the path in T from the root to u. The
path from w to u in T together with the edge {u, w} forms
a cycle in G. By our choice of w, this cycle also includes
the node v. Therefore, the edge {u, v} is a chord in the cycle.
This contradicts the assumption that G is critically 2-N ODE -
C ONNECTED since by theorem 5.2, no cycle in G can have a
chord. The lemma follows. 
We now prove several additional lemmas that are used in
our proof of lemma 5.1. Consider the given instance I of
the U NDIR , 2-N ODE C ONNECTED, T OTAL P problem and
let V denote the set of transceivers. Fix an optimal solution to
the instance I and let p denote the maximum power value as-
signed to a node in this optimal solution. Let the chosen opti-
mal power assignment induce the graph G (V , E ). Note that
G is 2-N ODE -C ONNECTED. Let G1 (V , E1 ) be an edge sub-
graph of G such that G1 is critically 2-N ODE -C ONNECTED.
(Such a subgraph can be obtained by starting with G and re-
peatedly removing edges until no further edge deletion is pos-
sible without violating the 2-N ODE -C ONNECTED property.)
For each edge {u, v} of G1 , we assign a weight w1 (u, v) as
follows.
1. Let r be a node such that (r) = p . Using r as the
root, perform a DFS of G1 . Let T (V , ET ) be the resulting
spanning tree. Thus, each edge of G1 is either a tree edge
or a back edge.
2. For each tree edge {u, v} where v is the parent of u, let
w1 (u, v) = (u).
3. For each back edge {u, v} where v is an ancestor of u, let
w1 (u, v) = (u).
2 It should be noted that the graph theoretic terminology used in [10,21] is
different from ours. The statement of theorem 5.2 given above is from [32].
26 LLOYD ET AL.

The following lemma bounds the total weight W1 (G1 ) of
all the edges in G1 under the edge weight function w1 chosen
above.
Lemma 5.3. W1 (G1 )  (2 2/n)OPT(I ).
Proof. As mentioned above, each edge of G1 is either a tree
edge or a back edge. Consider the tree edges first. For each
tree edge {u, v}, where v is the parent of u, w1 (u, v) = (u).
Thus, the weight (u) is assigned to at most one tree edge
(namely, the edge that joins u to the parent of u if any
in T ). The power value of the root r in the optimal solution,
namely p , is not assigned to any tree edge (since the root
has no parent). Thus, the total weight of all of the tree edges
under the weight function w1 is bounded by OPT(I ) p .
Now consider the back edges. For each back edge {u, v},
where v is an ancestor of u, w1 (u, v) = (u). Since G1
is critically 2-N ODE -C ONNECTED, by lemma 5.2, each node
has at most one back edge to an ancestor. Thus, the weight
(u) is assigned to at most one back edge. Again, the
power value p of the root r in the optimal solution is not
assigned to any back edge. Thus, the total weight of all of the
back edges under the weight function w1 is also bounded by
OPT(I ) p .
Therefore, the total weight W1 (G1 ) of all of the edges
in G1 under the edge weight function w1 is at most 2OPT(I )
2p . Since p is the largest power value assigned to a node
in the optimal solution, p is at least OPT(I )/n. Hence,
W1 (G1 ) is bounded by (2 2/n)OPT(I ) as required.  to show that  (2 1/n). The proof of this result is almost
The following lemma relates the weight w1 (u, v) of an
edge {u, v} to the power threshold p(u, v) needed for the ex-
istence of the edge.
is at most W1 (G1 ). By lemma 5.3, W1 (G1 ) is bounded
by (2 2/n)OPT(I ). Therefore, W (G1 ) is also bounded
by (2 2/n)OPT(I ). In other words, the edge subgraph
G1 (V , E1 ) is 2-N ODE - CONNECTED and the total weight of
all its edges is at most (2 2/n)OPT(I ). This completes the
proof of lemma 5.1. 
The following is a direct consequence of the above discus-
sion.
Theorem 5.3. There is a polynomial time approximation al-
gorithm with a performance guarantee of 2(22/n)(2+1/n)
(which approaches 8 asymptotically from below) for the
U NDIR , 2-N ODE C ONNECTED, T OTAL P problem.
5.2.2. An approximation algorithm for U NDIR , 2-E DGE
C ONNECTED, T OTAL P
A result analogous to theorem 5.3 can also be obtained for
U NDIR , 2-E DGE C ONNECTED, T OTAL P where the goal is
to induce a graph that has the monotone property 2-E DGE
C ONNECTED. This problem has also been shown to be NP-
complete in [4]. To obtain an approximation algorithm for
this problem from the general framework, we use an ap-
proximation algorithm of Khuller and Vishkin [15]. Their
approximation algorithm produces a 2-edge-connected sub-
graph whose cost is at most twice that of a minimum 2-edge
connected subgraph. In the notation of theorem 5.1, we have
 2. Again using the notation of theorem 5.1, it is possible
identical to that for the 2-Node-Connected case, except that
we need an analog of theorem 5.2. Before stating this ana-
log, we have the following definition (which is analogous to
definition 5.2).

Lemma 5.4. For any edge {u, v} in G1 , p(u, v)  w1 (u, v).
Proof. Consider any edge {u, v} in G1 . Since G1 is an edge
subgraph of G (the graph induced by the chosen optimal
power assignment), {u, v} is also an edge in G . Also, re-
call that the minimum power threshold values are symmetric.
Therefore, (u)
p(u, v) and (v)
p(u, v). Hence
min{ (u), (v)}
p(u, v). The weight assigned to the
edge {u, v} by the edge weight function w1 is either (u) or
(v). Therefore, w1 (u, v)
min{ (u), (v)}. It follows
that w1 (u, v)
p(u, v).
Definition 5.3. An undirected graph G(V , E) is critically
2-E DGE -C ONNECTED if it satisfies both of the following
conditions. (i) G is 2-E DGE -C ONNECTED. (ii) For every
edge e E, the subgraph of G obtained by deleting the edge e
is not 2-E DGE -C ONNECTED.
We can now state and prove the analog of theorem 5.2 for
critically 2-edge connected graphs.
Lemma 5.5. If a graph G is critically 2-E DGE -C ONNECTED
 then no cycle of G has a chord.

We are now ready to complete the proof of lemma 5.1.
Proof of lemma 5.1. Starting from an optimal power assign-
ment to the instance I , construct the graph G1 (V , E1 ) as de-
scribed above. Since the graph Gc constructed in step 1 of
the heuristic (Figure 1) is a complete graph, every edge in
G1 is also in Gc . Consider the edge subgraph G1 (V , E1 ) of
Gc where E1 = E1 . Since G1 is 2-N ODE -C ONNECTED,
so is G1 . By lemma 5.4, for each edge {u, v} in E1 ,
p(u, v)  w1 (u, v). Therefore, the total weight W (G1 )
of all of the edges in G1 under the edge weight function p
Proof. The proof is by contradiction. Suppose G is crit-
ically 2-E DGE -C ONNECTED but there is a cycle C=
v1 , v2 , . . . , vr , with r
4, with a chord {vi , vj }. Con-
sider the graph G obtained from G by deleting the chord
{vi , vj }. We will show that G is 2-E DGE -C ONNECTED, thus
contradicting the assumption that G is critically 2-E DGE -
C ONNECTED.
To show that G is 2-E DGE -C ONNECTED, it suffices to
show that G cannot be disconnected by deleting any single
edge. Consider any edge {x, y} of G , and let G denote the
graph created by deleting {x, y} from G . Since we deleted
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS
only one edge from G , all the nodes of the cycle C are in
the same connected component of G . Thus, if we create the
graph G1 by adding the chord {vi , vj } to G , the two graphs
G1 and G have the same number of connected components.
However, G1 is also the graph obtained by deleting the edge
{x, y} from G. Since G is 2-E DGE -C ONNECTED, G1 is con-
nected. Thus, G is also connected. We therefore conclude
that G is 2-E DGE -C ONNECTED, and this contradiction com-
pletes the proof of lemma 5.5.  randomly generated networks and on networks derived from
The remainder of the proof to show that  (2 1/n)
is identical to that for the 2-Node-Connected case. With
 (2 1/n) and  2, the following theorem is a di-
rect consequence of theorem 5.1.
Theorem 5.4. There is a polynomial time approximation
algorithm with a performance guarantee of 8(1 1/n)
(which approaches 8 asymptotically from below) for the
U NDIR , 2-E DGE C ONNECTED, T OTAL P problem.
Our performance guarantee results are somewhat pes-
simistic since they are derived from a general framework.
Using a different method of analysis, Calinescu and Wan [4]
have shown recently that both of our heuristics provide a per-
formance guarantee of 4.
As in the case of minimizing maximum power, our general
framework for minimizing total power can also be used to
obtain polynomial time approximation algorithms for topol-
ogy control problems wherein the connectivity requirements
are specified using proper functions. To obtain this result,
we use the general method outlined in [1,12] as the algo-
rithm in step 2 of our general heuristic. The method of [1,12]
gives a 2-approximation algorithm for network design prob-
lems specified using proper functions. Using the notation of
theorem 5.1, = 2. It is also straightforward to show that the
complete graph constructed in step 1 of our heuristic has a re-
quired subgraph of weight at most the optimal solution value.
In other words,  1. Thus, we obtain a 4-approximation
algorithm for the general class of problems defined in [1,12].
An important example of a problem in this class is the Steiner
variant of connectivity, where the goal is to assign power lev-
els so as to connect only a specified subset of nodes of a graph
rather than all the nodes. An approximation
algorithm with a
performance guarantee of (1+ln 3 ) is known for the Steiner
tree problem in graphs [25]. Thus, using this approximation
algorithm, our approach yields a (2 + ln 3)-approximation for
the Steiner variant.
6. Experimental results
In the preceding section, we showed that our algorithm for
U NDIR , 2-N ODE C ONNECTED, T OTAL P provides a con-
stant factor approximation. In this section, we report on
the experimental performance of this algorithm. Since
27
there are no existing approximation algorithms3 specifically
for U NDIR , 2-N ODE C ONNECTED, T OTAL P, in the exper-
iments described here we compare the performance of our
algorithm with Ramanathan and Rosales-Hains algorithm
in [23]. Recall that their algorithm finds an optimum so-
lution for U NDIR , 2-N ODE C ONNECTED, M AX P in which
the power level of each node is minimal. Our experiments
were conducted using a customized implementation on both
realistic data generated by the TRANSIMS project [29].
6.1. Randomly generated networks
6.1.1. Experimental environment
The experimental setup used here is similar to the one de-
scribed in [23]. The radio wave propagation model used is
the Log-distance Path Loss Model:



Gt Gr 2 d
P L(d) = 10 log10 + 10 log ,
(4)2 d02
10
d0
where is the path loss exponent, d0 is the close-in reference
distance, is the radio wavelength, Gt is the transmitter an-
tenna gain, Gr is the receiver antenna gain, and d is the sepa-
ration distance between transmitter and receiver (see [24] for
detailed descriptions of these parameters). All of the parame-
ters are chosen to emulate a 2.4 GHz wireless radio, and if d
is less than a certain threshold, the transmission power is set
to the minimum transmission power of 1 dBm.
The experiments are conducted by varying the density of
the network and the spatial distribution of the nodes. In total
there are 38 sets of experiments, and 10 trials are run on each
set. Each of the results we cite is the average over the 10 trials.
The node density varies from 0.625 node/sq mile to
6.25 nodes/sq mile (10 nodes to 100 nodes) in a 4 mile by
4 mile area. The experiments are conducted using two node
distributions: one uniform and one skewed. Specifically, in
the uniformly distributed networks, all nodes are placed us-
ing a random uniform distribution. In the networks with a
skewed distribution, the network area is equally divided into
a 2 by 2 grid, with 80% of the nodes uniformly distributed
in two diagonal squares, and the other 20% of the nodes uni-
formly distributed in the other two diagonal squares.
In each experiment, after generating a placement of the
nodes, both our approximation algorithm (M IN T OTAL) and
the algorithm of [23] (M IN M AX) are run on the network
consisting of those nodes. Each algorithm assigns powers
to nodes such that the resulting network is 2-N ODE C ON -
NECTED . For each algorithm we measure both the maximum
and average power assigned, as well as the maximum and av-
erage degrees of the nodes in the resulting network.
Prior to discussing the results, we first provide figure 3 that
shows the actual topologies for one simulated network with
60 nodes. Figures 3(a) and (b) are respectively the topologies
3 While this paper was under review, an algorithm that provides a con-
stant factor approximation for the geometric version of U NDIR, 2-N ODE
C ONNECTED, T OTAL P was presented in [4].
28
Figure 3. Examples of network topologies.
resulting from our approximation algorithm (M IN T OTAL)
and Ramanathan and Rosales-Hains algorithm (M IN M AX).
6.1.2. Experimental results and discussion
In reporting our experimental results, we plot four different
quantities: (i) average power assigned to a node, (ii) maxi-
mum power assigned to any node, (iii) average degree of a
node and (iv) maximum degree of all of the nodes. The ex-
perimental results on power and node degree are shown in
figure 4.
In figures 4(a) and (c), Min Max AVG (Min Max
MAX) and Min Total AVG (Min Total MAX) are
the average (maximum) power using the M IN M AX and
M IN T OTAL algorithms, respectively.
In figures 4(b) and (d), Min Max AVG (Min Max
MAX) and Min Total AVG (Min Total MAX) are the
average (maximum) degrees using the M IN M AX and the
M IN T OTAL algorithms, respectively.
Figure 4 illustrates the results on power and node de-
gree. In the cases where nodes are uniformly distrib-
uted, our M IN T OTAL algorithm consistently outperforms the
M IN M AX algorithm in [23] in regard to average power by
5%19%. This improvement increases as the density of the
network increases. In contrast, the maximum power assigned
by our algorithm is 14%37% larger than that of [23]. The
average power is about 60%83% of the maximum power
using the M IN M AX algorithm, and about 39%70% using
our algorithm. Those numbers decrease as the density of the
network increases, which implies that the average power de-
creases faster than the maximum power, and a smaller per-
centage of nodes have the maximum power as the network
density increases.
In skewed placements of nodes, our M IN T OTAL algorithm
outperforms the M IN M AX algorithm with respect to aver-
age power by 6%25%. We observe that the difference be-
tween average power and maximum power is larger in skewed
placements than in uniform placements. The average power is
LLOYD ET AL.
about 40%76% of the maximum power using M IN M AX al-
gorithm of [23], and about 25%64% using our algorithm. In
other words, for a given average node density, the maximum
power in a skewed network is higher than that in a uniformly
distributed network, while the average power in the skewed
network is lower. The reason is that in a skewed network the
node density varies significantly from region to region. With
a larger number of nodes in a smaller area, the average dis-
tance between two nodes is less, hence the required power
levels are, on the average, smaller.
As a general rule, smaller is better in regard to node de-
grees in the network induced by the power assignments (e.g.,
increases spatial spectrum reuse). In that context, in the case
where nodes are uniformly distributed, the average (maxi-
mum) degree of the network with power assigned by our
M IN T OTAL algorithm is consistently smaller than the aver-
age (maximum) degree of the network with power assigned
by the M IN M AX algorithm in [23]. When using either of
the algorithms, the average degree does not vary much as the
network density changes. Specifically, the average degree is
around 2.73 using our algorithm, which is very close to the
smallest possible degree, since in a 2-node-connected graph,
the degree of each node must be at least 2.
The results in regard to node degrees under the skewed
node distribution are similar to those for the uniform case.
6.2. The TRANSIMS networks
In addition to the randomly generated networks, we also con-
ducted an experimental study on a more realistic network
obtained from the TRANSIMS Portland Study by the Los
Alamos National Laboratory [29]. This data set contains
locations of 1716 nodes over a 3 km by 3 km area. The
locations were generated by carrying out a detailed simula-
tion of the traffic in the Portland, OR, metropolitan area us-
ing the TRANSIMS simulation tool. Since the running time
of our algorithm would be prohibitively high if run on all
1716 nodes, we selected two characteristic areas of this net-
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 29

Figure 4. Experimental results.

Table 1 Table 2
Experimental results in area 1. Experimental results in area 2.

Max range Average range Max degree Average degree Max range Average range Max degree Average degree
M IN M AX 158 m 67.75 m 12 4.80 M IN M AX 153 m 73.59 m 28 7.94
M IN T OTAL 193 m 55.07 m 5 2.72 M IN T OTAL 222 m 51.95 m 7 2.73

work and conducted experiments on those two areas. By

works. So, if one utilizes the propagation model we used in
so doing, the spatial effects of the network are preserved
previous experiments, most nodes would use the minimum
and the experimental results can be obtained in a reason-
transmission power of 1 dbm. The results are presented in the
able time frame. Area 1 is a 1 km by 1 km square, where
following tables and figures.
284 nodes are somewhat uniformly distributed. Area 2 is a
600 meter by 1650 meter rectangle, where the majority of the
Tables 1 and 2 present the experimental results for area 1
271 nodes are concentrated along a curve and the others are
and area 2, respectively.
sparsely distributed over the remaining area. Similar to ran-
dom networks, for each area, we conducted two experiments: Figure 5(a) shows the entire network of 1716 nodes given
One uses our approximation algorithm (M IN T OTAL) for that every node has a 75 meter transmission range. The
U NDIR , 2-N ODE C ONNECTED, T OTAL P; the other uses two selected areas are highlighted.
Ramanathan and Rosales-Hains algorithm (M IN M AX) for Figures 5(b) and (c) illustrate the topologies of area 1 and
U NDIR , 2-N ODE C ONNECTED, M AX P. However, instead area 2, respectively, after using our algorithm (M IN T O -
of measuring transmission power, we measure the transmis- TAL ). Note that in figure 5(b), several nodes appear not to
sion range. That is because the nodes in the TRANSIMS data be 2-node-connected (e.g., the node at the top middle part
set are much more dense than our randomly generated net- of the figure). The reason is that three nodes that are on or
30
Figure 5. Topologies of the TRANSIMS network.
almost on a straight line, are connected to each other, and
the edges between them overlap in the figure.
Our experiments with TRANSIMS data show that topol-
ogy control can significantly reduce the average transmis-
sion power. In figure 5(a), where each node has a transmis-
sion range of 75 meters, the induced graphs in areas 1 and 2
are not even connected. After the application of M IN T O -
TAL algorithm, the induced graphs in both areas are 2-node-
connected and the average range for the two areas is reduced
to 55.07 meters and 51.95 meters, respectively.
In area 1, the average range assigned by our M IN T OTAL
algorithm is 18.7% lower than that assigned by the M IN M AX
algorithm, while the maximum transmission range of our al-
gorithm is 22.2% higher than the M IN M AX algorithm. The
induced maximum and average degrees are always smaller
using the M IN T OTAL algorithm than using the M IN M AX
algorithm. For area 2, our M IN T OTAL algorithm assigns av-
erage range 29.4% lower, but 45% higher maximum range.
The contrast on the induced maximum and average degrees
by using the two algorithms is even larger in area 2.
These results are consistent with the experimental results
on randomly generated networks. Our M IN T OTAL algorithm
constantly outperforms the M IN M AX algorithm on average
power (transmission range), and the margin is larger when
the network is more skewed.
LLOYD ET AL.
7. Directions for future research
Our work provides several directions for future research.
First, it will be of interest to investigate whether approxima-
tion algorithms with performance guarantees better than 4 can
be developed for inducing 2-node connected and 2-edge con-
nected graphs. Second, it will be useful to consider topology
control problems for other graph properties. In that direc-
tion, some complexity and approximation results for proper-
ties such as bounded diameter and lower bounds on node de-
grees under the objective of minimizing total power are pre-
sented in [17]. A third direction is to investigate the behav-
ior of topology control problems under the asymmetric power
threshold model. Some results in that direction are also pre-
sented in [17]. Finally, it will be of interest to develop distrib-
uted versions of algorithms for topology control problems.
References [4,18,23] present some results along that direc-
tion.
Acknowledgements
We thank the reviewers for a careful reading of the manuscript
and for providing valuable suggestions.
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS
Appendix A. Proof of lemma 4.2
We first restate the lemma.
Lemma 4.2. To find a power assignment such that the result-
ing undirected graph G is a tree is NP-complete.
By abuse of terminology, we use U NDIR , T REE,  to de-
note this problem. The NP-hardness of this problem is es-
tablished using a reduction from the X3C problem defined
below.
Exact cover by 3-sets (X3C).
Instance: A set S = {x1 , x2 , . . . , xn } of elements, where
n = 3r for some integer r; a collection C = {C1 , C2 , . . . ,
Cm } of subsets of S such that |Cj | = 3, 1  j  m.
Question: Does C contain an exact cover for S, that is, is
there a subcollection C  of C such that the sets in C  are
pairwise disjoint and their union is equal to S?
Note that whenever there is a solution to an instance of
X3C, the number of sets in the solution is exactly r (i.e. n/3).
Proof of lemma 4.2. In the U NDIR , T REE,  problem, we
are given a collection of nodes, and a (symmetric) power
threshold p(u, v) for each pair of nodes. The question is
whether there exists a power assignment such that the graph
induced by the power assignment is a tree.
It is easy to see that U NDIR , T REE,  is in NP since one
can guess a power assignment and verify in polynomial time
that the resulting graph is a tree. We prove the NP-hardness
of the problem by a reduction from X3C (defined above).
Given an instance I of X3C consisting of a set S with n
elements and a collection C of m subsets, we construct an
instance I  of the U NDIR , T REE,  problem as follows. The
node set V of I  contains a total of n + m + 1 nodes: There is
one node (called an element node) ui corresponding to each
element xi of S (thus, there are totally 3r element nodes),
one node (called a set node) vj corresponding to each set Cj
of C (thus, there are totally m set nodes), and a special node
(called the root node) denoted by R. The power thresholds
are chosen as follows. (The reader should bear in mind that
the power thresholds are symmetric; that is, for any pair of
nodes u and v, p(u, v) = p(v, u).)
p(R, vj ) = 1, 1  j  m,
p(ui , vj ) = 2, if xi Cj , 1  i  n, 1  j  m.
For all other pairs of nodes, the power thresholds are set
to 3. This completes the construction of the instance I  of
U NDIR , T REE, . It is easy to verify that the construction
can be carried out in polynomial time. We now argue that
there is a solution to the U NDIR , T REE,  instance if and
only if there is a solution to the X3C instance.
If. Suppose the X3C instance has a solution C  . We choose
the following power assignment: p (R) = 1, p (ui ) = 2 (1 
i  n), p (vj ) = 2 if Cj is in C  and p (vj ) = 1 otherwise
31
(1  j  m). It can be seen that the graph G resulting from
this power assignment contains only the following edges:
(a) The edge {R, vj }, for each j , 1  j  m.
(b) For each node vj whose corresponding set Cj is in C  ,
there are edges from vj to the three nodes corresponding
to the elements in Cj .
By choosing R as the root and using the fact that C  is an
exact cover, it can be verified that G is a tree: the root node R
is adjacent to each of the set nodes; and, each element node
appears as one of the three children of a set node correspond-
ing to a subset in the collection C  .
Only if. Now, suppose the U NDIR , T REE,  instance has
a solution. Let p (x) denote the power assigned to node x and
let G denote the graph induced by the power assignment.
We first observe that p (R)
1; otherwise, R would be
an isolated node and thus G cannot be a tree. Similarly,
p (vj )
1 for every set node vj and p (ui )
2 for every
element node ui . As a consequence, the root node R is adja-
cent to each of the set nodes v1 , v2 , . . . , vm , and the maximum
power assigned is at least 2. Therefore, there are two cases to
consider:
Case 1. The maximum power assigned is 2. Let X =
{vjk : p (vjk ) = 2}. We claim that the collection C  =
{Cjk : vjk X} is an exact cover for S. We prove this by first
showing that each element xi appears in some subset of C  .
To see this, we note that the graph G is connected (since it is
a tree). Thus, there is at least one edge from the element node
ui (corresponding to element xi ) to some other node of G.
Since the maximum power assigned to any node is 2 and the
power threshold for the element node ui to have an edge to R
or an edge to any other element node is 3, ui must be adjacent
to a set node vj . Further, because the threshold values are
symmetric, p (vj ) = 2. Thus, vj X and the correspond-
ing subset Cj is in C  . Hence, each element appears in some
subset in the collection C  .
We now show that the subsets in the collection C  are pair-
wise disjoint. Suppose some pair of subsets Ca and Cb in C 
have a common element xi . By our choice of C  , the power
values assigned to the corresponding set nodes va and vb are
both 2. Further, the power assigned to node ui is also 2. Thus,
in the graph G, ui is adjacent to both va and vb . As observed
earlier, the root node R is adjacent to both va and vb . Now,
the four edges {R, va }, {va , ui }, {ui , vb } and {vb , R} create a
cycle in G. This contradicts the assumption that G is a tree.
So, the subsets in C  are pairwise disjoint, and C  is indeed an
exact cover for S.
Case 2. The maximum power assigned is 3. First, note
that at most two nodes can have power 3, since if three nodes
have power 3, then they are mutually adjacent, and thus G is
not a tree.
Second, if the power assignment is as in the following
cases, we argue that there is an equivalent assignment in
which the maximum power is 2. These cases are: only one
node has power 3; R and one set node vi have power 3; and,
one element node ui and one set node vj have power 3 where
32
xi Cj . In any of these cases, the resulting graph G has
no edge with power threshold 3, so an assignment with max-
imum power 2 can be obtained by reducing the power level
of the nodes with power 3 while keeping the assignments to
all of the other nodes unchanged. The induced graph does not
change. Thus, the new assignment is a solution with maxi-
mum power 2 to the instance of U NDIR , T REE, . Follow-
ing the argument in case 1, a solution to X3C can be con-
structed.
Finally, we claim that there are no such valid power as-
signments in the remaining cases (i.e. R and ui have power 3;
vi and vj have power 3; ui and uj have power 3; or, ui and vj
have power 3 where xi / Cj ). The reasons are the following:
1. If two set nodes vi and vj have power 3, then the edges
{R, vi }, {R, vj } and {vi , vj } form a cycle.
2. If the root node R and one element node ui have po-
wer 3, the edge {R, ui } is in G. Therefore, edge{ui , vj },
1  j  m, is not in G, otherwise R, ui , and vj form a
cycle. Recall that p (ui )
2 for every element node ui ,
therefore each vj with power 2 is adjacent to exactly 3 el-
ement nodes. No two set nodes can be adjacent to the
same element node, otherwise those three nodes and R
form a cycle. Hence, totally 3k (where k is the number
of set nodes with power 2) element nodes are adjacent to
some set node. Further, no two element nodes can be ad-
jacent to each other since the power thresholds between
such nodes are 3. Thus, there are 3k + 1 element nodes.
This is a contradiction since we know in this instance of
U NDIR , T REE, , the number of element nodes is a mul-
tiple of 3.
3. If two element nodes ui and uj have power 3, the edge
{ui , uj } is in G. Recall that all set nodes must be adjacent
to R, so one and only one of ui and uj is adjacent to a
set node. Suppose it is ui . We know from above that 3k
element nodes are adjacent to some set node. So, together
with uj , there are 3k + 1 element nodes a contradiction.
4. If one element node ui and one set node vj have power 3,
where xi / Cj , then ui is adjacent to vj . Therefore, there
are 4 nodes adjacent to vj , which are ui and three element
nodes whose corresponding elements are in set Cj . Hence,
there are totally 3k+1 element nodes a contradiction.
This completes the proof of the case 2 as well as that of
lemma 4.2.  Hence, node vi is only one hop away from w.
Appendix B. Proof of theorem 4.4
Proof. We use a reduction from S ET COVERING (S C), a
well-known NP-complete problem [11].
Set covering (Sc).
Instance: A set S = {x1 , x2 , . . . , xn }, a collection C =
{C1 , C2 , . . . , Cm }, where Ci is a subset of S (1  i  m),
and a positive integer K  m.
LLOYD ET AL.
Question: Does there exist a subcollection C  C, such
that |C  |  K and the union of the sets in C  is equal to S?
Let P be the property THE DIAMETER OF G IS LESS
THAN OR EQUAL TO 6. This property implies that in G,
each node is at most 6 hops away from any other node. Ob-
viously, P is monotone, and can be tested in O(n3 ) time by
using the FloydWarshall algorithm, where n is the number
of nodes in the graph [9]. Thus, Max-power users is in NP. To
prove the NP-hardness, we provide a reduction from S C.
Given an instance I of S C, an instance I  of Max-power
users is constructed as follows: For each element xi of S,
create a node ui in V and for each Ci of C, create a node
vi in V . Further, V also contains four special nodes: w, s1 ,
s2 , s3 . The power threshold function p is defined as follows.
(It should be noted that the power thresholds are symmetric.)
p(ui , vj ) = 1, if xi Cj ,
p(w, vj ) = P , 1  j  m,
p(w, s1 ) = p(s1 , s2 ) = p(s2 , s3 ) = 1.
For all other pairs of nodes x and y, p(x, y) = P + 1.
The value of M is set to K + 1. This completes the con-
struction of an instance I  of Max-power users. It is clear that
the construction can be done in polynomial time. Now, we
show that there is a solution to the Max-power users instance
if and only if there is a solution to S C.
If. Suppose C  is a solution to the instance of S C. We
construct a power assignment p as follows.
p (w) = P ,
p (vi ) = P , if Ci C 
(Note: there are at most K such nodes),
p (x) = 1 for any other node x.
We now argue that p is a solution to the instance of Max-
power users. Obviously, the maximum power assigned is P
and at most M (i.e., K + 1) nodes have power P . To establish
that the resulting graph G(V , E) satisfies P (i.e., the graph has
diameter at most 6), we show that w is within 3 hops of every
other node. This follows from the following observations.
1. Nodes s1 , s2 , and s3 are respectively 1, 2, and 3 hops away
from w.
2. For any Ci C, if Ci C  , then the edge {vi , w} E.
3. For any xi S, node ui is 2 hops away from w, since ui
is adjacent to some node vi that has an edge to w. (Other-
wise, C  does not cover the element xi .)
4. For any Ci C, if Ci / C  , then vi is 3 hops away from w,
since vi is adjacent to some uj .
Only if. Suppose we have a power assignment p that
is a solution to the instance of Max-lpower users, and that
G(V , E) is the resulting graph. We construct a solution to the
S C instance as follows. If there is an edge between w and
vi in E (there are at most M 1 such edges), then include
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS
set Ci in C  . We claim that C  is a solution to the instance of
S C. Since |C  |  M 1 = K, we need only show that C 
covers S. Since the diameter of G(V , E)  6, s3 is at most
6 hops away from any other node. It follows that w must be
within 3 hops of every other node. For each vi , if edge {vi , w}
E, vi is one hop away from w. However, if edge {vi , w} / E. Kranakis, Montreal, Canada (October 2003), Lecture Notes in Com-
E, vi is at least 3 hops away from w. Now suppose there is
an element xi S that is not in any set of C  . Then, ui is not
adjacent to any node vj that is one hop away from w. Thus,
ui must be adjacent to some node vj that is at least 3 hops
away from w. Thus, node ui is at least 4 hops away from w
a contradiction. This completes the proof of theorem 4.4. 
References
[1] A. Agrawal, P. Klein and R. Ravi, When trees collide: An approxima-
tion algorithm for the generalized Steiner problem on networks, SIAM
J. Computing 24(3) (1995) 440456.
[2] D.M. Blough, M. Leoncini, G. Resta and P. Santi, On the symmetric
range assignment problem in wireless ad hoc networks, in: Proc. of the
2nd IFIP Internat. Conf. on Theoretical Computer Science, Montreal
(August 2002) pp. 7182.
[3] G. Calinescu, I. Mandoiu and A. Zelikovsky, Symmetric connectivity
with minimum power consumption in radio networks, in: Proc. of the
2nd IFIP Internat. Conf. on Theoretical Computer Science, Montreal
(August 2002) pp. 119130.
[4] G. Calinescu and P.-J. Wan, Symmetric high connectivity with min-
imum total power consumption in multihop packet radio networks,
in: Proc. of the Internat. Conf. on Ad Hoc and Wireless Networks
(ADHOC-NOW03), eds. S. Pierre, M. Barbeau and E. Kranakis, Mon-
treal, Canada (October 2003) Lecture Notes in Computer Science,
Vol. 2865 (Springer, New York, 2003) pp. 235246.
[5] W. Chen and N. Huang, The strongly connecting problem on multihop
packet radio networks, IEEE Trans. Commun. 37(3) (1989) 293295.
[6] X. Cheng, B. Narahari, R. Simha, M. Cheng and D. Liu, Strong min-
imum energy topology in wireless sensor networks: NP-completeness
and heuristics, IEEE Trans. Mobile Computing 2(3) (2003) 248256.
[7] A.E.F. Clementi, P. Penna and R. Silvestri, Hardness results for the
power range assignment problem in packet radio networks, in: Proc.
of the 3rd Internat. Workshop on Randomization and Approximation
in Computer Science (APPROX 1999) (July 1999), Lecture Notes in
Computer Science, Vol. 1671 (Springer, New York, 1999) pp. 195208.
[8] A.E.F. Clementi, P. Penna and R. Silvestri, The power range assign-
ment problem in packet radio networks in the plane, in: Proc. of the
17th Annual Symposium on Theoretical Aspects of Computer Science
(STACS 2000) (February 2000) pp. 651660.
[9] T. Cormen, C. Leiserson, R. Rivest and C. Stein, Introduction to Algo-
rithms, 2nd ed. (MIT Press/McGraw-Hill, Cambridge, MA, 2001).
[10] G.A. Dirac, Minimally 2-connected graphs, J. Reine Angewandte
Math. 228 (1967) 204216.
[11] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide
to the Theory of NP-Completeness (Freeman, San Francisco, CA,
1979).
[12] M. Goemans and D.P. Williamson, A general approximation technique
for constrained forest problems, SIAM J. Computing 24(2) (1995) 296
317.
[13] L. Hu, Topology control for multi-hop packet radio networks, IEEE
Trans. Commun. 41(10) (1993) 14741481.
[14] S. Khuller and B. Raghavachari, Improved approximation algorithms
for uniform connectivity problems, J. Algorithms 21 (1996) 434450.
[15] S. Khuller and U. Vishkin, Biconnectivity approximations and graph
carvings, J. ACM 41 (1994) 214235.
[16] L.M. Kirousis, E. Kranakis, D. Krizanc and A. Pelc, Power consump-
tion in packet radio networks, in: Proc. of the 14th Annual Sympo-
sium on Theoretical Aspects of Computer Science (STACS 97) (Feb-
33
ruary 1997) Lecture Notes in Computer Science, Vol. 1200 (Springer,
New York, 1997) pp. 363374.
[17] S.O. Krumke, R. Liu, E.L. Lloyd, M.V. Marathe, R. Ramanathan and
S.S. Ravi, Topology control problems under symmetric and asymmet-
ric power thresholds, in: Proc. of the Internat. Conf. on Ad Hoc and
Wireless Networks (ADHOC-NOW03), eds. S. Pierre, M. Barbeau and
puter Science, Vol. 2865 (Springer, New York, 2000) pp. 187198.
[18] M. Kubisch, H. Karl, A. Wolisz, L. Zhong and J. Rabaey, Distributed
algorithms for transmission power control in wireless sensor networks,
in: Proc. of the IEEE Wireless Communications and Networking Con-
ference (WCNC 2003), New Orleans, LA (March 2003) pp. 558563.
[19] L. Li and J.Y. Halpern, Minimum energy mobile wireless networks
revisited, in: Proc. of the IEEE Conf. on Communications (ICC01)
(June 2001) pp. 278283.
[20] L. Li, J.Y. Halpern, P. Bahl, Y. Wang and R. Wattenhofer, Analysis of
cone-based distributed topology control algorithm for wireless multi-
hop networks, in: Proc. of the ACM Principles of Distributed Comput-
ing Conference (PODC01) (August 2001) pp. 264273.
[21] M.D. Plummer, On minimal blocks, Trans. AMS 134 (October
December 1968) 8594.
[22] V. Radoplu and T.H. Meng, Minimum energy mobile wireless net-
works, IEEE J. Selected Areas Commun. 17(8) (1999) 13331344.
[23] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
wireless networks using transmit power adjustment, in: Proc. of the
IEEE INFOCOM 2000, Tel Aviv, Israel (March 2000) pp. 404413.
[24] T.S. Rappaport, Wireless Communications: Principles and Practice
(Prentice-Hall, Englewood Cliffs, NJ, 1996).
[25] G. Robins and A. Zelikovsky, Improved Steiner tree approximation in
graphs, in: Proc. of the 11th Ann. ACMSIAM Symposium on Discrete
Algorithms (SODA 2000), San Francisco, CA (2000) pp. 770779.
[26] E.M. Royer, P. Melliar-Smith and L. Moser, An analysis of the opti-
mum node density for ad hoc mobile networks, in: Proc. of the IEEE
Internat. Conf. on Communication (ICC01), Helsinki, Finland (June
2001) pp. 857861.
[27] E.M. Royer and C. Perkins, Transmission range effects on AODV mul-
ticast communication, ACM Mobile Networks Appl. (Special Issue on
Multipoint Communication in Wireless Networks) 7(6) (2002) 455
470.
[28] H. Takagi and L. Kleinrock, Optimal transmission ranges for randomly
distributed packet radio terminals, IEEE Trans. Commun. 32(3) (1984)
246257; also see in: Multiple Access Communications, Foundations
for Emerging Technologies, ed. N. Abramson (IEEE Press, New York,
1992) pp. 342353.
[29] TRANSIMS, http://transims.tsasa.lanl.gov/.
[30] J. van Leeuwen, Graph algorithms, in: Handbook of Theoretical Com-
puter Science, Vol. A, ed. J. van Leeuwen (MIT Press/Elsevier, Cam-
bridge, MA, 1990) chapter 10.
[31] R. Wattenhofer, L. Li, P. Bahl and Y. Wang, Distributed topology con-
trol for power efficient operation in multihop wireless ad hoc networks,
in: Proc. of the IEEE INFOCOM 2001, Anchorage, Alaska (April
2001) pp. 13881397.
[32] D.B. West, Introduction to Graph Theory (Prentice-Hall, Englewood
Cliffs, NJ, 1996).
Errol L. Lloyd is a Professor of Computer and Infor-
mation Sciences at the University of Delaware. Pre-
viously he served as a faculty member at the Univer-
sity of Pittsburgh and as Program Director for Com-
puter and Computation Theory at the National Sci-
ence Foundation. From 1994 to 1999 he was Chair
of the Department of Computer and Information Sci-
ences at the University of Delaware. Concurrently,
from 1997 to 1999 he was Interim Director of the
University of Delaware Center for Applied Science
and Engineering in Rehabilitation. Professor Lloyd received undergraduate
degrees in both computer science and mathematics from Penn State Uni-
34
versity, and a Ph.D. in computer science from the Massachusetts Institute
of Technology. His research expertise is in the design and analysis of al-
gorithms, with a particular concentration on approximation algorithms for
computationally difficult problems. He has published over thirty journal pa-
pers and numerous conference papers. In 1989 Professor Lloyd received an
NSF Outstanding Performance Award, and in 1994 he received the Univer-
sity of Delaware Faculty Excellence in Teaching Award.
E-mail: elloyd@cis.udel.edu
Rui Liu received the BS degree in mathematics from
Peking University, Beijing, China, in 1998; the MS
degree in applied mathematics from University of
Delaware in 2000. He is a doctoral candidate in
computer and information sciences at the University
of Delaware. His research interests include design
and analysis of algorithms for combinatorial opti-
mization problems, parallel and distributed comput-
ing and computer networks.
E-mail: ruliu@cis.udel.edu
Madhav V. Marathe received his B. Tech in com-
puter science from IIT Madras and his Ph.D. in
computer science from the University at Albany
SUNY in 1994. Since that time, he has been with
Los Alamos National Laboratory, where he currently
leads the Mathematics and Computer Science team
in the Basic and Applied Simulation Science Group.
His research interests include mobile computing,
simulation of large socio-technical systems such as
transportation, telecommunication and markets and
design and analysis of algorithms.
E-mail: marathe@lanl.gov
LLOYD ET AL.
Ram Ramanathan is a Division Scientist at BBN
Technologies. His research interests are in the area
of wireless and ad hoc networks, in particular, rout-
ing, medium access control and directional anten-
nas. He is currently the principal investigator for a
project on architecture and protocols for opportunis-
tic access of spectrum using cognitive radios. Re-
cently, he was one of one of two principal inves-
tigators for the DARPA project UDAAN (Utilizing
Directional Antennas for Ad hoc Networking), and
the co-investigator on NASAs Distributed Spacecraft Network project. Ram
is actively involved in the evolution of mobile ad hoc networking, and has
recently served on the program committees of the ACM MobiHoc sympo-
sium, and ACM Mobicom. He is on the editorial board of Ad Hoc Networks
journal. He has won three best paper awards at international conferences
at ACM SIGCOMM 92, at IEEE INFOCOM 96, and IEEE MILCOM 02.
Dr. Ramanathan holds a B.Tech from the Indian Institute of Technology,
Madras, and an M.S. and a Ph.D. from the University of Delaware. He is
a senior member of the IEEE.
E-mail: ramanath@bbn.com
S.S. Ravi received his Ph.D. in computer science
from the University of Pittsburgh in 1984. Since that
time, he has been with the Department of Computer
Science, University at Albany State University of
New York, where he is currently a Professor. His re-
search interests include design and analysis of algo-
rithms, mobile computing, fault-tolerance and VLSI
design and testing.
E-mail: ravi@cs.albany.edu
 Mobile Networks and Applications 10, 3545, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Wireless ATM Layouts for Chain Networks ,

MICHELE FLAMMINI
Dipartimento di Informatica, University of LAquila, Via Vetoio loc. Coppito, I-67100 LAquila, Italy

GIORGIO GAMBOSI
Dipartimento di Matematica, University of Rome Tor Vergata, Via della Ricerca Scientifica, I-00133 Rome, Italy

ALFREDO NAVARRA
Dipartimento di Informatica, University of LAquila, Via Vetoio loc. Coppito, I-67100 LAquila, Italy, and
MASCOTTE project, I3S-CNRS, INRIA, Universit de Nice, Sophia Antipolis, route des Lucioles, B.P. 93 F-06902, Sophia Antipolis Cedex, France

Abstract. In this paper we consider the problem of constructing ATM layouts for wireless networks in which mobile users can move along
a chain of base stations. We first show that deciding the existence of a layout with maximum hop count h, load l and channel distance d
is NP-complete for every fixed value of d greater or equal to 1. We then provide optimal layout constructions for the case d
2. Finally,
optimal layout constructions are obtained also for any d within the class of the so-called canonic layouts, that so far have always been shown
to be the optimal ones.
Keywords: capacity planning, ATM networks, wireless networks, mobile users, chains

1. Introduction VPs belonging to a VC, which represents the number of VP

The Asynchronous Transfer Mode (ATM for short) is the
most popular networking paradigm for Broadband ISDN
[18,19,24]. It transfers data in the form of small fixed-size
cells, and in order to achieve the stringent transfer rate re-
quirements, is based on two types of predetermined routes in
the network: virtual paths or VPs, constituted by a sequence
of successive edges or physical links, and virtual channels or
VCs, each given by the concatenation of a proper sequence
of VPs. Routing in virtual paths can be performed very effi-
ciently by dedicated hardware, while a cell passing from one
virtual path to another one requires more complex and slower
elaboration.
Given a network and a set of connections to be estab-
lished, to provide the performance required by B-ISDN ap-
plications it is important that routing is performed in a hard-
ware fashion in most of the nodes a cell traverses, at the same
time limiting the number of paths sharing a same physical
link [1,4,15,25,26].
A graph theoretical model related to this ATM design prob-
lem has been first proposed in [7,15]. In such a framework,
the VP layouts determined by the VPs constructed on the net-
work are evaluated mainly with respect to two different cost
measures: the hop count, that is the maximum number of
Work supported by the IST Programme of the EU under contract num-
ber IST-1999-14186 (ALCOM-FT), by the EU RTN project ARACNE,
by the Italian project REAL-WINE, partially funded by the Italian Min-
istry of Education, University and Research, by the French MASCOTTE
project I3S-CNRS/INRIA/Univ. Nice, Sophia Antipolis and by the Ital-
ian CNR project CNRG003EF8 Algoritmi per Wireless Networks
(AL-WINE).
Preliminary version of this paper appeared in [11].
Corresponding author.
changes of messages along their route to the destination, and
the load, given by the maximum number of virtual paths shar-
ing an edge, that determines the size of the VP routing tables
(see, e.g., [8]). For further details and technical justifications
of the model for ATM networks see for instance [1,15].
While the problem of determining VP layouts with
bounded hop count and load is NP-hard under different as-
sumptions [10,15], many optimal and near optimal construc-
tions have been given for various interconnection networks
such as chain, trees, grids and so forth [3,7,9,13,14,21,29]
(see [30] for a survey).
The integration of wireless and ATM networks is emerging
as one of the most promising approaches able to support users
mobility while maintaining the quality of service offered by
the classical ATM. This combination occurs at different lev-
els and yields different scenarios, such as End-to-End WATM
and WATM Interworking, applied respectively to create new
wireless networks with ATM virtual channels extending until
the mobile terminals and at a more external level for intercon-
necting different existing wireless subnets [16]. In both sce-
narios, the mobility facility requires the efficient solution of
several problems, such as handover (users movement), rout-
ing, location management, connection control and so forth.
A detailed discussion of these and other related issues can be
found in [2,5,6,16,23,27].
An extension of the basic ATM model of [7,15] able to
combine quality of service and mobility aspects in wireless
ATM networks has been proposed in [12]. In this model a
subset of the nodes of the network represents the base sta-
tions and users are allowed to move between them according
to an adjacency graph expressing their adjacencies in the ge-
ographic space. Such a graph, in general, can differ from
36
the physical topology of the infrastructured network. For in-
stance, in nowadays cellular systems like GSM, the physical
graph G is a tree, stations correspond to its leaves and the ad-
jacency graph is an hexagonal grid (see, for instance, [22]).
Standard ATM layouts must be constructed in order to estab-
lish a different VC for each station, but their performance is
evaluated by means of a further parameter, the virtual channel
distance, that measures the time needed to reconstruct virtual
channels during handover phases, that is when mobile termi-
nals switch between adjacent stations. More precisely, the
distance between the virtual channels of two adjacent nodes
is equal to the number of VPs that must be deleted and added
to one VC in order to obtain the other one. In order to make
the rerouting phase imperceptible to users and thus to obtain
a sufficient quality of service, the maximum distance between
two virtual channels must be maintained as low as possible.
Therefore, a natural combinatorial problem arises in which
suitable trade-offs must be determined between the different
performance measures.
In [12] it has been shown that the layout construction prob-
lem is intractable, that is NP-hard. Moreover, optimal lay-
out constructions are given when the physical and adjacency
graphs are coincident and correspond to basic interconnection
networks, such as chains and rings. Such results hold under
the assumption that all the VCs induce shortest paths in the
underlying network.
In this paper we consider the determination of WATM lay-
outs for chains in the non-shortest path case in which the
lengths of the paths induced by the VPs is not constrained.
We first show that deciding the existence of a layout with
maximum hop count h, load l = 1 and distance d = 1 is
NP-complete even when the adjacency graph is a chain of
base stations with the source coinciding with one of its end-
points. Moreover, such a hardness result is extended to every
fixed value of d. We then consider the case in which the phys-
ical and adjacency graph coincide with chains and provide
optimal layout constructions for d
2. Finally, optimal lay-
out constructions are obtained also for any d within the class
of the so-called canonic layouts, that so far have been always
shown to be the optimal ones.
The paper is organized as follows. In the next section we
introduce the model, the notation and the necessary defini-
tions. In section 3 we provide the above mentioned hardness
results for the layout construction problem. In section 4 we
provide the optimal layouts for chains when d = 2 and in
section 5 the optimal canonic ones for any d. Finally, in sec-
tion 6, we give some conclusive remarks and discuss some
open questions.
2. The WATM model
We model the network as an undirected graph G = (V , E),
where nodes in V represent switches and edges in E are point-
to-point communication links. In G there exists a subset of
nodes U V constituted by base stations, i.e., switches
adapted to support mobility and having the additional capa-
FLAMMINI ET AL.
bility of establishing wireless connections with mobile termi-
nals. A distinguished source node s V provides high speed
services to the users moving along the network. We observe
that, according to the wireless nature of the system, during
the handover phase mobile terminals do not necessarily have
to move along the network G, but they can switch directly
from one station to another, provided that they are adjacent in
the physical space. It is thus possible to define a (connected)
adjacency graph A = (U, F ), whose edges in F represent
adjacencies between stations.
A layout
for G = (V , E) with source s V is a col-
lection of simple paths in G, termed virtual paths (VPs for
short), and a mapping that defines, for each station u U ,
a unique virtual channel VC(u) connecting s to u, i.e., a sim-
ple path from s to u in the virtual topology defined by the VPs
of
. In other words, VC(u) is a collection of VPs whose
concatenation forms a path in G from s to u.
Definition 2.1 [15]. The hop count h(u) of a node u U in
a layout
is the number of VPs contained in VC(u), that
is |VC(u)|. The maximal hop count of
is Hmax (
)
maxuU {h(u)}.
Definition 2.2 [15]. The load l(e) of an edge e E in a
layout
is the number of VPs
that include e. The
maximal load Lmax (
) of
is maxeE {l(e)}.
As already observed, when passing from a station u U
to an adjacent one v U , the virtual channel VC(v) must
be reconstructed from VC(u) changing only a limited number
of VPs. Once fixed VC(u) and VC(v), denoted as VC(u, v)
the set of VPs in the subchannel corresponding to the longest
common prefix of VC(u) and VC(v), this requires the deletion
of all the VPs of VC(u) that occur after VC(u, v), plus the
addition of all the VPs of VC(v) after VC(u, v). The number
of removed and added VPs, denoted as D(VC(u), VC(v)), is
called the distance of VC(u) and VC(v) and naturally defines
a channel distance measure d between pairs of adjacent nodes
in A.
Definition 2.3 [12]. The channel distance of two nodes u
and v, such that, {u, v} F (i.e., adjacent in A) is d(u, v) =
D(VC(u), VC(v)) = h(u)+h(v)2|VC(u, v)|. The maximal
distance of
is Dmax (
) max{u,v}F {d(u, v)}.
It is now possible to give the following definition concern-
ing layouts for WATM networks.
Definition 2.4. A layout
with Hmax (
)
h, Lmax (
)

l and Dmax (
)
d is a h, l, d-layout for G, s and A.
In the following we will always assume that all the VPs
of
are contained in at least one VC. In fact, if such prop-
erty does not hold, the not used VPs can be simply removed
without increasing the performance measures h, l and d.
Before concluding the section, let us remark that for prac-
tical purposes and quality of services guarantees, it makes
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS
sense to consider the case where d  h. In fact, while a
little communication delay proportional to the hop count in
general can be tolerated, connections gaps due to rerouting of
virtual channels must not be appreciated by mobile users. On
the other hand, when d  2h, our model coincides with the
classical one presented in [15] for standard ATM networks,
since the difference between any two virtual channels is al-
ways at most equal to 2h.
3. Hardness and approximation results
In this section we show that constructing optimal dynamic
layouts is in general an NP-hard problem, even when l = 1,
d = 1 and the adjacency graph is a chain of stations with the
source being one of its endpoints.
Before proving our results, let us briefly outline the basic
characteristics of a layout with maximum delay d = 1.
Given any two stations u1 , u2 U adjacent in A =
(U, F ), during an handover from u1 to u2 if d = 1 by de-
finition only one VP can be modified. This means that either
VC(u2 ) is a prefix of VC(u2 ) and thus VC(u2 ) is obtained
from VC(u1 ) adding a new VP from u1 to u2 , or vice versa.
In any case, a VP between u1 and u2 must be contained in the
layout. As a direct consequence, the virtual topology defined
by the VPs of
coincides with the adjacency graph A. More-
over, A must be acyclic. In fact, moving from a station in one
direction along a cycle it is not possible to rebuild the virtual
channel of the station itself when it is reached twice. Finally,
if the source coincides with a base station, the maximum hop
count of
is the eccentricity of s in A, that is, the maximum
distance in A between s and the other stations.
We are now ready to prove our first hardness result.
Theorem 3.1. Given a network G = (V , E), a source s V ,
a chain adjacency graph A = (U, F ) and a positive inte-
ger h, deciding the existence of a h, 1, 1-layout for G with
source s is an NP-complete problem.
Proof. First of all, observe that, for any h, l, d, the prob-
lem of deciding the existence of a h, l, d-layout is in NP, as
given G = (V , E), s V , A = (U, F ) and a layout
, it is
possible to check in polynomial time whether Hmax (
)
h,
Lmax (
)
l and Dmax (
)
d.
We prove the claim by providing a polynomial time re-
duction from Disjoint Paths problem (DP), known to be
NP-complete [20]. An instance of this problem is consti-
tuted by a graph G = (V , E) and a collection of node pairs
{(s1 , t1 ), . . . , (sk , tk )}. We want to determine whether there
exist k edge-disjoint paths in G, each connecting a different
pair (s1 , t1 ), 1
i
k.
Without loss of generality, it is possible to assume that
all the pairs (si , ti ), 1
i
k, are disjoint, i.e., all
nodes s1 , . . . , sk , t1 , . . . , tk are different. In fact, any instance
not satisfying this property can be trivially modified into an
equivalent one in which every node v occurring in k 
k
pairs is connected in G to k  new nodes v1 , . . . , vk  and the
k  pairs contain in the order v1 , . . . , vk  instead of v.
37
Starting from an instance of DP, we construct a network
G = (V  , E  ), a source s V  and a chain adjacency graph
A = (U, F ) that admit a h, 1, 1-layout with h = 2k 1 if
and only if there exist the requested k edge-disjoint paths in
the instance of DP.
Let G = (V  , E  ) be such that, given k 1 nodes
w1 , . . . , wk1 not contained in the initial graph G, V  = V
{w1 , . . . , wk1 } and E  = E{{ti , wi }{wi , si+1 } | 1
i < k}.
Concerning A = (U, F ), let U = {s1 , . . . , sk , t1 , . . . , tk } and
F = {{si , ti } | 1
i
k} {{ti , si+1 } | 1
i < k}. Finally,
the source s = s1 .
Assume first that there is a 2k 1, 1, 1-layout
for
G = (V  , E  ), s and A = (U, F ). By the considerations
at the beginning of this section, for each e F , a VP in

must exist connecting the two endpoints of e. We can assume
that for each i, 1
i < k, the VP connecting ti to si+1 is
ti , wi , si+1 , i.e., it is constituted by the new added path in
G that goes from ti to si+1 through the new node wi . In
fact, if this does not hold, it is possible to add to
the new
VP ti , wi , si+1 , deleting the old one and then, in order to
keep l = 1, if there is another VP stepping through wi , it is
modified in such a way that its subpath between ti and si+1
coincides with the old deleted VP.
Therefore, since l = 1 and for all i, 1
i
k, the VP
between si and ti does not step through any of the nodes
w1 , . . . , wk1 , there must exist k edge-disjoint paths in G
connecting the pairs (s1 , t1 ), . . . , (sk , tk ).
Vice versa, if there are k edge-disjoint paths in G connect-
ing the pairs (s1 , t1 ), . . . , (sk , tk ), a 2k 1, 1, 1-layout
for
G = (V  , E  ), s and A = (U, F ) can be constructed as fol-
lows. For each i, 1
i
k, the VP between si and ti is given
by the corresponding path in G, edge-disjoint with all the oth-
ers. The VP between ti and si+1 , 1
i < k, is ti , wi , si+1 .
Since s = s1 and the eccentricity in A = (U, F ) of the sta-
tion s1 is 2k 1, the layout
thus constructed gives directly
a h, 1, 1-layout with h = 2k 1. 
Notice that in the above construction the source s corre-
sponds to an endpoint of the chain A = (U, F ), so as already
remarked the NP-completeness holds also under this restric-
tion.
The above result can generalized to any fixed d > 0 as
follows.
Theorem 3.2. For any fixed integer d > 0, given a network
G = (V , E), a source s V , a chain adjacency graph
A = (U, F ) and a positive integer h, deciding the existence
of a h, 1, d-layout for G with source s is an NP-complete
problem.
Proof. Given G, a source s and an adjacency graph A =
(U, F ), it is sufficient to construct in polynomial time G , s 
and A = (U  , F  ), such that, G, s, A admit a h, 1, 1-layout
if and only if G , s  , A admit a h , 1, d-layout for a suitable
h > 0.
By theorem 3.1, it is possible to assume that A = (U, F )
is a chain of the nodes u1 , . . . , uk and s = u1 .
38
G = (V  , E  ) is obtained from G = (V , E) by adding
for each ui , 1
i < k, d 1 other stations ui,1 , . . . , ui,d1
connected by edges {ui , ui,1 } and {ui,j , ui,j +1 }, 1
j <
d 1, in such a way that ui , ui,1 , . . . , ui,d1 form a chain of
d nodes. s  = u1 and the new adjacency graph A = (U  , F  )
is, such that, U  = U {ui,1 , . . . , ui,d1 | 1
i
k} and
F  = {{ui , ui,1 | 1
i
d 1} {{ui,j , ui,j +1 | 1
i < k,
1
j < d 1} {{ui,d1 , ui+1 | 1
i < k}. Hence,
A = (U, F ) is a chain.
Since l = 1, in any layout
 for G , s  , A , all edges edges
{ui , ui,1 } and {ui,j , ui,j +1 }, 1
i < k and 1
j < d 1,
must be VPs, as they are the only simple paths connecting
the respective endpoint stations. Then, during the handover
from a station ui,d1 to ui+1 , the d 1 VPs {ui , ui,1 } and
{ui,j , ui,j +1 }, 1
j < d 1, must be deleted and then a
single VP must be added from ui (the last station in the com-
mon prefix of the virtual channels VC(ui,d1 ) and VC(ui+1 ))
to ui+1 .
Since s  = u1 and u1 has eccentricity h = max{h, k 2 +
d 2} in A = (U  , F  ), then G, s, A admit a h, 1, 1-layout
if and only if G , s  , A admit a h , 1, d-layout, hence the
theorem holds.  the adjacency graph A coincides with Cn . Moreover, without
Again, the NP-completeness still holds if the source s is an
endpoint of the chain adjacency graph.
Before concluding the section, let us, finally, show that for
d = 1 a stronger hardness result holds. To this aim observe
first that as remarked at the beginning of this section, the vir-
tual topology induced by any h, l, 1-layout coincides with
the adjacency graph. Moreover, if the source coincides with a
base station, h is equal to the eccentricity in A of s, otherwise
connecting s by a VP to a node of minimum eccentricity in A
it is possible to obtain a layout with a maximum hop count
equal to its eccentricity increasing the load at most of one.
Therefore, as far as approximation results are concerned, the
interesting parameter to be approximated remains the maxi-
mum load.
The problem of minimizing the maximum load is equiv-
alent from an approximation point of view to the optimiza-
tion version of the decision problem DP in which we want to
determine k paths connecting the k sourcedestination pairs
(s1 , t1 ), . . . , (sk , tk ) in such a way as to minimize the maxi-
mum number of paths sharing a same edge.
In fact, any r-approximation algorithm A for the layout
problem yields directly a O(r)-approximation algorithm ADP
for DP. Informally, ADP simply consists in running A on the
instance of the layout problem obtained by adding a new
source s, connecting s to each si , 1
i
k, and letting
A = (U, F ) be, such that, U = {s, s1 , . . . , sk , t1 , . . . , tk } and
F = {{s, si } | 1
i
k} {{si , ti } | 1
i
k}. The
k VPs connecting each si to ti , 1
i
k, in the layout
returned by A correspond to an O(r)-approximate solution
for DP. A reverse reduction can be determined by observing
that an r-approximation algorithm ADP for DP yields directly
an O(r)-approximation algorithm A for the layout construc-
tion problem that consists in running ADP on the instance ob-
tained by associating a sourcedestination (s, t) to each edge
FLAMMINI ET AL.
{s, t} F . The paths returned by ADP plus an eventual path
connecting s to the node with minimum eccentricity in A if
s is not a base station form the VPs of an O(r)-approximate
solution for the layout problem.
To the best of our knowledge, the best general
algorithm
for DP has an approximation ratio r = O( |E| log |V |),
while r = O(polylog|V |) [28]. Therefore, an O( |E| log |V |)-
approximation algorithm exists for the maximum load min-
imization in layout with d = 1, while any algorithm with
an asymptotic better approximation ratio would improve
upon [28].
4. Optimal chain layouts for d
2
Starting from the hardness results shown in the previous sec-
tion, we now focus on specific topologies and provide optimal
layouts for chain networks when the maximum channel dis-
tance d is at most 2. More precisely, we consider the case
in which the physical graph is a chain Cn of n nodes, that is
V = {1, 2, . . . , n}, E = {{v, v + 1} | 1
v
n 1} and
loss of generality, we take the leftmost node of the chain as
the source, i.e., s = 1, as otherwise we can split the layout
construction problem into two equivalent independent sub-
problems for the left- and the right-hand sides of the source,
respectively.
Given fixed h,l,d and a h, l, d-layout
for a chain Cn ,
we say that
is optimal if no h, l, d-layout exists for any
chain Cm with m > n.
By the considerations of the previous section for d = 1, the
virtual topology induced by the VPs of any h, l, 1-layout

coincides with the adjacency graph A and thus with Cn . As a
consequence, the largest chain admitting a h, l, 1-layout is,
such that, n = h + 1. Therefore, in the remaining part of this
section we focus on the case d = 2.
In the following we denote by u, v the unique VP cor-
responding to the simple path from u to v in Cn and by
s, v1 v1 , v2  . . . vk , v or simply s, v1 , v2 , . . . , vk , v the
virtual channel VC(v) of v given by the concatenation of the
VPs s, v1 , v1 , v2 , . . . , vk , v.
The following lemma establishes that, when moving in one
direction along a chain, some VPs are accumulated, that is
they cannot be removed from the VCs of the successive nodes
encountered along the same direction.
Lemma 4.1. Given a h, l, 2-layout
for a chain network
and a node v, if VC(v) = s, v1 , v2 , . . . , vk , v and in VC(v)
there exist two consecutive VPs vi1 , vi , vi , vi+1  with
vi1 < v and vi < v (resp. vi1 > v and vi > v), then
for every u  v (resp. u
v), s, v1 , v2 , . . . , vi1 , vi  is a
prefix of VC(u).
Proof. Assume first that vi1 < v and vi < v and let
u > v be the first node, such that, vi , vi+1 
/ VC(u). Since
d(u 1, u)
2, to reach u we can only add vi , u. This
process can be iterated to every node w with w > u, hence the
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS
claim holds. An analogous argument applies when vi1 > v
and vi > v.  VC(u) for every u < v 1. In this case also vk , v is not
Another useful property of h, l, 2-layouts is that the pre-
fixes of a VC are the VCs of their final nodes.
Lemma 4.2. There exists an optimal h, l, 2-layout
for
a chain network, such that, for every node v with VC(v) =
s, v1 , v2 , . . . , vk1 , vk , v, VC(vi ) = s, v1 , v2 , . . . , vi  for
every i
k.
Proof. Let
be any optimal h, l, 2-layout. We now prove
that, if for a given node v each prefix of every virtual channel
VC(u) with u < v is the virtual channel of the corresponding
final node, then
can be modified in such a way that such
a property is satisfied also by VC(v). This clearly proves the
lemma.
Trivially the property is satisfied by the virtual channel of
source s, since it is empty. Therefore, let v  2 be any node,
such that, the property is true for all the VCs of the previous
nodes and let VC(v) = s, v1 , v2 , . . . , vk1 , vk , v. Recalling
that d(v 1, v)
2, it is possible to distinguish among the
following cases:
1. VC(v 1) = s, v1 , v2 , . . . , vk , v 1,
2. VC(v 1) = s, v1 , v2 , . . . , vk , v, v 1,
3. VC(v 1) = s, v1 , v2 , . . . , vk , that is vk = v 1,
4. VC(v 1) = s, v1 , v2 , . . . , vk , vk+1 , v, and
5. VC(v 1) = s, v1 , v2 , . . . , vk1 , that is vk1 = v 1.
Since by the hypothesis the claim is true for VC(v 1)
and s, v1 , v2 , . . . , vk  is a prefix of VC(v 1), VC(vi ) =
s, v1 , v2 , . . . , vi  for each vi with 1
i
k. Therefore,
every prefix of VC(v) is a VC.
If VC(v 1) = s, v1 , v2 , . . . , vk1 , that is vk1 = v 1,
we further distinguish the following two subcases.
(I) vk > v. In this case the VP v, vk  must be
deleted in the VC of a node u with v < u
vk . If
u = vk VC(vk ) = s, v1 , v2 , . . . , vk , otherwise, since
d(u 1, u)
2, VC(u) = s, v1 , v2 , . . . , vk , u and iterating
the same argument to the VP u, vk  we finally have that again
VC(vk ) = s, v1 , v2 , . . . , vk . Therefore, since for each vi
with i < k s, v1 , v2 , . . . , vi  is a prefix of VC(v1), we have
VC(vi ) = s, v1 , v2 , . . . , vi  for every vi with 1
i
k.
(II) vk < v 1. If the VP vk , v 1 is contained in the
VC of a node u < v 1, then starting from the source s, in
VC(u) vk , v 1 is not traversed from vk to v 1, otherwise
by hypothesis vk , v 1 would be contained in VC(v 1)
and thus it could not be added to VC(v 1) with vk , v to
obtain VC(v). Therefore, vk , v 1 is traversed from v 1 to
vk and again by hypothesis the prefix of VC(u) till v 1 coin-
cides with VC(v 1) and VC(vk ) = s, v1 , v2 , . . . , vk1 , vk .
As in the previous subcase, since for each vi with i < k
s, v1 , v2 , . . . , vi  is a prefix of VC(v 1), we have VC(vi ) =
s, v1 , v2 , . . . , vi  for every vi with 1
i
k.
39
Assume then that the VP vk , v 1 is not contained in
contained in any VC(u) with u < v 1, because otherwise,
with no matter of the sense in which vk , v is traversed in
VC(u), VC(v) could not contain vk , v 1, as by hypotheses
vk , v 1 does not belong to VC(u).
If vk , v is contained in VC(v + 1), then by lemma 4.1
VC(v) is a prefix of all the VCs VC(u) with u  v. Therefore,
the layout obtained by deleting the VPs vk , v1 and vk , v,
adding v1, v and modifying each VC(u) = s, v1 , v2 , . . . ,
v 1, vk , v, . . . , u with u  v as VC(u) = s, v1 , v2 , . . . ,
v 1, v, . . . , u does not increase the hop count of any node,
the load of any edge and the channel distance of the ad-
jacent nodes. Therefore, since in the new layout the vir-
tual channels of the nodes before v are not modified and
VC(v) = s, v1 , v2 , . . . , v 1, v = s, v1 , v2 , . . . , vk1 , v,
s, v1 , v2 , . . . , vi  is a prefix of VC(v 1) for every vi with
1
i < k and thus VC(vi ) = s, v1 , v2 , . . . , vi .
If vk , v is not contained in VC(v + 1), then by lemma 4.1
s, v1 , v2 , . . . , v 1, vk  is a prefix of all the VCs VC(u) with
u  v. Moreover, all the VPs starting at vk contained in
the VC of some node u > v are not contained in any VC
VC(w) with w
v 1, as otherwise by hypothesis VC(u)
would not contain vk , v 1. Notice also that the other
endpoint of each such VP is greater than v, as otherwise by
the maximum channel distance it cannot be used in the VCs
of the nodes after v. It is thus possible to modify the lay-
out
as follows. The VPs vk , v 1 and vk , v are deleted,
v 1, v is added, the VPs vk , w contained in the VC of
some node u > v are substituted with v, w and, finally, each
VC(u) = s, v1 , v2 , . . . , v 1, vk , w, . . . , u with u  v is
modified as VC(u) = s, v1 , v2 , . . . , v 1, v, w, . . . , u. By
the above considerations, the new layout does not increase the
hop count of any node, the load of any edge and the channel
distance of the adjacent nodes. Moreover, it does not modify
the virtual channels of the nodes before v. Therefore, again
VC(v) = s, v1 , v2 , . . . , v 1, v = s, v1 , v2 , . . . , vk1 , v,
s, v1 , v2 , . . . , vi  is a prefix of VC(v 1) for every vi with
1
i < k and thus VC(vi ) = s, v1 , v2 , . . . , vi . 
Motivated by the previous lemma, even if not explicitly
stated, in the remaining part of this section, we restrict our
attention to layouts in which all the prefixes of each VC are
the VCs of the corresponding final nodes. In fact, this does
not affect the correctness of our results, since the optimality
is preserved under such assumption.
The following corollary is a direct consequence of the pre-
vious lemma.
Corollary 4.3. Every VP of a h, l, 2-layout
for a chain
network is the final VP of exactly one of its two endpoints.
A last useful property that allows to suitably bound the
maximum size of a chain admitting a h, l, 2-layout is estab-
lished in the following lemma.
Lemma 4.4. Given a h, l, 2-layout
for a chain network
and any j
h, let u and v be the last nodes with hop count
40
h(u) = j 1 and h(v) = j , respectively. Then the last VPs
of all the VCs reaching the nodes from u+1 to v share a same
physical link.
Proof. Let us first prove the claim for j = 1, and thus with
u corresponding to the source s = 1.
In order to show that the lemma holds for j = 1, it is
sufficient to prove that no two edge-disjoint VPs exist in the
VCs of the nodes from u + 1 to v.
Assume by contradiction that such property does not hold
and let x, y,w, z be a pair of closest edge-disjoint VPs in
such VCs with x < y
w < z
v. Then y = w or
y = w 1, otherwise any other VP used to reach a node
between y and w would be disjoint from x, y or w, z and
closest to x, y or w, z, thus contradicting the hypothesis.
If y = w 1 then x, y is not used to reach y. In
fact, x, y cannot be contained in VC(w), as by lemma 4.1
it would be contained also in VC(v) against the hypothe-
sis h(v) = 1. Thus, if x, y is used to reach y, since
d(y, w)
2, when moving from y to w x, y must be re-
moved and the VP x, w must be added to reconstruct VC(w).
But then x, w would be a VP closer to w, z than x, y,
again contradicting the hypothesis.
By corollary 4.3, x, y is then used to reach x. If y is
reached by a VP q, y with q < y, then by corollary 4.3
w, z is used to reach z, and thus by lemma 4.1 q, y is con-
tained in VC(v) again contradicting the hypothesis h(v) = 1.
If q > y x, y and y, q would be closer than x, y and
w, z: a contradiction.
Therefore, y = w 1 cannot hold and it must be y = w.
Recalling that by corollary 4.3 every VP is used to reach
exactly one of its endpoints, i.e., it is the final VP of exactly
one of the VCs of its endpoints, it is possible to distinguish
the following cases:
x, y reaches x and y, z reaches z.
By corollary 4.3 y is reached by another VP q, y with
q = x and q = z. If q > y, then by lemma 4.1 q, y is
contained in all the VCs of the nodes before x, x included,
thus contradicting h(s) = 0. Similarly, if q < y, q, y is
contained in all the VCs of the nodes after z, z included,
thus contradicting h(v) = 1. Therefore, this case cannot
hold.
x, y reaches y and y, z reaches z.
Since by lemma 4.2 every prefix of VC(z) corresponds to
the VC of its final node, x, y is contained in VC(z) and
by lemma 4.1 x, y belongs to VC(v), thus contradicting
h(v) = 1.
x, y reaches x and y, z reaches y.
This case is symmetric to the previous one and completely
Figure 1. Optimal h, l, 2-layout for a chain network.
FLAMMINI ET AL.
analogous considerations show that it would contradict
h(s) = 0.
x, y reaches y and y, z reaches y. This case is clearly
impossible, since only one of the two VPs x, y and by
y, z can be the last one of VC(y).
In conclusion, neither y = w can hold and therefore no
two edge-disjoint VPs exist before v.
In order to extend the proof to every j
h, observe that
since u is the last node with hop count h(u) = j 1 and
d = 2, all the VCs of the nodes w  u have VC(u) as prefix.
Thus an identical proof shows that all the VPs after u in the
VCs of the nodes w with u < w
v share a same physical
edge. 
It is thus possible to prove the following lemma.
Lemma 4.5. Given any h, l, 2-layout
for a chain net-
work and j
h, the last node v with h(v) = j is, such
that, v
j l + 1.
Proof. The claim trivially holds for j = 0, since only the
source s = 1 has hop count h(s) = 0.
Assume by induction that the corollary holds for a given
j , such that, 1
j < h and let u
j l + 1 the last node
with hop count h(u) = j and v the last node with hop count
h(v) = j + 1. Since by lemma 4.4 the last VPs of all the VCs
reaching the nodes from u + 1 to v share a same physical link,
there cannot be more than l nodes from u + 1 to v, otherwise
the shared link would have load greater than l. Therefore,
v
u + l
(j + 1)l + 1. 
In conclusion, the following theorem holds.
Theorem 4.6. For every h, l  0, a h, l, 2-layout for a
chain network Cn is optimal if and only if n = hl + 1.
Proof. By lemma 4.5, n
hl + 1 holds for any h, l, 2-
layout for a chain Cn . A layout attaining n = hl + 1 is de-
picted in figure 1. 
Before concluding the section, we finally observe that the
layout in figure 1 is not the only optimal one. Another exam-
ple with h = 4, l = 4 and d = 2 is shown in figure 2.
5. Optimal canonic layouts for chain networks
In this section we provide h, l, d-layouts for chain networks
that are optimal within the class of the canonic layouts. In
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS 41

Figure 2. An alternative optimal 4, 4, 2-layout.

Figure 3. The recursive definition of T (h, l, d) for h > 0 and l > 0.

fact, such layouts have been shown to be the optimal ones If h > 0 and l > 0 T (h, l, d) contains at least two nodes
under different assumptions (see, for instance, [12,15]). In- and the lowest order node u, that is the root, is connected
formally speaking, a layout
is canonic if it does not contain by an edge to the highest order node v.
intersecting VPs and it induces a tree. More precisely, we Moreover, a chain of min{h, d/2
} trees T (hj, l 1, d)
have the following definitions. with 0
j
min{h, d/2
} 1 is attached to u in such a
way that the lowest order node of T (h, l 1, d) coincides
Definition 5.1. Two VPs u, v and w, z are crossing if u < with u and the lowest order node of each T (h j, l 1, d)
w < v < z. A layout
is crossing-free if it does not contain with 1
j
min{h, d/2
}1 coincides with the highest
any pair of crossing VPs. order node of T (h j + 1, l 1, d).
Finally, a chain of min{h 1, (d 1)/2
} reverse trees
Definition 5.2. A layout
is canonic if it is crossing-free
}
and the virtual topology induced by its VPs is a tree.
Let us say that a rooted tree is ordered if a total order
is defined on its nodes with the root being the lowest order
T r (h j, l 1, d) with 1
j
min{h 1, (d 1)/2
is attached to v in such a way that the highest order node
of T r (h 1, l, d) coincides with v and the highest order
node of each T r (h j, l 1, d) with 2
j
min{h j,
node. Then there is a one-to-one corresponds between lay-
} coincides with the lowest order node of
outs for chains and ordered trees. Namely, each node of the
tree corresponds to a node of the chain, the root corresponds
to the source s, each edge to a VP of
and finally the to-
tal order of the nodes of the tree is given by the order of
the nodes along the chain. Clearly, not all the ordered trees
yield canonic layouts, as their induced VPs might be cross-
ing. However, the one-to-one correspondence between or-
dered trees and canonic layouts is maintained if we restrict
to ordered trees in which every subtree contains a subset of
nodes that forms an interval according to the node ordering.
In other words, each subtree corresponds to a segment of the
chain not touched by the other subtrees.
Given any ordered tree T , let the reverse tree T r be the
symmetric ordered tree obtained from T by inverting the or-
der of the nodes (hence the root becomes the highest or-
der node). We now introduce a new class of ordered trees
(h, l, d) that allows to completely define the structure of an
optimal h, l, d-layout.
The definition of (h, l, d) is recursive and the solution
of the associated recurrence gives the exact number of nodes
reached by an optimal canonic h, l, d-layout. Before intro-
ducing (h, l, d), let us define another ordered subtree that is
exploited in its definition.
Definition 5.3. Given any h, l, d, T (h, l, d) is an ordered tree
recursively defined as follows.
(d 1)/2
T r (h j + 1, l 1, d).
An example of T (h, l, d) is depicted in figure 3. Infor-
mally speaking, a T (h, l, d) corresponds to the sublayout of
a canonic layout
induced by all the VPs occurring under
a given VP, with the lowest order node being closer to the
source. Thus, T (h, l, d) is the subtree induced by all the VPs
whose endpoints occur from the first endpoint of the given VP
until the second endpoint.
Directly from the definition, it follows that all the nodes in
T (h, l, d) are at distance at most h from u, and thus at most
h additional hops from the node corresponding to u in
are
sufficient to reach the other nodes corresponding to T (h, l, d)
in the chain. Moreover, the load yielded by T (h, l, d) on its
segment of the chain is bounded by l. Finally, the channel
distance between two consecutive nodes belonging to the sub-
chain of T (h, l, d) is always at most equal to d. In fact, it
is given by the maximum distance in T (h, l, d) between two
nodes adjacent in the ordering. Therefore, assuming by induc-
tion that such property holds inside the subtrees T (j, l 1, d)
(and thus T r (j, l 1, d)), in order to show that it holds also
in T (h, l, d) it is sufficient to prove that the final node of the
chain of subtrees attached to u and the final node of the other
reverse chain attached to v, that is the only not yet considered
adjacent pair of nodes, is at distance at most d. But such nodes
are at distance min{h, d/2
}+1+min{h1, (d 1)/2
}

If h = 0 or l = 0 T (h, l, d) consists of a single node. d/2
+ 1 + (d 1)/2
= d. Therefore, also the channel
42
Figure 4. (h, l, d) in terms of trees of type T (a) and the alternative recursive definition (b).
distance within the subchain of T (h, l, d) is bounded by d.
Clearly, symmetric considerations hold for each T r (h, l, d).
We are now ready to define the final tree (h, l, d).
Definition 5.4. The ordered tree (h, l, d) is formed by the
chain of h trees T (j, l, d), 1
j
h, such that, the lowest
order node of T (j, l, d) coincides the highest order node of
T (j + 1, l, d) for 1
j < h (see figure 4).
Notice that, if h = 0 or l = 0, (h, l, d) consists of just a
single node. Moreover, an alternative recursive definition of
(h, l, d) is given by a T (h, l, d) attached to a (h 1, l, d)
tree (again see figure 4).
Let Tn (h, l, d) denote the number of nodes of T (h, l, d)
(and thus of T r (h, l, d)) minus one. Then, directly from
definition 5.3, Tn (h, l, d) = 0 if h = 0 or l = 0, other-
FLAMMINI ET AL.
Assume then that the claim is true for l  1, that is, for all
the VPs of a canonic h, l, d-layout
with l l  + 1 VPs
over them and let u, v be a VP of
with h(u) = h h ,
h(v) = h h + 1 and l l  VPs over it. Let w be the
last node with u
w
v (that is under u, v) reached
by a VC stepping through u and not from v, and consider
the subchain of d1 VPs w1 , . . . , wd1 +1  with w1 = u and
wd1 +1 = w connecting u to w. Similarly, let zd2 +1 , . . . , z1 
with z1 = w + 1 and zd2 +1 = v the subchain of d2 VPs
connecting v to w + 1. Since w and w + 1 are adjacent and
the maximum channel distance is d, it must be d1 +d2 +1
d.
Moreover, since h(w)
h and h(w + 1)
h, d1
h and
d2
h 1. Therefore, since such subchains and with all their
VPs occur under u, v, by applying the inductive assumption
it follows that
v u = (w u) + (v w)

min{h,d/2
}1
wise Tn (h, l, d) = 1 + j =0 Tn (h j, l 1, d) +  
1 +1
d 2 +1
d

min{h1,(d1)/2
}
j =1 Tn (h j, l 1, d). = (wi wi1 ) + 1 + (zi zi1 )
Moreover, by definition 5.4, denoted as n (h, l, d) the i=2 i=2
number of nodes in (h, l, d), n (h, l, d) = 1 +

1 +1
d
h  
k=1 Tn (k, l, d).
1+ Tn h i + 2, l  1, d
Clearly, by the above observations, (h, l, d) corresponds i=2
to a canonic h, l, d-layout for a chain network. Actually, a
2 +1
d
stronger result holds.  
+ Tn h d2 + i 2, l  1, d
i=2
Lemma 5.5. The layout induced by (h, l, d) is optimal
1 1
d
within the class of the canonic h, l, d-layouts for chain net-  
works. =1+ Tn h i, l  1, d
i=0

Proof. Let
be any canonic h, l, d-layout for a chain Cn . 

d2
 
It is sufficient to show that n
n (h, l, d). + Tn h i, l  1, d
Let VC(n) = v1 , . . . , vk  with v1 = s, vk = n and k
i=1
h + 1 the VC of the last node of the chain in
. We prove  , d/2}1
min{h
 
that vi vi1
Tn (h i + 2, l, d) for every i,
such that,
1+ Tn h i, l  1, d
2
i
k. In fact, this implies n = vk = v1 + ki=2 (vi i=0



vi1 )
1 + ki=2 Tn (h i + 2, l, d) = 1 + k2 i=0 Tn (h min{h 1,

h1
h (d1)/2}  
i, l, d)
1 + i=0 Tn (h i, l, d) = 1 + i=1 Tn (i, l, d) = + Tn h i, l  1, d
n (h, l, d). i=1
In order to show that vi vi1
Tn (h i + 2, l, d) for  
 
= Tn h , l , d .
every i, such that, 2
i
k, it suffices to prove that, given
any VP u, v of a canonic h, l, d-layout, such that, h(u) = A completely symmetric proof shows that vu
Tn (h , l  , d)
h h , h(v) = h h + 1 or vice versa and there exist l l  for every VP u, v of
with h(u) = hh +1, h(v) = hh
VPs over it, that is of the form w, z with w
u and z > v and l l  VPs over it. 
or w < u and z  v, it is v u
Tn (h , l  , d). If l  = 1,
it must be v = u + 1, otherwise the nodes between u and v Starting from lemma 5.5, in order to determine the largest
could not be reached from the source without exceeding the chain admitting a canonic h, l, d-layout, it is sufficient to
maximum load l. Recalling definition 5.3, v u = 1 = estimate the number of nodes contained in the tree (h, l, d),
Tn (h , 1, d). that is, n (h, l, d).
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS 43

 d/21 i
Before solving the recurrence on Tn (h, l, d) and conse- 2x + 2x d/22 + + 2x 2 + 2x + 1
quently estimate n (h, l, d), we recall that given n + 1 pos- i  
 i  d/21 k
itive integers m, k1 , . . . , kn , such that, m = k1 + + kn , = + 2x d/22 + + 2x 2 + 2x 1
m 
2x
the multinomial coefficient k1 ,...,k n
is defined as m!/(k1 ! k1
k1 =0
k2 ! kn !) (see, for instance, [17]).  

i
i  d/22 k
= 2k1
x + x d/23 + + x + 1 1 x k1
k1
Lemma 5.6. For every h > 0, l > 0 and d > 1, if d is even k1 =0


i  k1  
Tn (h, l, d) i k1
= 2k1 (x d/23 + x d/24 +

l 
h1  k1 k2
k1 =0 k2 =0
= 2k1 k2 k1 +k2
i=1 j =0 0
kd/21
kd/22

k2
k1
i
+ x + 1) x =
    
k1 +k2 ++kd/21 =j
i  k1
i k1
  = 2k1
i k1 k2
, k1 =0 k2 =0
i k1 , k1 k2 , . . . , kd/22 kd/21, kd/21
 kd/22
kd/22
while if d is odd x k1 +k2 ++kd/21
kd/21
kd/21 =0
Tn (h, l, d)

i 
k1 kd/22
   

l 
h1  i k1
= 2 k1 k(d1)/2 = 2k1
k1 k2
i=1 j =0 0
k(d1)/2
k(d1)/21

k2
k1
i k1 =0 k2 =0 kd/21 =0
k1 +k2 ++k(d1)/2 =j
 
kd/22 k1 +k2 ++kd/21
  x ,
i kd/21
.
i k1 , k1 k2 , . . . , k(d1)/21 k(d1)/2, k(d1)/2 that can be rewritten as

2k1
0
kd/21
kd/22

k2
k1
i
Proof. Let M be the matrix defined as follows:     
i k1 kd/22 k1 +k2 ++kd/21
x
1 if i = 0 and j = 0, k1 k2 kd/21



0 if i = 0 and j > 0,


i(d/21) 

 j
Mi1,t = 2k1
Mi,j = j =0 0
kd/21
kd/22

k2
k1
i

t =max{0,j d/2 +1}
k1 +k2 ++kd/21 =j

    

j
+
i k1 kd/22 j


Mi1,t otherwise. x .
t =max{1,j (d1)/2 } k1 k2 kd/21

Note that a generic element Mi,j represents the number Therefore, recalling the definition of multinomial coefficient
of subtrees T (h j, l i, d) and T r (h j, l i, d) that and that Mi,j is the coefficient of x j in (2x d/21 + 2x d/22 +
occur in T (h, l, d) or analogously in the expansion of the re- + 2x 2 + 2x + 1)i , it follows that

cursive definition of T (h, l, d) until obtaining only trees of Mi,j = 2k1
load

l l
i. Moreover, by the recurrence of Tn , it results that 0
kd/21
kd/22

k2
k1
i
h1
i=1 j =0 Mi,j is exactly the number of nodes in T (h, l, d) k1 +k2 ++kd/21 =j
minus one, that is the value Tn (h, l, d).  
i
In order to determine the sum of the first h columns and .
i k1 , k1 k2 , . . . , kd/22 kd/21, kd/21
the l rows without the first of M, we observe that each row i
of M corresponds to the coefficients of the ith power of the For the case of odd d we obtain
polynomial ((x d/2 1 +x d/2 2 + +x +1)+(x (d1)/2 +  i
x (d1)/2 + 2x (d1)/21 + + 2x 2 + 2x + 1
x (d1)/2 1 + + x))i . More precisely, a generic element
i  
Mi,j is equal to the coefficient of x j in the expansion of the i  (d1)/2 k
= x + 2x (d1)/21 + + 2x 2 + 2x 1
polynomial ((x d/2 1 +x d/2 2 + +x +1)+(x (d1)/2 + k1
k1 =0
x (d1)/2 1 + + x))i .
i  

If d is even, by
  d/2 1 times the well-known
applying
=
i 
x (d1)/21
equality (a +b)i = ik=0 ki a k bik to (2x d/21 +2x d/22 + k1
k1 =0
+ 2x 2 + 2x + 1)i with a = 2x d/21 + 2x d/22 + + 2x 2 k1
+ 2x and b = 1 and iterating the same argument, we obtain + 2x (d1)/22 + + 2x + 2 x k1
44 FLAMMINI ET AL.

i  
 k1  
i k1  n (h, l, d)
= x (d1)/22 + 2x (d1)/23 +
k1 k2 
h 
l 
h1 
k1 =0 k2 =0 =1+ 2k1 k(d1)/2
k2 k1 k2 k1 +k2
+ 2x + 1 2 x k=1 i=1 j =0 0
k(d1)/2
k(d1)/21

k2
k1
i
k1 +k2 ++k(d1)/2 =j
     k2    
i  k1  k2  (d1)/23
i k1
= x i
,
k1 =0
k1
k2 =0
k2
k3 =0
k3 i k1 , k1 k2 , . . . k(d1)/21 k(d1)/2, k(d1)/2
k
+ 2x (d1)/24 + + 2x + 1 3 2k1 k3 x k1 +k2 +k3 if d is odd.
=
 i  k1 k(d1)/21
    Unfortunately, n (h, l, d) in general cannot be expressed
i k1
= 2k1 k(d1)/2 by means of a more compact closed formula. However, in
k1 k2 some cases it can be significantly simplified. For instance,
k1 =0 k2 =0 k(d1)/2 =0
 
k(d1)/21 k1 +k2 ++k(d1)/2 d = 2: n (h, l, 2) = h l + 1.
x ,
k(d1)/2 In fact, by the definition of the matrix M in the proof of
lemma 5.6, the only non null elements of M belong to the
that can be rewritten as first column and their value is always equal to one. Hence,
 the number of the nodes of every Tn (k, l, 2) is l and
2k1 k(d1)/2
0
k(d1)/2
k(d1)/21

k2
k1
i 
h
     n (h, l, 2) = 1 + Tn (k, l 1, 2) = 1 + h l.
i k1 k(d1)/21 k1 +k2 ++k(d1)/2
x k=1
k1 k2 k(d1)/2
This coincides with the result obtained in the previous sec-

i((d1)/2)  tion, and in fact (h, l, 2) coincides with the layout con-
k1 k(d1)/2
= 2
struction depicted in figure 1.
j =0 0
k(d1)/2
k(d1)/21

k2
k1
i
min{h,l} i1 hl  1
k1 +k2 ++k(d1)/2 =j d  2h: n (h, l, d) = i=0 2 i i + 2.
    
i k1 k(d1)/21 j In fact, in this case, our model and constructions coincide
x . with the ones in [9].
k1 k2 k(d1)/2

Therefore,
 6. Conclusion
Mi,j = 2k1 k(d1)/2
0
k(d1)/2
k(d1)/21

k2
k1
i The main question left open in the paper is if the family of
k1 +k2 ++k(d1)/2 =j the canonic layouts contains optimal layouts for d > 2. Even
 
i if not claimed explicitly, our constructions show that this is
. true for d
2 and the previous results shown in the literature
i k1 , k1 k2 , . . . , k(d1)/21 k(d1)/2, k(d1)/2
seem to confirm this conjecture.
In every


case, the claim follows by recalling that Tn (h, l, d) = Moreover, it would nice to extend our results to more gen-
l h1
i=1 j =0 Mi,j .  eral topologies and to the case in which the physical and ad-
jacency graphs are not coincident.
Theorem 5.7. For every h > 0, l > 0 and d > 1, the max- Another worth investigating issue is the extension to other
imum number of nodes reachable in a chain network by a communication patterns like multicast and all-to-all.
canonic h, l, d-layout is Finally, it would be worth to investigate the approximabil-
ity of the layout construction problem for d > 1.
n (h, l, d)

h
=1+ Tn (k, l, d) References
k=1
[1] S. Ahn, R.P. Tsang, S.R. Tong and D.H.C. Du, Virtual path layout de-

h 
l 
h1  sign on ATM networks, in: Proc. of the 13th INFOCOM Conf. (1994)
=1+ 2 k1
pp. 192200.
k=1 i=1 j =0 0
kd/21
kd/22

k2
k1
i [2] B.A. Akyol and D.C. Cox. Rerouting for handoff in a wireless ATM
k1 +k2 ++kd/21 =j network, in: Proc. of the IEEE Internat. Conf. on Universal Personal
  Communications (1996).
i
, [3] L. Becchetti, P. Bertolazzi, C. Gaibisso and G. Gambosi, On the de-
i k1 , k1 k2 , . . . , kd/22 kd/21, kd/21 sign of efficient ATM schemes, in: Proc. of the SOFSEM Conf., Lec-
ture Notes in Computer Science, Vol. 1338 (Springer, New York, 1997)
if d is even, and pp. 375382.
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS
[4] J. Burgin and D. Dorman, Broadband ISDN resource management: The
role of virtual paths, IEEE Communicatons Magazine 29 (1991).
[5] M. Cheng, S. Rajagopalan, L. Chang, G. Pollini and M. Barton. PCS
mobility support over fixed ATM networks, IEEE Communications
Magazine 35 (1997) 8292.
[6] C. Chrysostomou, A. Pitsillides and F. Pavlidou, A survey of wireless
ATM handover iusses, in: Proc. of the Internat. Symposium of 3G In-
frastructure and Services, 3GIS, Vol. 2/3 (2001) pp. 3439.
[7] I. Cidon, O. Gerstel and S. Zaks, A scalable approach to routing in ATM
networks, in: Proc. of the 8th Internat. Workshop on Distributed Algo-
rithms, eds. G. Tel and P.M.B. Vitnyi, Terschelling, The Netherlands
(October 1994), Lecture Notes in Computer Sience, Vol. 857 (Springer,
New York, 1994) pp. 209222; submitted for publication in IEEE/ACM
Transactions on Networking.
[8] R. Cohen and A. Segall, Connection management and rerouting in
ATM networks, in: Proc. of the 13th INFOCOM Conf. (1994) pp. 184
191.
[9] Y. Dinitz, M. Feighelstein and S. Zaks, On optimal graphs embedded
into path and rings, with analysis using l1 -spheres, in: Proc. of the 23rd
Internat. Workshop on Graph-Theoretic Concepts in Computer Science
(WG), Lecture Notes in Computer Science, Vol. 1335 (Springer, New
York, 1997) pp. 171183.
[10] T. Eilam, M. Flammini and S. Zaks, A complete characterization
of the path layout construction problem for ATM networks with
given hop count and load, in: Proc. of the 24th Internat. Collo-
quium on Automata, Languages and Programming (ICALP), Lecture
Notes in Computer Science, Vol. 1256 (Springer, New York, 1997)
pp. 527537.
[11] M. Flammini, G. Gambosi and A. Navarra, Wireless ATM layouts for
chain networks, in: Proc. of the 17th Internat. Parallel and Distributed
Processing Symposium (IPDPS), 3rd Workshop on Wireless, Mobile
and Ad Hoc Networks (WMAN), IEEE Computer Society (2003)
p. 220.1.
[12] M. Flammini, A. Gasparini, G. Gambosi and A. Navarra, Dy-
namic layouts for wireless ATM, in: Proc. of the 9th Internat.
Conf. on Parallel and Distributed Computing (Euro-Par), Lecture
Notes in Computer Science, Vol. 2790 (Springer, New York, 2003)
pp. 10561063.
[13] O. Gerstel, I. Cidon and S. Zaks, The layout of virtual paths in
ATM networks, IEEE/ACM Transactions on Networking 4(6) (1996)
873884.
[14] O. Gerstel, A. Wool and S. Zaks, Optimal layouts on a chain ATM
network, in: 3rd Annual European Symposium on Algorithms (ESA),
Corfu, Greece (September 1995), Lecture Notes in Computer Science,
Vol. 979 (Springer, New York, 1995) pp. 508522; to appear in
Discrete Applied Mathematics.
[15] O. Gerstel and S. Zaks, The virtual path layout problem in fast
networks, in: Proc. of the 13th ACM Symposium on Principles of Dis-
tributed Computing, Los Angeles, USA (August 1994) pp. 235243.
[16] J.D. Gibson, The Mobile Communications Handbook, 2nd ed. (CRC
Press/IEEE Press, 1999).
[17] R.L. Graham, D.E. Knuth and O. Patashnik, Concrete Mathematics
(Addison-Wesley, Reading, MA, 1989).
[18] R. Hndler and M.N. Huber, Integrated Broadband Networks: An
Introduction to ATM-Based Networks (Addison-Wesley, Reading, MA,
1991).
[19] ITU recommendation, I series, Blue Book (November 1990).
[20] R.M. Karp, On the computational complexity of combinatorial
problems, Networks 5 (1975) 4568.
[21] E. Kranakis, D. Krizanc and A. Pelc, Hop-congestion tradeoffs in
ATM networks, in: Proc. of the 9th IEEE Symposium on Parallel and
Distributed Processing (1995) pp. 662668.
[22] M. Mouly and M.B. Pautet, The GSM System for Mobile Communica-
tions (Cell & Sys, 1993).
45
[23] G. Parry, Wireless ATM MAC protocols a literature survey, WARP
Project URL, http://vera.ee.und.ac.za/coe/warp
(1999).
[24] C. Partridge, Gigabit Networking (Addison-Wesley, Reading, MA,
1994).
[25] K.I. Sato, S. Ohta and I. Tokizawa, Broad-band ATM network archi-
tecture based on virtual paths, IEEE Transactions on Communications
38(8) (1990) 12121222.
[26] Y. Sato and K.I. Sato, Virtual path and link capacity design for ATM
networks, IEEE Journal on Selected Areas in Communications 9
(1991).
[27] D. Sobirk and J.M. Karlsson, A survey of wireless ATM MAC
protocols, in: Proc. of the Internat. Conf. on the Performance
and Management of Complex Communication Networks (PMCCN)
(Chapman & Hall, London, 1997).
[28] A. Srinivasan, Improved approximations for edge-disjoint paths,
unsplittable flow, and related routing problems, in: Proc. of the 38th
Annual IEEE Symposium on Foundations of Computer Science (FOCS)
(IEEE Computer Society Press, Los Alamitos, CA, 1997) pp. 416425.
[29] L. Stacho and I. Vrto, Virtual path layouts for some bounded degree
networks, in: Proc. of the 3rd Colloquium on Structural Information
and Communication Complexity (SIROCCO) (Carleton Univ. Press,
1996) pp. 269278.
[30] S. Zaks, Path layouts in ATM networks, in: Proc. of the SOFSEM
Conf., Lecture Notes in Computer Science, Vol. 1338 (Springer, New
York, 1997) pp. 144160.
Michele Flammini received the degree in computer
science at the University of LAquila in 1990 and the
Ph.D. degree in computer science at the University of
Rome La Sapienza in 1995. He is associate profes-
sor at the computer science Department of the Uni-
versity of LAquila since 2000. His research inter-
ests include algorithms and computational complex-
ity, communication problems in interconnection net-
works and routing. He has authored and co-authored
more than 50 papers in his fields of interest published
in the most reputed international conferences and journals.
E-mail: flammini@di.univaq.it
Giorgio Gambosi received the degree in electronic
engineering at the University of Rome La Sapienza
in 1980. He is full professor at the Department of
Mathematics of the University of Rome Tor Ver-
gata. His research interests include distributed sys-
tems, network management algorithms and routing.
He has coauthored about 70 research papers in rele-
vant international conferences and journals.
E-mail: gambosi@mat.utovrm.it
Alfredo Navarra received the degree in computer
science at the University of LAquila in 2000. He
spent one year at the research institute INRIA of
Sophia Antipolis (France) collaborating with the
MASCOTTE project group and now he is ending his
Ph.D. degree in computer science at the University of
Rome La Sapienza. His research interests include
algorithms and computational complexity, ATM, op-
tical and wireless communication networks.
E-mail: navarra@di.univaq.it;
navarra@sophia.inria.fr
 Mobile Networks and Applications 10, 4759, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Ad Hoc Multicast Routing Algorithm with Swarm Intelligence

CHIEN-CHUNG SHEN
Department of Computer and Information Sciences, University of Delaware, Newark, DE 19716, USA

CHAIPORN JAIKAEO
Department of Computer Engineering, Faculty of Engineering, Kasetsart University, 50 Phahonyothin Rd., Lardyaw, JatuJak, Bangkok 10900, Thailand

Abstract. Swarm intelligence refers to complex behaviors that arise from very simple individual behaviors and interactions, which is
often observed in nature, especially among social insects such as ants. Although each individual (an ant) has little intelligence and simply
follows basic rules using local information obtained from the environment, such as ants pheromone trail laying and following behavior,
globally optimized behaviors, such as finding a shortest path, emerge when they work collectively as a group. In this paper, we apply
this biologically inspired metaphor to the multicast routing problem in mobile ad hoc networks. Our proposed multicast protocol adapts
a core-based approach which establishes multicast connectivity among members through a designated node (core). An initial multicast
connection can be rapidly setup by having the core flood the network with an announcement so that nodes on the reverse paths to the core
will be requested by group members to serve as forwarding nodes. In addition, each member who is not the core periodically deploys a small
packet that behaves like an ant to opportunistically explore different paths to the core. This exploration mechanism enables the protocol to
discover new forwarding nodes that yield lower total forwarding costs, where cost is abstract and can be used to represent any metric to suit
the application. Simulations have been conducted to demonstrate the performance of the proposed approach and to compare it with certain
existing multicast protocols.
Keywords: ad hoc networks, multicast routing, swarm intelligence

1. Introduction from a sender to a receiver to increase robustness at the price

Mobile wireless ad hoc networks consist of mobile nodes
that autonomously establish connectivity via multihop wire-
less communications. Without relying on any existing, pre-
configured network infrastructure or centralized control, they
are useful in many situations where impromptu communi-
cation facilities are required, such as battlefield communi-
cations and disaster relief missions. In many applications,
nodes are likely to collaborate to achieve common goals and
are expected to communicate as a group rather than as pairs
of individuals (point-to-point). For instance, soldiers roam-
ing in the battlefield may need to keep listening to their
group commander (point-to-multipoint), or a group of com-
manders exchange current mission scenarios with one another
(multipoint-to-multipoint). Therefore, multicast communica-
tion serves as one critical operation to support these applica-
tions.
Many different multicast protocols have been proposed for
ad hoc networks. Some protocols are based on constructing
a tree spanning all the group members. A node then accepts
packets only when they are coming from another node with
which a tree branch has been established. However, since
there is only a single path between a pair of sender and re-
ceiver, the scheme is vulnerable to network dynamics. Con-
sequently, several protocols aim to construct a mesh that al-
lows data packets to be transmitted over more than one path
This work is supported in part by National Science Foundation under grant
ANI-0240398.
Corresponding author.
of redundancy in data transmission. Multicast protocols can
also be classified by how multicast connectivity is established
and maintained. In a source-based approach, a tree or a mesh
is constructed per multicast sender, where the construction
process is often initiated by the sender. While in a group-
shared tree/mesh approach, a single multicast connection is
shared by all senders of the same group. One common tech-
nique used in this approach is to assign a node, known as the
rendezvous point or the core [2], to accept join requests from
members. The multicast connection then consists of shortest
paths from the core to each of the members.
In this paper, we propose a novel multicast routing pro-
tocol for mobile ad hoc networks that adopts swarm intelli-
gence to reduce the number of nodes used to establish mul-
ticast connectivity. We name the protocol Multicast for Ad
Hoc Networks with Swarm Intelligence or MANSI for short.
Swarm intelligence refers to complex behaviors that arise
from very simple individual behaviors and interactions, which
is often observed in nature, especially among social insects
such as ants and honeybees. Although each individual (for in-
stance, an ant) has little intelligence and simply follows basic
rules using local information obtained from the environment,
global optimization objectives1 emerge when they work col-
lectively as a group. Similarly, MANSI utilizes small con-
trol packets equivalent to ants in the physical world. These
packets, traveling like biological ants, deposit control infor-
1 An example of these is that ants often find a shortest path from their nest
to the food source.
48
mation at nodes they visit, similar to the way ants laying
pheromone trails. This information, in turn, affects the be-
havior of other ant packets. With this form of indirect com-
munication (known as stigmergy), the deployment of ant-like
packets resembles an adaptive distributed control system that
evolves itself to a more efficient state, accommodating the
current condition of the environment.
For each multicast group, MANSI determines a set of in-
termediate nodes, forming a forwarding set, that connect all
the group members together and are shared among all the
group senders. By adopting a core-based approach, the for-
warding set is initially formed by nodes that are on the short-
est paths between the core and the other group members,
where the core may be one of the group members or senders.
In addition, during the lifetime of the multicast session (i.e.,
when there is at least one active sender), the forwarding set
will evolve, by means of swarm intelligence, over time into
states that yield lower cost, which is expressed in terms of
total cost of all the nodes in the forwarding set. This evolv-
ing, including exploring and learning, mechanism differen-
tiates MANSI from other existing ad hoc multicast routing
protocols. Since a nodes cost is abstract and may be defined
to represent different metrics, MANSI can be applied to many
variations of multicast routing problems for ad hoc networks
such as load balancing, secure routing, and energy conserva-
tion.
The remainder of the paper is organized as follows. We
first describe the motivation and overview of the MANSI pro-
tocol in the next section. Section 3 explains the protocol in
details. Simulation results are then presented and discussed
in section 4. Related works are reviewed in section 5. And
section 6 concludes the paper with future research efforts.
2. Overview of MANSI
MANSI is an on-demand multicast routing protocol that cre-
ates a multicast connection among group members by deter-
mining a set of intermediate nodes that serve as forwarding
nodes. This set, called a forwarding set, is shared among all
the senders of the group. The protocol exploits a core-based
technique where each member joins the group via the core
node in order to establish a connection with the other group
members. Unlike the core-based tree (CBT) protocol [2],
however, the core of each group is not statically assigned to a
particular node in the network and is not known in advance
by the members. Instead, the first member who becomes
an active source (i.e., starts sending data to the group) takes
the role of the core and announces its existence to the others
by flooding the network with a C ORE A NNOUNCE packet.
Each member node then relies on this announcement to
reactively establish initial connectivity by sending a J OIN
R EQUEST back to the core via the reverse path. Nodes who
receive a J OIN R EQUEST addressed to themselves become
forwarding nodes of the group and are responsible for accept-
ing and rebroadcasting non-duplicated data packets, regard-
less of which node the packets were received from. There-
fore, MANSI does not rely on any unicast routing protocol.
SHEN AND JAIKAEO
Figure 1. Examples of multicast connectivity among three group members:
(a) a forwarding set of six nodes formed by shortest paths from the core to the
other two members, and (b) another forwarding set when A partially shares
the same path to the core with B, which results in more efficient data packet
forwarding.
To maintain connectivity and allow new members to join, the
core floods C ORE A NNOUNCE periodically as long as there
are more data to be sent. As a result, these forwarding nodes
form a mesh structure that connects the group members to-
gether, while the core serves as a focal point for forwarding
set creation and maintenance. Since this process is performed
only when there is an active source sending data to the group,
we do not waste valuable network bandwidth to unnecessarily
maintain group connectivity in such dynamic environments.
Similar to other core-based protocols, this process creates
a forwarding set consisting of all the intermediate nodes on
the paths on which C ORE A NNOUNCEs are accepted and for-
warded from the core to the other members, which are often
shortest paths, as illustrated in figure 1(a). However, group
connectivity can be made more efficient by having A choose
another path that is partially shared by B to reduce the size
of the forwarding set, as shown in figure 1(b), which low-
ers the total cost of forwarding data packets. Note that the
cost is considered on a per-node basis, not per-link, due to the
fact that wireless communication is broadcast in nature (i.e.,
a single data packet broadcast by a node is expected to arrive
at all of its immediate neighbors in one transmission). In gen-
eral, the cost of the forwarding set does not always reflect the
number of nodes in the set. Instead, the cost associated with
each node can represent different measurements, depending
on the desired properties of the forwarding set. For instance,
if we aim to reduce the number of nodes in the forwarding
set for efficient data forwarding, the cost associated with each
node could be one. Table 1 lists a few more examples of what
node cost would represent when MANSI is applied to other
variations of the multicast routing problem in wireless ad hoc
networks.
We adopt the swarm intelligence metaphor to allow nodes
to learn a better multicast connection that yields a lower (to-
tal) forwarding cost. Each member who is not the core pe-
riodically deploys a small packet, called a F ORWARD A NT,
that opportunistically explores different, and hopefully better
paths toward the core. This exploring process is illustrated in
figure 2. If a F ORWARD A NT arrives at a node who is cur-
rently serving as a forwarding node for the group (node D in
this case), it turns itself into a BACKWARD A NT and travels
AD HOC MULTICAST ROUTING ALGORITHM
Table 1
A few variations of the multicast routing problem and how each node would
compute its cost in MANSI.
Problem Cost calculation per node
Load balancing Current traffic load or the current queue size
Power-aware routing Nodes transmission power
Energy conservation Inverse of the remaining energy of the node
Secure routing Security risk of the area the node is located in
Figure 2. Behavior of forward and backward ants: (1) a F ORWARD A NT
deployed from the member A choosing C as the next hop and encounter-
ing a forwarding node D, and (2) at D, the F ORWARD A NT becoming a
BACKWARD A NT and following the reverse path back to A while depositing
pheromone along the way.
back to its originator via the reverse path. When the BACK -
WARD A NT arrives at each intermediate node, it estimates the
cost of having the node it is currently at join the forwarding
set via the forwarding node it previously found. The com-
puted cost, as well as a pheromone amount that is inversely
proportional to the cost, are updated on the nodes local data
structure. These pheromone amounts are then used by sub-
sequent F ORWARD A NTs that arrive at this node to make a
decision which node they will travel to next, similar to how
pheromone is used by biological ants. Let us consider the
same example shown in figure 2, when the BACKWARD A NT
leaves D and arrives at C, the cost of having C join the for-
warding set via D is zero since D is already a forwarding
node and is directly connected to C. When the ant comes
back to A, the cost of having A join the forwarding set via
D is the same as the cost associated with C because C would
be required to become a forwarding node to allow A to join
the group via D. If A sees that the pheromone amount on the
link to C becomes the highest among links to all neighboring
nodes, it will switch to join the group via C by sending a J OIN
R EQUEST to C. Consequently, C will become a forwarding
node, while E, F and G will remove themselves from the
forwarding set (since they no longer hear requests from A),
which is similar to the connectivity shown in figure 1(b).
To prevent the race condition where members attempt to
establish group connectivity via one anothers forwarding
path and nobody remains connected to the core, each for-
warding node is associated with a height which is identical
to the highest ID of the nodes that use it to connect to the
core. In addition, the core has its height set to infinity. Fig-
ure 3 shows an example illustrating how heights are assigned
49
Figure 3. An example illustrating how heights are assigned to forwarding
nodes used by the members with IDs 3, 6 and 8.
to forwarding nodes. A F ORWARD A NT must stop and turn
into a BACKWARD A NT only when it encounters a forwarding
node whose height is higher than the ID of the member who
originated the ant. That means a member is allowed to con-
nect to the core via an existing path that belongs to another
member with a higher ID, but not vice versa, to assure that
the core, whose height is always the highest, will eventually
be connected to all the other members.
By following these simple rules, a majority of F ORWARD
A NTs from each member will choose a path that connects to
an existing forwarding node with a smaller total path cost.
Nodes on this path are then used to forward multicast data
packets, resulting in a lower data forwarding cost. This ex-
ploring and learning mechanism enables MANSI to learn a
better forwarding set for each group, depending on how node
cost is defined, as well as differentiates MANSI from other
existing ad hoc multicast routing protocols. Note that, by do-
ing so, MANSI attempts to evolve multicast connectivity into
states that yield lower cost. It, however, does not guarantee
that minimum-cost connectivity can be achieved.
3. MANSI protocol description
This section explains the operations of MANSI in details.
3.1. Local data structures
Each node in the network is assigned a unique ID. A node
with a unique ID i maintains a list of neighboring nodes,
ntab(i), obtained via a neighbor discovery protocol such as
periodic hello messaging. The node cost associated with
i is denoted by cost(i), where cost(i)
0, which should
be appropriately defined to reflect the performance metric
subject to minimization. In addition, for each multicast
group g, MANSI maintains the following data structures at
each node i.
Join table: maintains a list of nodes that have requested to
join a multicast group via node i. The join table of node i
for multicast group g is denoted by joing (i). This table is
updated when i hears a J OIN R EQUEST packet intended to
itself. Each entry in joing (i) is of the form r, hr , where
50 SHEN AND JAIKAEO

Figure 4. Sample network snapshots illustrating the operations of MANSI: (a) network setup with three members: nodes 1 (lower-left), 47 (upper-right), and
50 (upper-left), where node 1 is the core, (b) dissemination of C ORE A NNOUNCE indicated by arrows, (c) initial multicast connectivity using reverse paths to
the core, resulting in a forwarding set of ten nodes (shown in gray), and (d) forwarding set of four nodes learned by ants later in time.

r is a requesting nodes ID and hr is its height (as de- Height: represents the height of i if it is currently a mem-
scribed in section 2) that it has sent along with its J OIN ber or a forwarding node of the group g, defined as:
R EQUEST. The join table is initially empty for each node.

The node i becomes a forwarding node of the group g as if i = coreg (i),

 
long as joing (i) = . When a neighbor j is removed from

max i, max{hr | r, hr  joing (i)}
ntab(i) due to a link failure, i will remove all the corre- heightg (i) = if i is a member of group g,
sponding entries j, hj  from all the join tables.
 


Core ID: denoted by coreg (i) to indicate the current core max hr | r, hr  joing (i)

of group g. coreg (i) is initially set to INVALID_AD- otherwise.
(1)
DRESS. As described in section 2, the height of i is the highest ID
Core sequence number: keeps track of the latest C ORE of the nodes, including i itself if it is a member, that are
A NNOUNCEs sequence number, denoted by seqNog (i), using i to connect to the core, and the core has an infinite
and initially set to zero. height.
AD HOC MULTICAST ROUTING ALGORITHM
Pheromone table: maps neighboring nodes and heights
to pheromone intensities. For node i, the pheromone in-
tensity associated to the height h of the link (i, j ) for
the multicast group g is denoted by g (i, j, h), where
0  g (i, j, h)  1. This table is initially empty. Similar
to maintaining the join table, if a neighbor j is removed
from ntab(i), all entries g (i, j, h), g, h are removed as
well. The maximum pheromone intensity of one is de-
fined to prevent pheromone trails from being overly inten-
sified, which, therefore, gives ants enough probabilities to
explore different paths in a timely manner.
Best cost table: keeps track of how close node i thinks it
is to forwarding nodes of certain heights in terms of path
costs. The cost of the best path to any forwarding node of
height h for group g that i has seen so far is represented
by bestCostg (i, h). This best cost information is used to
determine whether a BACKWARD A NT has returned from
a good path or a bad path. Initially, this table is also empty.
3.2. Forwarding set initialization
Since MANSI is a reactive protocol, it does not send any
control packet out (except hello packets for neighbor dis-
covery) when there is no active source of multicast traf-
fic. When a member c of a group g has data to send and
it sees that the core does not exist for the group yet (i.e.,
coreg (c) = INVALID_ADDRESS), it sets coreg (c) to its
own ID and floods the network with a C ORE A NNOUNCE
packet to announce that it is becoming the core. The C ORE
A NNOUNCE contains the node ID, c, the multicast group ID,
g, a sequence number, and a cost which is initially set to
zero, as shown in figure 5. Upon receiving this C ORE A N -
NOUNCE , each node i discards the packet if it has seen an
announcement from the same node with the same sequence
number before, or if coreg (i) > c. This is to assure that du-
plicate C ORE A NNOUNCEs will not be processed, and only
one C ORE A NNOUNCE is allowed to be flooded if more than
one node are attempting to become the core and flooding their
C ORE A NNOUNCEs simultaneously. Algorithm 1 presents
the pseudo code of how node i processes a C ORE A NNOUNCE
packet. If the conditions are satisfied, i sets its coreg (i) to
c, increases the packets cost field by its own cost, then re-
broadcasts the packet. In addition, i updates the best cost
table, as well as the pheromone amount corresponding to
the height (i.e., the cores height) and the neighbor from
which the C ORE A NNOUNCE was received, by invoking the
procedure UpdatePheromoneAndCost shown in algorithm 2.
The operations of algorithm 2 will be explained later in de-
tails.
For every node i, coreg (i) is reset back to INVALID_
ADDRESS if it has not heard any C ORE A NNOUNCE within
the ANNOUNCE_INTERVAL time period. The core node
c also keeps sending out an announcement packet for group
g every ANNOUNCE_INTERVAL time period as long as
coreg (c) = c and it had at least one data packet for the group
to send within the last ANNOUNCE_INTERVAL time pe-
riod.
51
Figure 5. C ORE A NNOUNCE packet format.
Figure 6. J OIN R EQUEST packet format.
Algorithm 1. Node i processing a C ORE A NNOUNCE packet
1: Input:
2: announce incoming C ORE A NNOUNCE
3: lastHop the node from which announce was received
4: Begin:
5: g announce.group
6: if coreIdg (i) = INVALID_ADDRESS
OR coreIdg (i)  announce.core
OR seqNog (i)  announce.seqNo
7: Update local information:
coreIdg (i) announce.coreId
seqNog (i) announce.seqNo
8: Invoke UpdatePheromoneAnd
Costg (lastHop, , announce.cost, TRUE)
9: Update cost in the announcement packet:
announce.cost announce.cost + cost(i)
10: Rebroadcast announce
11: end if
As long as a member or a current forwarding node i of
group g keeps hearing C ORE A NNOUNCE from the core node,
i.e., coreg (i) = INVALID_ADDRESS, it periodically broad-
casts a J OIN R EQUEST packet to its neighbors. The J OIN
R EQUEST packet contains an entry g, k, heightg (i), where
k is defined as:
 g (i, n, h)
k = arg max . (2)
nntab(i) bestCostg (i, h) + 1
h>heightg (i)
The above formula implies that node i who is willing to join
a group should send a request to a neighbor whose good-
ness was recently confirmed by BACKWARD A NTs (i.e., hav-
ing high pheromone intensity) and also potentially yields the
lowest joining cost. In addition, node i only takes into ac-
count the best cost information and pheromone intensities of
heights greater than its own height since it is not allowed to
connect to an existing forwarding node of a smaller height, as
discussed in section 2. At this moment, however, no actual
ant packets are involved and each node has only one entry,
whose height is (i.e., the cores height), in each of its best
cost table and pheromone table. In other words, each node
has just enough information to establish a connection directly
to the core via the reverse path. As a result, the initial for-
warding set generally consists of all the nodes that are on the
(often times, shortest) paths on which the C ORE A NNOUNCE
are forwarded to the members. Figures 4(a)(c) illustrate the
forwarding set initialization process.
52
Algorithm 2. Procedure UpdatePheromoneAnd
Costg (next, height, cost, detFlag) executed by node i
1: Parameters:
2: next neighbor ID indicating
which pheromone table entry to be updated
3: height height associated with this update
4: cost cost of joining the group g at a forwarding
node of height height via next
5: detFlag flag indicating whether this
update is deterministic
6: Begin:
7: if g (i, next, height) is not defined then
8: g (i, next, height) 0
9: end if
10: if detFlag = TRUE then
11: bestCostg (i, height) cost
12: g (i, next, height) g (i, next, height)
+1/(2(1 + cost))
13: else
14: if bestCostg (i, height) is not defined OR
cost < bestCostg (i, height) then
15: bestCostg (i, height) cost
16: g (i, next, height) 1 /* set intensity to max */
17: else
18: g (i, next, height) g (i, next, height)
+1/(1 + cost)
19: end if
20: end if
21: g (i, next, height) min{g (i, next, height), 1)}
/* pheromone intensity is at most one */
If i is a member or a forwarding node belonging to more
than one group, it can combine multiple join entries into a
single J OIN R EQUEST packet, as shown in figure 6. When a
node j receives a J OIN R EQUEST from i and sees that its ID
is in the packet, it realizes that it should become a forward-
ing node for the group g. It then inserts the senders ID and
height in its join table joing (j ) and broadcasts its own J OIN
R EQUEST containing the ID of the next hop obtained by the
same formula above. Therefore, requests made by members
will eventually be propagated to the core, thus creating multi-
cast connectivity among all the members. On the other hand,
if node j hears a J OIN R EQUEST from i again without its ID,
or i is removed from ntab(j ) by neighbor discovery due to
a link failure, it removes i from its join table. Each node i
remains to serve as a forwarding node for group g as long as
joing (i) is not empty.
SHEN AND JAIKAEO
Figure 7. Ant packet format used by both F ORWARD A NT and BACKWARD
A NT.
3.3. Forwarding set evolution
Once the initial forwarding set is formed, each group member
who is not the core attempts to learn a better connection to the
core, in order to minimize the overall cost of the forwarding
set, by deploying a F ORWARD A NT every ANT_INTERVAL
time period. A F ORWARD A NT packet deployed by member i
for multicast group g, whose format is shown in figure 7, con-
tains the following fields:
group: multicast group ID.
height: height of the forwarding node found by this ant.
This field is used only after this ant has been turned to a
BACKWARD A NT.
f : forwarding flag indicating whether this ant is a F OR -
WARD A NT or a BACKWARD A NT (since they share the
same structure). Since i is deploying a F ORWARD A NT,
this flag is set to TRUE.
exLimit: the number of times the ant is allowed to proba-
bilistically pick a next hop that is not the current best one
in order to prevent it from aimlessly traversing the net-
work. This field is initially set to EXPLORE_LIMIT and
decrements every time the ant makes a decision on a next
hop probabilistically, instead of deterministically choosing
the next hop given by (2).
d: deterministic flag indicating whether the ant should al-
ways follow the current best path in order to obtain the
actual current cost for the best cost table. The reason for
using deterministic ants is that costs in the best cost table
may no longer reflect the actual costs due to node mobility,
dynamics of nodes costs, or dynamics of the forwarding
set itself. If this flag is set, the exLimit field is always
ignored. Every other ant deployed by each member is de-
terministic.
cost: the total cost of the nodes this ant has visited, initially
set to zero.
costLimit: the cost limit of the path that the ant is allowed
to traverses after leaving its originator. This field is used
in conjunction with the cost field to prevent the ant from
traversing forward after the accumulated cost exceeds the
limit. Usually this limit is set to minh>i bestCostg (i, h),
the lowest known cost to a current forwarding node of
group g that i is allowed to connect to, plus some thresh-
old. By this way, the ant can stop proceeding once it is
certain that it will not find any better path than what its
originator currently has. This cost limit is ignored if the
ant is deterministic since its goal is not to find a better
cost, but to find the actual current best cost.
visitedNodes: the set of nodes visited by the ant, initially
set to {i}.
AD HOC MULTICAST ROUTING ALGORITHM 53

Algorithm 3. Procedure ReleaseForwardAntg (fant) Algorithm 4. Node i processing a F ORWARD A NT packet.

executed by node i.
1: Input:
1: Parameter:
2: fant incoming F ORWARD A NT
2: fant a F ORWARD A NT to be released
3: Begin:
3: Begin:
4: if i = last entry in fant.visitedNodes then
4: Compute a desirability, dn , for node n, n ntab(i) from
5: g fant.group
summations of only entries whose heights are higher than
fant.height in the main pheromone table: 6: if joing (i) = AND fant.visitedNodes[0]

0 if n fant.visitedNodes, < heightg (i) then



g (i, n, h)
1+ 7: Convert fant to a BACKWARD A NT
dn = bestCostg (i, h) + 1 (3)


h>fant.height
fant.cost 0




if g (i, n, h) exists,
fant.height heightg (i)
1 otherwise.
5: if n, dn = 0 then fant.f FALSE

6: return /* ant has no place to go */ 8: Remove last entry from fant.visitedNodes

7: end if and broadcast fant

8: If fant is not deterministic (fant.d = FALSE) and it is 9: else

allowed to explore (fant.exLimit > 0), with probability
0.5, fant decides to randomly choose a next hop n, where
the probably of choosing n depends on its desirability as
follows:
dn
Prob(n) =  (4)
kntab(i) dk
fant.exLimit fant.exLimit 1
/* ant just performs one more exploration */
9: Otherwise, the next hop is set to the one whose
desirability is maximum:
n arg maxkntab(i) dk
10: append n to fant.visitedNodes and broadcast fant
A node deploying a F ORWARD A NT invokes the procedure
ReleaseForwardAnt described in algorithm 3 to find the next
hop that the ant will travel to. A desirability, defined in (3),
is computed for each the neighboring nodes by giving higher
values to neighbors that have higher pheromone intensities
and potentially yield lower costs to connect to an existing for-
warding node. On the other hand, zero desirability is given to
all the nodes that have been visited before. If the ant is not
deterministic and is still allowed to explore, these desirabil-
ities are then normalized to obtain a probability of choosing
each of the neighboring nodes. Otherwise the neighbor node
that gives the maximum desirability is chosen, which has the
same effect as using (2) except that it excludes all the nodes
in the visitedNodes field. Once a next hop is chosen, its ID
is appended to the end of visitedNodes and the ant is broad-
cast.
10: fant.cost fant.cost + cost(i)
11: if fant.cost < fant.costLimit OR
fant.d = TRUE then
12: Invoke ReleaseForwardAntg (fant)
13: end if
14: end if
15: end if
When a node j receives a F ORWARD A NT, it checks if its
ID matches the ID at the end of the ants visitedNodes field.
If not, the ant is discarded. Otherwise, j knows that this ant
is intended to itself and accepts it. Algorithm 4 shows how
a F ORWARD A NT is processed. First, j checks if it is cur-
rently a forwarding node of the group and its height is higher
than the ID of the ants originator. If so, j realizes that the
member who deployed the ant is eligible to join the group via
j itself. This ant is then turned into a BACKWARD A NT by
resetting its f flag. Its cost is then reset to zero in order to
start computing the total cost on the way back, and its height
field is set to j s height. The last entry of its visitedNodes
is removed in order to send this ant back to the previous
hop.
If the condition is not satisfied to convert the ant to a
BACKWARD A NT, j increases the ants cost field by its own
cost cost(j ). It then invokes the procedure ReleaseForward
Ant to forward the ant to a next hop, if the updated cost does
not exceed the limit or the ant is deterministic.
When a node k hears a BACKWARD A NT from j , it in-
vokes the procedure UpdatePheromoneAndCost, described in
algorithm 2, which updates the entries in ks pheromone and
54
best cost tables in accordance with j and the height field. If
the ant is deterministic, the cost that it carries back is the ac-
tual cost of the path its originator is currently using to join the
group. Therefore, the best cost corresponding to the height
field is updated to this value. If the ant is not deterministic,
however, the best cost is updated only when it is higher than
the returned cost, which means that the ant has found a better
path to join the group from this node. The pheromone inten-
sity on this link is also updated to the maximum in order to
encourage subsequent F ORWARD A NTs to use the same link,
as well as to redirect join request to this link instead. If the
ant comes back with a higher cost, a pheromone amount of
1/(1 + cost) is added instead. In case of deterministic ant,
the added amount is reduced by half since this link already
has the highest pheromone intensity as it has just been chosen
by a deterministic F ORWARD A NT. Note that we have men-
tioned this procedure before when we explained how a node
uses it while processing a C ORE A NNOUNCE (line 8 of algo-
rithm 1). This is because a C ORE A NNOUNCE more or less
serves as a deterministic BACKWARD A NT returning from the
core.
Algorithm 5. Node i processing a BACKWARD A NT packet.
1: Input:
2: bant incoming BACKWARD A NT
3: lastHop the node from which bant
was received
4: Begin:
5: g fant.group
6: Invoke UpdatePheromoneAnd
Costg (lastHop, bant.height, bant.cost, bant.d)
7: if i = last entry in fant.visitedNodes then
8: Remove the last entry from bant.visitedNodes
9: if bant.visitedNodes = then
10: bant.cost bant.cost + cost(i)
11: broadcast bant
12: end if
13: end if
After updating the pheromone and the best cost tables,
k checks if the BACKWARD A NT was intended to itself by
examining the last entry in the visitedNodes field. If its ID
matches, it adds its cost into the cost field, removes the last
entry from visitedNodes, and rebroadcasts as long as there is
at least one entry left in visitedNodes. Algorithm 5 presents
the pseudo code of how a node processes a BACKWARD A NT.
Similar to pheromone evaporation of biological ants, each
node i updates all the entries g (i, j, h) in its pheromone table
SHEN AND JAIKAEO
by reducing their values by DECAYING_FACTOR at every
DECAY_INTERVAL time period:
g (i, j, h) = (1 DECAYING_FACTOR)g (i, j, h), (5)
where 0 < DECAYING_FACTOR < 1.
By probabilistically selecting next hops, the majority of
the F ORWARD A NTs will choose paths with high pheromone
intensity, while some of them may explore totally different
new paths. If a BACKWARD A NT comes back with a better
cost on a new branch, the pheromone amount on that branch
will be increased significantly. As a result, a change in multi-
cast connectivity (i.e., forwarding set) is triggered due to the
periodic broadcast of J OIN R EQUEST packets, as illustrated
in figure 4(d).
MANSI also takes advantage of broadcast nature of wire-
less communication to speed up the learning process as
follows. When a node i overhears a J OIN R EQUEST for
group g from j but not intended to itself, it invokes Up-
datePheromoneAndCostg (j, hj , 0, TRUE), where hj is the
height that j reports in its J OIN R EQUEST. This implies
that i could join the group via j with no cost, given that its
height is less than hj . However, a drawback of this idea is that
some members who are not forwarding nodes will broadcast
J OIN R EQUESTs as well and might be mistaken as forwarding
nodes by its neighbors.
3.4. Multicast data forwarding
Since MANSI is a mesh-based protocol which allows for-
warding nodes and members to accept data packets arriving
from any node, each data packet is assigned a unique se-
quence number when it is transmitted from the source. The
sequence numbers are checked by each forwarding node and
member node to make sure that no duplicate data packets are
rebroadcast or delivered to the application. When a node i
receives a non-duplicate data packet of group g, it checks
whether it is currently a forwarding node of the group, i.e.,
joing (i) = . If so, it rebroadcasts the packet. Otherwise, the
packet is silently discarded.
3.5. Handling mobility
In MANSI, mobility and other network dynamics are handled
inherently rather than as exceptions. With the pheromone lay-
ing/following behavior of BACKWARD A NTs and F ORWARD
A NTs, each path comprising the forwarding set keeps being
reinforced as long as no link on the path is broken. How-
ever, network dynamics can cause optimal connectivity to
change from time to time even though the current connectiv-
ity may still be valid. With the probabilistic nature of F OR -
WARD A NT s to explore new paths, the multicast forwarding
set should be able to evolve into a configuration that is more
efficient for the new topology.
When a link currently used by a member or a forwarding
node to send J OIN R EQUESTs breaks, the pheromone table
entries corresponding to that link are also removed. There-
fore, all subsequent F ORWARD A NTs will be redirect to other
AD HOC MULTICAST ROUTING ALGORITHM
Figure 8. A network of 50 nodes moving at 10 m/s, where members are in black and forwarding nodes are in gray: (a) without mobility-adaptive mechanism,
and (b) with mobility-adaptive mechanism where NLFF_THRESHOLD is 0.01.
paths, while the majority of them will take the next hop whose
pheromone intensity was the second highest before the link
failure. If this next hop leads to a forwarding node of a higher
height, BACKWARD A NTs will return and update pheromone
on the new path, hence reestablishing a connection to the
group. However, in case that F ORWARD A NTs fail to find a
new path, C ORE A NNOUNCEs flooded periodically will even-
is defined as:
tually restore the connectivity.
Although MANSI is considered a mesh-based protocol by
= arg max . (8)
its way of forwarding data packets, connectivity of the for-
warding set may still be fragile if the network is sparse and
members are far apart from each other, especially with the
presence of mobility. To make data forwarding more effective
under mobility, while maintaining good efficiency when the
network is static, we incorporate a mobility-adaptive mecha-
nism into MANSI. With this mechanism, each node i keeps
track of the normalized link failure frequency, denoted by
nlff (i), which reflects the dynamic condition of the area sur-
rounding i in terms of the number of link failures per neigh-
bor per second. A calculation of nlff (i) is performed every
NLFF_TIME_WINDOW time period as follows:
current_nlff (i)
f
= , (6)
NLFF_TIME_WINDOW |ntab(i)|
current_nlff (i) + nlff (i)
nlff (i) = , (7)
2 QualNet simulator [10]. Ten random networks were gener-
where f is the number of link failures detected during the last
NLFF_TIME_WINDOW time period. Initially nlff (i) is set
to zero.
Each member or forwarding node then uses this nlff to de-
termine the stability of its surrounding area. If its nlff is lower
than a threshold NLFF_THRESHOLD, the node will consider
its area stable and join the group by sending J OIN R EQUESTs
55
toward its best next hop as usual. If nlff exceeds the threshold,
however, it will add another entry for the second best next hop
into its J OIN R EQUESTs. Since all the neighbors are ranked
by their goodness in terms of pheromone intensities, the sec-
ond best next hop can be easily determined. Formally, if k is
the best next hop for i to join the group g, as defined in (2),
then the second best next hop k
 g (i, n, h)
k
nntab(i),n=k h>height (i) bestCostg (i, h) + 1
g
Figure 8(a) illustrates the forwarding set created by MANSI
without the mobility-adaptive mechanism for a multicast
group of three members in a network of 50 nodes, where
each node is moving at 10 m/s. The group connectivity is
almost a straight line and is vulnerable to link failures. With
the mobility-adaptive mechanism enabled, most members and
forwarding nodes request two of their neighbors to be in the
forwarding set, as shown in figure 8(b), so that the group con-
nectivity becomes more robust.
4. Experimental results and discussion
To study the characteristics and evaluate the performance of
MANSI, we have conducted simulation experiments using the
ated with 50 nodes uniformly distributed over a terrain of
size 1000 1000 m2 . Each node was equipped with a ra-
dio transceiver which was capable of transmitting signals up
to approximately 250 meters over a 2 Mbps wireless chan-
nel, using the two-ray path loss model without fading. We
used IEEE 802.11DCF as the MAC layer protocol, and IP
as the network layer. Since MANSI does not rely on any
56 SHEN AND JAIKAEO

Table 2 Table 3
Parameter values for MANSI. Average size of the forwarding set formed in
MANSI, CORE, and FLOOD for each network.
HELLO_INTERVAL 1 sec
ANNOUNCE_INTERVAL 10 sec Network Average size
ANT_INTERVAL 2 sec MANSI CORE FLOOD
EXPLORE_LIMIT 3
DECAY_INTERVAL 1 sec 1 7.89 9.49 50.00
DECAYING_FACTOR 0.1 2 4.00 3.67 50.00
NLFF_THRESHOLD 0.01 3 4.00 4.97 50.00
4 4.46 4.68 50.00
5 6.51 8.46 50.00
6 5.52 6.25 50.00
7 6.90 7.83 50.00
8 6.04 7.46 50.00
9 5.16 7.67 50.00
10 5.02 6.95 50.00
Average 5.55 6.74 50.00

Figure 9. Average size of the forwarding set as a function of time for CORE
and MANSI.
unicast routing protocol, no other routing protocols were em-
ployed. For each network, a multicast groups of 5 members
was setup, where each member generated a constant bit rate
(CBR) traffic at 2 packets/sec to the group for 20 minutes.
The size of data payload was 512 bytes. The MANSI parame-
ter values used in our simulation are shown in table 2. Note
that NLFF_THRESHOLD is used only when the mobility-
adaptive mechanism is enabled.
Our first set of experiments were setup without mobility in
order to study how MANSI maintains forwarding sets in static
environments. For comparison purposes, we used two base-
line protocols: FLOOD and CORE, as references. FLOOD is
a simple flooding protocol where a data packet is rebroadcast
by every node in the network. And CORE is a generic core-
based protocol that operates exactly like MANSI, but with
no ants deployed, where C ORE A NNOUNCEs are periodically
flooded as usual. The cost of each node was set to one, which
implies that MANSI would attempt to reduce the size of the
forwarding set.
We first look at the average size of forwarding sets main-
tained by CORE and MANSI over time for the ten sample
networks, as shown in figure 9. Due to random delays added
to avoid packet collisions when broadcasting, the dissemina-
tion pattern of a C ORE A NNOUNCE is unpredictable when it
is flooded, which causes a forwarding set to be formed dif-
ferently for each announcement. Consequently, the average
size of forwarding sets keeps changing from time to time in
CORE. In contrast, forwarding sets maintained by MANSI
start of at around the same size as that of CORE but keep re-
ducing in size during the first 200 seconds. Their size then
becomes stable and stays low most of the time as each mem-
ber or forwarding node tends to join the group via a low-cost
path (i.e., small hop count in this case), whose existence was
recently confirmed by BACKWARD A NTs. Although another
C ORE A NNOUNCE may arrive at a member from a differ-
ent node, the member will not send a J OIN R EQUEST to this
new node as long as the current joining cost is low and the
pheromone intensity on the link it currently uses to join the
group is high.
Table 3 summarizes the sizes, averaged over the entire sim-
ulation time, of the forwarding sets maintained by MANSI,
CORE, and FLOOD on each simulated network. (FLOOD
does not really maintain a forwarding set, but the set con-
sists of every node in the network.) The results show that in
all cases, except one, MANSI yields forwarding sets that are
approximately 15%20% smaller than those of CORE, and
much smaller than FLOOD. Since the size of the forward-
ing set indicates how many nodes are involved to relay a data
packet from one member to the others, this demonstrates the
efficiency of MANSI in terms of data forwarding.
We have performed another set of experiments to compare
the performance of MANSI, in terms of effectiveness and ef-
ficiency, with ODMRP. ODMRP [1] is an on-demand, mesh-
based multicast protocol that attempts to establish a forward-
ing group similar to a forward set in MANSI only when
a source of the group has data to send. The nodes in the for-
warding group form a mesh that connects the group members
together. When a multicast source has data to send for the
first time, it broadcasts to its neighbors a J OIN Q UERY packet,
which is a data packet with the query flag set. Upon receiving
a non-duplicate J OIN Q UERY, each node stores the upstream
node ID in its routing table and rebroadcasts the packet. When
a member of the multicast group receives a J OIN Q UERY, it
constructs and broadcasts a J OIN R EPLY packet containing
the source ID and the upstream node ID to all of its neighbors.
Upon receiving a J OIN R EPLY, a node whose ID matches the
upstream ID in the packet realizes that it is on the path be-
tween the source and a member, so it becomes a forward-
ing node for the group by setting its FG_FLAG (Forwarding
Group Flag). It then constructs and broadcasts its own J OIN
AD HOC MULTICAST ROUTING ALGORITHM 57

R EPLY using its corresponding upstream node ID. The broad-

casting of J OIN R EPLY packets therefore propagates the infor-
mation from all the members back to the source on the reverse
paths. Once the source has sent out a J OIN Q UERY, it sends
all subsequent data packets normally with no query flag set.
This will allow only nodes that are currently in the forward-
ing group to rebroadcast these data packets, thus reducing
data forwarding overhead. To deal with dynamics of the net-
work topology and group membership, each source floods the
network with J OIN Q UERYs every REFRESH_INTERVAL as
long as it still has data to be sent to the group. The FG_FLAG
on each node will be reset if it has not been refreshed by J OIN
R EPLY for some period of time, which implies that the source
has no data to send, or it is no longer needed as a forwarding
node. If nodes are equipped with GPS, a mobility prediction
method can also be used to adaptively adjust the value of RE- Figure 10. Packet delivery ratio as a function of mobility speed.
FRESH_INTERVAL to suit the current mobility condition.
In this comparison, we used QualNets implementation
of ODMRP, which followed the specification in the Inter-
net Draft draft-ietf-manet-odmrp-02.txt [7] but
without mobility prediction which requires GPS. The value
of REFRESH_INTERVAL was fixed at 3 seconds. Each node
moved constantly with the predefined speed, which was var-
ied from 0 m/s to 20 m/s. The following statistics were col-
lected and used in the comparison, where each measurement
will be shown with a 95% confidence interval:
Packet delivery ratio. The ratio of the number of non-
duplicate data packets successfully delivered to the re-
ceivers versus the number of packets supposed to be re-
ceived. This metric reflects the effectiveness of a protocol.
Number of total packets transmitted per data packet re-
ceived. The ratio of the number of data and control packets
transmitted versus the number of data packets successfully Figure 11. Total packets transmitted per data packet received at the destina-
tions as a function of mobility speed.
delivered to the application. HELLO packets are also con-
sidered as packets transmitted. This measure shows effi-
ciency of a protocol in terms of channel access. The lower and FLOOD. Although the delivery ratio is a bit lower than
the number, the more efficient the protocol. that of the other two protocols, more than 90% of data packets
can be delivered at every mobility speed.
Number of total bytes transmitted per data byte received.
In terms of efficiency, both MANSI-Basic and MANSI-
This metric is similar to the second metric except that
Mobile give significantly better performance than ODMRP
number of bytes is considered instead. Here, bytes trans-
and FLOOD at low mobility in both channel access and
mitted include everything that is sent to the MAC layer
bandwidth utilization aspects, as shown in figures 11 and
(i.e., IP and UDP headers, as well as HELLO packets),
12, respectively. The reason is that every multicast sender
where data bytes received involve only the data payloads.
floods J OIN Q UERY packets periodically in ODMRP, while in
This metric presents efficiency of a protocol in terms of
MANSI, only the core of the group performs periodic flood-
bandwidth utilization. Similar to the second metric, the
ing. Moreover, ODMRP requires each member to send a J OIN
lower the number, the more efficient the protocol.
R EPLY toward each sender via the reverse path on which the
Figure 10 presents packet delivery ratio of the protocols J OIN Q UERY was received, resulting in a fairly large forward-
at different mobility speeds. MANSI without the mobility- ing group, especially with high number of senders. MANSI,
adaptive mechanism, denoted by MANSI-Basic, shows sig- in contrast, has each member establish connectivity toward
nificant performance degradation as mobility increases due to the core, which keeps the number of forwarding nodes low.
the fact that the forwarding set lacks redundant paths when Without adapting their behaviors to mobility, data forward-
each member and forwarding node always requests only one ing characteristics of MANSI-Basic, ODMRP, and FLOOD
of its neighbor to be part of the forwarding set. However, remain almost the same regardless of mobility speeds, where
when the mobility-adaptive mechanism is enabled, as denoted FLOOD employs the highest number of forwarding nodes,
by MANSI-Mobile, its results are comparable with ODMRP and MANSI-Basic uses the least number, but suffers low
58
Figure 12. Total bytes transmitted per data byte received at the destinations
as a function of mobility speed.
packet delivery ratio under high mobility. With its mobility-
adaptive mechanism, MANSI-Mobile is shown to perform as
efficiently as MANSI-Basic at low mobility2 and as ODMRP
at high mobility, while yielding consistently high packet de-
livery ratio for the entire range of speeds.
5. Related work
There have been numerous multicast routing protocols pro-
posed for ad hoc networks. Protocols such as AMRoute [8],
AMRIS [12], and MAODV [9] are based on constructing
a tree spanning all the group members, where a node can
only accept packets coming from a node with which a tree
branch has been established. Since a tree structure provides
only one forwarding path between a pair of sender and re-
ceiver, group connectivity may suffer from frequent topology
changes in dynamic networking environments. Other proto-
cols such as CAMP [5] and ODMRP [1], including MANSI,
employ a mesh-based approach to increase redundancy by al-
lowing packets to be forwarded over more than one path, thus
giving a higher chance of successful delivery.
Based on the way they establish and maintain connec-
tivity within each multicast group, multicast protocols can
also be broadly classified as either a source-based approach
or a group-shared tree/mesh approach. In a source-based
approach, a multicast tree or mesh is constructed for each
sender. The construction process is usually initiated by a
sender that floods a request message to all other nodes in
the network so that the other members of the group can es-
tablish connectivity via the reverse paths. In ODMRP, each
sender also exploits periodic flooding of control packets to
refresh group connectivity and handle mobility. This is suit-
able with dense multicast groups but yields high overhead as
the network size and the number of senders increase. In con-
trast, a group-shared tree/mesh approach aims to construct a
tree/mesh for each multicast group, which is shared by all
2 In fact, they behave exactly the same.
SHEN AND JAIKAEO
senders within the group. A common technique for creating
group connectivity is to designate a node in the network the
role of the rendezvous point, or the core, for each group. Each
member then establishes connectivity, often via the shortest
path, to the core, which in turn connects all the group mem-
bers together. For each group, one of the members, the first
sender, or any node in the network can take the role of the
core. Examples of ad hoc multicast protocols that are based
on this technique are MAODV, AMRIS, and CAMP. In con-
trast, in MANSI, group connectivity can be made more ef-
ficient by having some members share common paths to the
core with other members in order to further reduce the to-
tal cost of forwarding data packets. Moreover, the forwarding
cost may adopt different performance metrics for different ob-
jectives, in general, in addition to the number of nodes used
to forward data. MANSI extends a core-based technique by
adopting the metaphor of swarm intelligence to learn a better
multicast connection that yields a lower total forwarding cost.
Swarm intelligence appears in biological swarms of cer-
tain insect species. It gives rise to complex and often intel-
ligent behavior through simple, unsupervised interactions be-
tween a sheer number of autonomous swarm members. The
end result is the emergence of very complex forms of so-
cial behavior which fulfill a number of optimization objec-
tives and other tasks. Its metaphor has been applied to many
combinatorial optimization problems like the traveling sales-
man problem (TSP) and the quadratic assignment problem
(QAP). In communications networks, a number of routing and
load balancing mechanisms based on swarm intelligence have
been proposed. Ant-Based Control (ABC) [11] has applied
swarm intelligence to achieve load balancing in telecommu-
nications networks. Simulated on a model of the British Tele-
com (BT) telephone network, ABC has been shown to result
in fewer call failures than other methods such as shortest-
path routing. In [3,4], a distributed adaptive routing for data-
gram networks, called AntNet, has been described. Several
variations of AntNet have been developed but all of them
rely on the same concept where forward ants are launched
toward destinations and backward ants travel back and up-
date pheromone along the backward paths. The amount of
added pheromone is proportional to the goodness of the path
measured by the forward ant. The same concept has been
extended and applied to Adaptive Swarm-based Distributed
Routing (Adaptive-SDR) [6] for routing in wireless and satel-
lite networks, which incorporates a mechanism to cluster
nodes into colonies so as to resolve the scalability issue in
large networks.
We exploit the concept of forward and backward ant de-
ployment in the MANSI protocol to provide multicast support
for ad hoc networks. Within a multicast group, each member
launches a forward ant in order to find an existing forwarding
node where it can use to establish connectivity to the group
with lower cost. Once such a node is found, the forward ant
turns into a backward ant and returns to its origin via the re-
verse path, while depositing pheromone along the way to at-
tract more future forward ants. To our best knowledge, no ad
AD HOC MULTICAST ROUTING ALGORITHM
hoc multicast routing protocol has been proposed to exploit
the concept of swarm intelligence.
6. Conclusion and future work
Inspired by swarm intelligence, we have introduced an alter-
native approach to solving the multicast routing problem in
mobile ad hoc networks. Our protocol, called MANSI (Mul-
ticast for Ad hoc Networks with Swarm Intelligence), is an
on-demand multicast routing protocol that creates a multicast
mesh shared by all the members within each group. The pro-
tocol uses a core-based scheme, where each member initiates
a request to the core node to establish multicast connectivity
with other members. Intermediate nodes who receive such a
request become forwarding nodes that are used to relay data
packets from one member to the others. Unlike other core-
based protocols, MANSI does not always rely on the shortest
paths between the core and the members to establish group
connectivity. Instead, each member who is not the core peri-
odically deploys a small packet that behaves like an ant to op-
portunistically explore different paths. This exploring mech-
anism enables the protocol to discover paths that comprise a
better set of forwarding nodes yielding a lower total cost of
data forwarding, where the cost of forwarding (nodes) can
be defined in terms of different application specific perfor-
mance metrics. MANSI also incorporates a mobility-adaptive
mechanism that allows the protocol to remain effective as
mobility increases. The simulation results have shown that
MANSI performs both effectively and efficiently in static or
low-mobility environments, yet still effectively in highly dy-
namic environments.
Research is in progress to apply MANSI with other objec-
tives such as load balancing, energy conservation, and secu-
rity.
References
[1] S. Bae, S. Lee, W. Su and M. Gerla, The design, implementation, and
performance evaluation of the on-demand multicast routing protocol in
multihop wireless networks, IEEE Network (Special Issue on Multicas-
ting Empowering the Next Generation Internet) 14(1) (2000) 7077.
[2] T. Ballardie, P. Francis and J. Crowcroft, Core-based trees (CBT): An
architecture for scalable inter-domain multicast routing, in: Commu-
nications, Architectures, Protocols, and Applications, San Francisco,
CA, USA (1317 September 1993).
[3] G. Di Caro and M. Dorigo, AntNet: A mobile agents approach to adap-
tive routing, Technical Report IRIDIA/97-12, Universit Libre de Brux-
elles, Belgium (1997).
[4] G. Di Caro and M. Dorigo, Two ant colony algorithms for best-effort
routing datagram networks, in: Tenth IASTED Internat. Conf. on Par-
59
allel and Distributed Computing and Systems (PDCS98), Las Vegas,
NV (2831 October 1998).
[5] J. Garcia-Luna-Aceves and E. Madruga, The core-assisted mesh proto-
col, IEEE Journal on Selected Areas in Communications 17(8) (1999).
[6] I.N. Kassabalidis, M.A. El-Sharkawi, R.J. Marks II, P. Arabshahi and
A.A. Gray, Adaptive-SDR: Adaptive swarm-based distributed routing,
in: IEEE WCCI 2002, IJCNN 2002 Special Session: Intelligent Sig-
nal Processing for Wireless Communications, Honolulu, Hawaii (12
17 May 2002).
[7] S.-J. Lee, W. Su and M. Gerla, On-demand multicast routing
protocol (ODMRP) for ad hoc networks, IETF Internet Draft,
http://www.ietf.org/proceedings/00jul/I-D/manet
-odmrp-02.txt (2000).
[8] M. Liu, R.R. Talpade and A. McAuley, AMRoute: Adhoc multicast
routing protocol, Technical Report 99, The Institute for Systems Re-
search, University of Maryland (1999).
[9] R. Royer and C. Perkins, Multicast using ad-hoc on demand distance
vector routing, in: MOBICOM99, Seattle, WA (August 1999) pp. 207
218.
[10] Scalable Network Technologies, QualNet Simulator, http://www.
scalable-networks.com.
[11] R. Schoonderwoerd, O. Holland, J. Bruten and L. Rothkrantz, Ant-
based load balancing in telecommunications networks, Technical Re-
port HPL-96-76, Hewlett-Packart Laboraties Bristol, Bristol, UK
(21 May 1996).
[12] C.W. Wu and Y.C. Tay, AMRIS: A multicast protocol for ad hoc wire-
less networks, in: IEEE Military Communications Conf. (MILCOM),
Atlantic City, NJ (November 1999) pp. 2529.
Chien-Chung Shen received his B.S. and M.S. de-
grees from National Chiao Tung University, Taiwan,
and his Ph.D. degree from UCLA, all in computer
science. He was a research scientist at Bellcore Ap-
plied Research working on control and management
of broadband networks. He is now an assistant pro-
fessor in the Department of Computer and Informa-
tion Sciences of the University of Delaware, and a
recipient of NSF CAREER Award. His research in-
terests include ad hoc and sensor networks, control
and management of broadband networks, distributed object and peer-to-peer
computing, and simulation.
E-mail: cshen@cis.udel.edu
Chaiporn Jaikaeo received his B.Eng. degree in
computer engineering from Kasetsart University,
Bangkok, Thailand, in 1996, and his M.S. and Ph.D.
degrees in computer and information sciences from
the University of Delaware in 1999 and 2004, re-
spectively. He is now a faculty member in the De-
partment of Computer Engineering at Kasetsart Uni-
versity, Bangkok, Thailand. His research interests
include unicast and multicast routing, topology con-
trol, peer-to-peer computing, and network manage-
ment for ad hoc and sensor networks.
E-mail: fengchj@ku.ac.th
 Mobile Networks and Applications 10, 6177, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Regional Gossip Routing for Wireless Ad Hoc Networks

XIANG-YANG LI, KOUSHA MOAVENINEJAD and OPHIR FRIEDER
Department of Computer Science, Illinois Institute of Technology, Chicago, IL 60616, USA

Abstract. Many routing protocols have been proposed for wireless ad hoc networks, and most of them are based on some variants of
flooding. Thus many routing messages are propagated through the network unnecessarily despite various optimizations. Gossip based
routing method has been used and re-investigated to reduce the number of messages in both wired networks and wireless ad hoc networks.
However, the global gossiping still generates many unnecessary messages in the area that could be far away from the line between sender
node and receiver node. We propose a regional gossip approach, where only the nodes within some region forward a message with some
probability, to reduce the overhead of the route discovery in the network. We show how to set the forwarding probability based on the region
and the network density both by theoretical analysis and by extensive simulations. Our simulations show that the number of messages
generated using this approach is much less than the simple global gossiping method, which already saves many messages compared with
global flooding. We expect that the improvement should be even more significant in larger networks.
Keywords: gossip, fault tolerance, routing, wireless ad hoc networks

1. Introduction recent years, a variety of routing protocols [16,1832], tar-

Recent years saw a great amount of research in wireless net-
works, especially ad hoc wireless networks due to its po-
tential applications in various situations such as battlefield,
emergency relief, and so on. There are no wired infrastruc-
tures or cellular networks in ad hoc wireless network. Two
nodes can communicate directly if they are within the trans-
mission range of the other. Otherwise, they communicate
through multi-hop wireless links by using intermediate nodes
to relay the message. Consequently, each node in the wire-
less network also acts as a router, forwarding data packets
for other nodes. In addition, we assume that each node has
a low-power Global Position System (GPS) receiver, which
provides the position information of the node itself. If GPS
is not available, the distance between neighboring nodes can
be estimated on the basis of incoming signal strengths and
the direction of arrival. Relative co-ordinates of neighboring
nodes can be obtained by exchanging such information be-
tween neighbors [1].
The devices in the wireless ad hoc networks are often pow-
ered by batteries only. Thus, the power supply is limited and
it is often difficult to recharge the batteries, which motivates
many researches in designing power efficient protocols for
power assignment [27], topology control [814] and rout-
ing [1517]. In addition, the bandwidth available is much
less compared with the wired networks counterpart due to its
unique transmission characteristics. Moreover, since nodes
can be mobile, routes may constantly change. Thus, the de-
signed routing protocols for wireless ad hoc networks should
use as less messages as possible, which will reduce power
consumption (thus enlong network life), and signal interfer-
ence (thus increase the throughput).
One of the key challenges in the design of ad hoc networks
is the development of dynamic routing protocols that can ef-
ficiently find routes between two communication nodes. In
geted specifically for ad hoc environment, have been devel-
oped. For the review of the state of the art of routing pro-
tocols, see surveys by Royer and Toh [33], by Ramanathan
and Steenstrup [34], and by Mauve et al. [35]. Some routing
protocols assume that the each node knows its own positions
(e.g., equipped with GPS receivers). These category of pro-
tocols are called Location-Aided Routing (LAR) protocols in
which the overhead of route discovery is decreased by uti-
lizing location information. Some protocols do not rely on
position information, and make use flooding (or some vari-
ants of flooding). Thus many routing messages are propa-
gated through the network unnecessarily despite possible var-
ious optimizations. Gossip based routing method has been
used and re-investigated to reduce the number of messages in
both wired networks and wireless ad hoc networks. When-
ever a node receives a message, it tosses a coin to decide
whether to forward a message or not in order to reduce the
total number of routing messages sent by all nodes. However,
the global gossiping still generates many unnecessary mes-
sages in the area that could be far away from the line between
sender node and receiver node. We propose a regional gossip
approach, where only the nodes within some region forward
a message with some probability, to reduce the overhead of
route discovery in the network.
The key observation for all gossiping based routing meth-
ods is that the gossiping exhibits a bimodal behavior, which
is well-known in the percolation theory [36,37]. This can
be rephrased as follows. Let p be the uniform probability
that a node will forward the routing message to its neighbors.
Then, there is a threshold value p0 such that, in sufficiently
large random networks, the gossip message quickly dies out
if p < p0 (p is slightly less than p0 ) and the gossip message
spreads to all network nodes if p > p0 (p is slightly greater
than p0 ). In other words, in almost all executions, either al-
62
most no node receives the message or almost all of them do.
So ideally, we would set the gossiping probability to some
value slightly larger than p0 to reduce the routing messages
propagated. When the network is sufficiently large, we can set
p sufficiently close to p0 , thus save about (1 p0 )n messages
overhead compared with the flooding, since about p0 n nodes
will forward the message in gossiping based method com-
pared with n nodes forwarding in flooding. Hass et al. [24]
conducted extensive simulations to investigate the extent to
which this gossiping probability can be lowered. They found
that using gossiping probability between 0.6 and 0.8 suffices
to ensure that almost every node gets the message in almost
every routing. They report of up to 35% fewer messages than
flooding (close to our previous explanation). Notice that their
experimental setting of the network has some special config-
urations [24].
Although gossiping reduces the routing messages com-
pared with flooding, it still produces lots of unnecessary mes-
sages in regions that are far from the line between sender
node and receiver node. Notice that, the traditional gossip
will propagate the message to the whole network. To further
reduce the number of forwarding messages, we propose re-
gional gossiping, in which essentially only nodes inside some
region (derived from the source and target) will execute the
gossiping protocol, and nodes outside the region will not par-
ticipate in the gossiping at all. The region we select in our
simulations are some ellipses using the source and target as
foci. Notice that here we assume source node knows either
the exact or the approximate location of the destination node,
we will discuss this later in section 2 in detail. We also dy-
namically adjust the forwarding probability based on the node
density estimated by the current node. Our results show that,
by using appropriate optimization heuristics, we can save up
to 94% messages even compared with the global flooding
method.
The remaining of this paper is organized as follows. In sec-
tion 2, we review some known location services techniques
for wireless ad hoc networks. We study our regional gossip
method in detail in section 3. We demonstrate its effective-
ness by both theoretical study and extensive simulations in
section 4 . We also study the effectiveness of the regional gos-
siping on constructing multiple paths for any pair of source
and destination nodes in section 5. We conclude our paper
and discuss possible future research directions in section 6.
2. Preliminaries
We consider a wireless ad hoc network (or sensor network)
with all nodes distributed in a two-dimensional plane. As-
sume that all wireless nodes have distinctive identities and
each static wireless node knows its position information1 ei-
ther through a low-power Global Position System (GPS) re-
ceiver or through some other way. For simplicity, we also as-
1 More specifically, it is enough for our protocol when each node knows the
relative position of its one-hop neighbors. The relative position of neigh-
bors can be estimated by the direction of arrival and the strength of signal.
LI ET AL.
sume that all wireless nodes have the same maximum trans-
mission range and we normalize it to one unit. Throughout
this paper, a broadcast by a node u means that u sends the
message to all nodes within its transmission range. Notice
that, in wireless ad hoc networks, the radio signal sent out
by a node u can be received by all nodes within the trans-
mission range of u. The main communication cost in wire-
less networks is to send out the signal while the receiving and
processing costs of a message is neglected here.
2.1. Location service
Several proposed routing algorithms [18,22] assume that the
source node knows the position information (or approximate
position) of the destination node. Our regional gossip method
also assumes that the source node knows the current posi-
tion information of the target approximately. Notice that, for
sensor networks collecting data, the destination node is often
fixed, thus, location service is not needed in those applica-
tions. However, the help of a location service is needed in
most application scenarios. Mobile nodes register their loca-
tions to the location service. When a source node does not
know the position of the destination node, it queries the loca-
tion service to get that information. In cellular networks, there
are dedicated position severs. It will be difficult to implement
the centralized approach of location services in wireless ad-
hoc networks. First, for centralized approach, each node has
to know the position of the node that provides the location
services, which is a chicken-and-egg problem. Second, the
dynamic nature of the wireless ad hoc networks makes it very
unlikely that there is at least one location server available for
each node. Thus, we will concentrate on distributed location
services.
For the wireless ad hoc networks, the location service pro-
vided can be classified into four categorizes: some-for-all,
some-for-some, all-for-some, all-for-all. Some-for-all service
means that some wireless nodes provide location services for
all wireless nodes. Other categorizations are defined simi-
larly.
An example of all-for-all services is the location services
provided in the Distance Routing Effect Algorithm for Mo-
bility (DREAM) by Basagni et al. [38]. Each node stores a
database of the position information for all other nodes in the
wireless networks. Each node will regularly flood packets
containing its position to all other nodes. A frequency of the
flooding and the range of the flooding is used as a control of
the cost of updating and the accuracy of the database.
Using the idea of quorum developed in the databases and
distributed systems, Hass and Liang [39] and Stojmenovic
[40] developed quorum based location services for wireless
ad-hoc networks. Given a set of wireless nodes V , a quorum
system is a set of subset (Q1 , Q2 , . . . , Qk ) of nodes whose
union is V . These subsets could be mutually disjoint or often
have equal number of intersections. When one of the nodes
requires the information of the other, it suffices to query one
node (called the representative node of Qi ) from each quo-
rum Qi . A virtual backbone is often constructed between
REGIONAL GOSSIP ROUTING
the representative nodes using a non-position-based methods
such as [4144]. The updated information of a node v is sent
to the representative node (or the nearest if there are many) of
the quorum containing v. The difficulty of using quorum is
that the mobility of the nodes requires the frequent updating
of the quorums. The quorum based location service is often
some-for-some type.
The other promising location service is based on the
quadtree partition of the two-dimensional space [45]. It di-
vides the region containing the wireless network into hierar-
chy of squares. The partition of the space in [45] is uniform.
However, we notice that the partition could be non-uniform
if the density of the wireless nodes is not uniform for some
applications. Each node v will have the position information
of all nodes within the same smallest square containing v.
This position information of v is also propagated to up-layer
squares by storing it in the node with the nearest identity to v
in each up-layer square containing v. Using the nearest iden-
tity over the smallest identity, we can avoid the overload of
some nodes. The query is conducted accordingly. It is easy to
show that it takes about O(log n) time to update the location
of v and to query another nodes position information.
If the location service is not provided, the nodes can cache
the location information of some other nodes. When the
source node wants to send a message to the target, it directly
uses the region gossip if the target location is known. Oth-
erwise, it will use flooding (with selective forwarding [46]
to control the number of messages sent) to send the message
to all nodes within k hops, where k is a parameter to be set.
Then if a node within k hops knows the destination location,
that node then starts the regional gossip to send message to
the destination.
2.2. Random deployment and connectivity
Energy conservation is critical for the life of the wireless net-
work. One approach to save energy is to use the minimum
power to transmit the signal without disconnecting the net-
work. The universal minimum power used by all wireless
nodes, such that the induced network topology is connected,
is called the critical power. Determining the critical power
for static wireless ad hoc networks is well-studied [5,7,13]. It
remains to study the critical power for connectivity for mo-
bile wireless networks. As the wireless nodes move around, it
is impossible to have a unanimous critical power to guarantee
the connectivity for all instances of the network configuration.
Thus, we need to find a critical power, if possible, at which
each node has to transmit to guarantee the connectivity of the
network almost surely, i.e., with high probability almost one.
The wireless nodes are randomly deployed in majority
wireless ad hoc networks either due to its massive number,
due to its emergency requirement, or due to harsh environ-
ment. For simplicity, we assume that the n wireless devices
are distributed in a unit area square (or disk) according to
some distribution function, e.g., random uniform distribution,
denoted by Xn , or Poisson process, denoted by Pn .
63
Let G(V , r) be the graph defined on V with edges uv E
if and only if uv
r where uv is the Euclidean dis-
tance between nodes u and v. Let G
(Xn , rn ) be the set of
graphs G(V , rn ) for n nodes V that are uniformly and in-
dependently distributed in a two-dimensional region
. The
problem considered by Gupta and Kumar [5] is then to deter-
mine the value of rn such that a random graph in G
(Xn , rn )
is asymptotically connected with probability one as n goes to
infinity, when
is a unit disk. Specifically, they showed that
G(V , rn ) is connected almost surely if nrn2  ln n + c(n)
for any c(n) with c(n) as n goes to infinity, and
G(Xn , rn ) is asymptotically disconnected with positive prob-
ability if nrn2 = ln n + c(n) and lim supn c(n) < +. It is
unknown whether the same result holds if the geometry do-
main in which the wireless nodes are distributed is a unit-area
square instead of the unit-area disk.
Independently, Penrose [47] showed that the longest edge
Mn of the minimum spanning tree of n points randomly and
uniformly distributed in a unit area square C satisfies that


lim Pr nMn2 ln n
= ee ,
n
for any real number . This result gives the probability of the
network to be connected if the transmission radius is set as a
positive real number r when n goes to infinity. For example,
if we set = ln ln n, we have


Pr nMn2
ln n + ln ln n = e1/ ln n .
It implies that the network is connected with probability at
least e1/ ln n if the transmission radius rn satisfies nrn2 =
ln n + ln ln n. Notice that e1/ ln n > 1 1/ ln n from ex >
1 x for x > 0. By setting = ln n, the probability that
the graph G(V , rn ) is connected is at least e1/n > 1 1/n,
where nrn2 = 2 ln n. Notice that the above probability is
only true when n goes to infinity. When n is a finite number,
then the probability of the graph being connected is smaller.
In [48], Li et al. presented the experimental study of the prob-
ability of the graph G(V , rn ) being connected for finite num-
ber n.
Gupta and Kumar [5] conjectured that if every node has
probability p of being fault, then the transmission range for
resulting a connected graph satisfies prn2 = log n/n. This
was recently confirmed by Wan et al. [49]. It is not difficult
to see that whether the global gossip can deliver the packet
is related to whether a set of randomly deployed nodes in a
region form a connected graph when each node has a uniform
faulting probability p. Consequently, given a wireless net-
work with n nodes , each with transmission range r, the relay
probability of a gossip routing protocol is p = log n/(nrn2 ),
when n goes to infinity. We conjecture that this is true for any
non-flat convex region
.
2.3. Fault tolerance and security
Fault tolerance is one of the central challenges in designing
the wireless ad hoc networks. To make fault tolerance pos-
sible, first of all, the underlying network topology must have
64
multiple disjoint paths to connect any two given wireless de-
vices. Here the path could be vertex disjoint or edge disjoint.
Considering the communication nature of the wireless net-
works, the vertex disjoint multiple paths are often used in the
literature. A graph is called k-vertex connected (k-connected
for simplicity) if, for each pair of vertices, there are k mu-
tually vertex disjoint paths (except end-vertices) connecting
them. A k-connected wireless network can sustain the failure
of k 1 nodes.
The connectivity of random graphs, especially the geomet-
ric graphs and its variations, have been considered in the ran-
dom graph theory literature [50], in the stochastic geometry
literature [47,5154], and the wireless ad hoc network litera-
ture [2,5,5561].
Penrose [53] showed that a graph of G(Xn , r) becomes
k-connected almost surely at the moment it has minimum de-
gree k. However, this does not mean to guarantee a graph over
n points is k-connected almost surely, we only have to con-
nect every node to its k nearest neighbors. Let V be a set of
n points randomly and uniformly distributed in a unit square
(or disk). Xue and Kumar [61] proved that, to guarantee that
a geometry graph over V is connected, the number of nearest
neighbors that every node has to connect must be asymptot-
ically (ln n). Dette and Henze [51] studied the maximum
length of the graph by connecting every node to its k nearest
neighbors asymptotically. For the unit volume sphere, their
result implies that, when k > 2,

2
lim Pr nrn,k
ln n + (2k 3) ln ln n 2 ln(k 1)!
n 
2(k 2) ln 2 + ln + 2 = ee .
Li et al. [48] showed that, given n random points V over
a unit-area square, to guarantee that a geometry graph
over V is (k + 1)-connected, the number of nearest neigh-
bors that every node has to connect is asymptotically (ln n+
(2k 1) ln ln n). Li et al. [48] derived a tighter bound on rn for
a set V of n two-dimensional points randomly and uniformly
distributed in C such that the graph G(V , rn ) is k-connected
with high probability.
The theoretical value of the transmission ranges gives us
insight on how to set the transmission radius to achieve the
k-connectivity with certain probability. These results also ap-
ply to mobile networks when the moving of wireless nodes
always generate randomly (or Poisson process) distributed
node positions. Bettstetter [2] conducted the experiments to
study the relations of the k-connectivity and the minimum
node degree using toroidal model. Li et al. [48] also con-
ducted experiments to study the probability that a graph has
minimum degree k and has vertex connectivity k simultane-
ously using Euclidean model. Recently, Bahramgiri et al. [8]
showed how to decide the minimum transmission range of
each node such that the resulted directed communication
graph is k-connected. Here it assumes that the unit disk graph
by setting each node with the maximum transmission range
is k-connected. Lukovszki [62] gave a method to construct a
spanner that can sustain k nodes or k links failures.
LI ET AL.
3. Regional gossip
Although gossiping reduces the routing messages compared
with flooding, it still produces lots of unnecessary messages
in regions that are far away from the line between the source
and the target node. Notice that, the traditional gossip will
propagate the message to the whole network. To further re-
duce the number of forwarding messages, we propose re-
gional gossiping, in which essentially only nodes inside some
region (derived from the source and target) will execute the
gossiping protocol, and nodes outside the region will not par-
ticipate the gossiping at all. The region we select in our sim-
ulations are some ellipses using the source and target as foci.
We now describe our regional gossiping routing method
in detail. Assume that wireless mobile hosts are a set V of
n points distributed in a two-dimensional space. Each node
has a fixed transmission range r: all nodes within distance r
to a node v can receive the signal sent by v. Thus, all mo-
bile hosts define a communication graph G(V , r) in which
there is an edge uv iff uv
r. From now on, we also
assume that the source node knows the position of the target
node, the global ellipse factor , in addition to its own posi-
tion. Every mobile host can get its own position through a
low-cost GPS. In many applications such as data-centric sen-
sor network, there is only a fixed number of destination nodes
(called sink), which is often static, thus every node knows the
positions of these possible target nodes. Otherwise, location
service is needed to find the location of the destination node.
The geometry information of the source node and the destina-
tion node and also the current route (i.e., the route from source
to the sender of the message) is piggybacked along with the
message packet. When a node, say v, receives a message, it
retrieves the geometry position of the source node and the tar-
get node. Node v then checks if it is inside the ellipse defined
by using the source point s and the destination point t as foci.
Notice that, a node v is inside this ellipse iff
vs + vt 
st,
which can be checked trivially. When a node is not inside the
ellipse, it will just simply discard this message. Otherwise,
with a fixed probability p, the node forwards this message to
all nodes within its transmission range. Hereafter, we call p
the relay probability and  the ellipse factor of our regional
gossiping method. Obviously, the probability that the desti-
nation node receives the message depends on the relay prob-
ability p, the ellipse factor , the number of nodes n, and the
transmission range r.
Gupta and Kumar [5] showed that a random graph G(V , r)
is connected whenever r is larger than some threshold
value rn . It is known that the global gossiping (by simply set-
ting  to ) exhibits some bimodal behavior: the destination
node receives the message if and only if the relay probability
is larger than some threshold value. We expect our regional
gossiping method to have the similar transmission phenom-
ena.
We then estimate the relay probability for a network of n
nodes. It was shown in [49] that given n wireless nodes dis-
REGIONAL GOSSIP ROUTING
tributed in a unit square and each node has transmission range
rn and being off or fault with probability p, then the network
is connected with high probability if pnrn2  2 ln n. Con-
sider the network of n nodes distributed in a square region
with side length a. Assume that the distance between the
source and the target is d and the ellipse factor is . The
number of nodes inside the ellipse is then about
n  2 1 2
Nd = 2 d .
a 4
Since each node inside the ellipse forwards the message with
probability p after it receives the message, to let the target
receive the message almost surely, the subnetwork composed
of the nodes inside the ellipse with fault probability p must
be connected. In other words, the relay probability in our
regional gossiping is at least
ln Nd + c(Nd )
p . in a square region. We tried unit disk graphs with different
Nd (r/a)2
Here r is the transmission range of each wireless node and
c(Nd ) is a number going to when Nd goes to . The
probability that the network (each node is chosen with prob-
c(N )
ability p) is connected is ee d . Substituting in Nd , we
have
4a 4 ln(n 2 1d 2 /(4a 2 )) ln(n 2 d 2 /4)
p = .
2 d 2 r 2  2 1 n n 2 2 d 2 r 2 /4
Here 2 =  2 1, d = d/a, and r = r/a. Since for a
random pair of source and target nodes, d
2a, we have
ln(n 2 /4)
p . Ellipse factor. In each iteration of the simulation, the source
n 2 2 r 2 /4
For example, consider a network of n = 1000 nodes distrib-
uted in a square of side length a = 15, and each node has
transmission range r = 1. For ellipse factor  = 1.2, we can
calculate the relay probability p such that the regional gossip-
ing routing can deliver the packets almost surely as
ln(n 2 /4)
p = 0.74.
n 2 2 r 2 /4
The actual relay probability should be larger since we omit
the number c(Nd ) here, which actually decides the success
probability of the regional gossiping. The percentage of all
vertices involved is at most
ln(n 2 d 2 /4)
p Nd /n =  0.46.
r 2 n
Since the distance d between most pairs of source and target is
small compared with a, the actual number of involved vertices
is much smaller. Let Pd be the probability that a pair of source
and target has distance d. The average percentage of number
a
of vertices (for all source and target pairs) is actually x=0 p
Nx Px /n dx. It is not difficult to show that the percentage of
vertices involved in regional gossiping is at most pNd /2n =
65
0.23. When the ellipse factor  = , we can estimate the
relay probability of the regional gossiping as
ln n
p = 0.495.
n r 2
The actual relay probability should be larger, so do the per-
centage of vertices involved in global gossiping. The exper-
iments discussed in the following sections verify the above
study.
4. Experimental studies
4.1. Simulation environment
We conducted extensive simulations to study the performance
of our region gossiping method. We model the network by
unit disk graph and the mobile hosts are randomly placed
number of vertices that are randomly placed in a 15 15
square. Notice that the density of the graph must be above
some threshold to see the effectiveness of the algorithm oth-
erwise the properties would be hidden and cannot be seen. In
other words, the algorithm works better for dense graphs than
sparse graphs with the same parameters p and .
There are different parameters involved in our simulations,
which are described as follows:
Number of vertices. We tried graphs with 1000, 1500 and
2000 vertices. For convenience, we use n to denote the num-
ber of vertices.
vertex and the target vertex are the foci of an ellipse with el-
lipse factor  chosen from 1.2, 1.4, 1.6, 1.8 and 2. We also
consider the case where the ellipse factor  is which is just
the traditional global gossiping method. The smaller the el-
lipse factor is, the narrower the ellipse will be. Notice that
ellipse factor must be greater than one.
Transmission range. Remember that to make the graph
G(V , r) connected, the transmission range has to be greater
than some threshold value rn . To study the effect of the graph
density on the delivery rate, we tried different values of trans-
mission range: 1, 1.5, 2, 2.5 and 3. From the result by Gupta
and Kumar [5], given 1000 nodes in a 15 15 square, the
transmission range should be at least about 0.7 to guarantee a
connected network G(V , r) theoretically.
Relay probability. In our simulation, we use different relay
probabilities p. First, we use the relay probabilities p from
0.1 to 1.0 with step 0.1 and we find that, when the network
is dense enough, the transmission phenomenon happens be-
tween two intervals of relay probabilities. To study this trans-
mission phenomenon in detail, we further refine our relay
probabilities. Specifically, we conduct further simulations us-
ing relay probabilities from 0.02 to 0.30 with step 0.02.
66
Beside the above parameters there are two more constant
metrics used in our simulations as follows:
Source-target pairs. To compute the exact value of the av-
erage delivery rate, we have to try all possible pairs for each
graph, which is n(n1), where n is the number of vertices. It
is not feasible to test all pairs when n is large. Instead we ran-
domly select 100 pairs for each graph and conduct regional
gossiping based routing for each pair. Although we are not
testing all possible pairs, choosing 100 random pairs would
give the results close enough to exact values.
Number of trys. The delivery probability (called delivery
rate also) of our regional gossiping method for a pair of nodes
is defined as the probability that the destination node receives
the message. To compute the delivery rate, we tried sending
the message 1000 times for each pair and then the delivery
rate is approximated by the total number of times that the
message reached the target divided by the total number that
the message is sent (which is 1000 in out simulations).
There are four different types of nodes in each iteration of
our simulations:
(1) Not in ellipse. Nodes that are out of the ellipse region.
(2) Blocked. Nodes that receive the message and do not relay
it.
(3) Relayed. Nodes that receive and relay the message.
(4) Initial hops nodes. The nodes within the initial hops al-
ways receive the message and from those, the ones whose
distance from source is less than some fix initial hops pa-
rameter, always relay the message. Other nodes inside the
ellipse relay the message with the given relay probability.
Here we want to involve as little nodes as possible. In other
words, we want to minimize the number of nodes that relay
(a)
Figure 1. Message delivery rate as a function of relay probability for different values of ellipse factor. Here transmission range is 1. (a) Number of vertices
is 1000. (b) Number of vertices is 2000.
LI ET AL.
the message. It is important because sending message con-
sumes energy and energy is a bottleneck for wireless nodes.
In all the figures of this paper the Y-axis is either the mes-
sage delivery rate or the percentage of vertices that are in-
volved in message delivery, and the X-axis is one of the pa-
rameters with respect to another parameter which is shown
in the legend and the remaining two parameters are fixed.
For example, we can show message delivery rate as a func-
tion of relay probability p for different values of ellipse fac-
tor , while the transmission range r and the number of ver-
tices n are fixed (see figure 1). Each point in each figure rep-
resents the average of the 100, 000 iterations since we will
test 100 different source-target pairs, and each pair is tested
1000 times, when all four parameters are fixed.
We believe that the relay probability and the graph density
are two major factors of message delivery rate. On the other
hand, the ellipse factor and the relay probability are the major
factors determining the number of vertices that are involved
in message delivery. Here a node is said to be involved if it
relays the message. In other words, when the Y-axis is the
message delivery rate and X-axis is either relay probability,
number of vertices or transmission range, we expect to see a
jump in the figures. It means that when the X-axis exceeds
some threshold, then the regional gossiping method almost
surely guarantees that the message arrives at the target. When
the X-axis is less than some threshold, the target almost never
gets the message.
4.2. Message delivery rate as a function of relay probability
We first conduct extensive simulations to study the effect of
the relay probability on the message delivery rate. Intuitively,
if we increase the relay probability, the message delivery rate
increases. Besides the relay probability, we vary either the el-
lipse factor , or the number of vertices n, or the transmission
range r. Now we discuss them one by one as follows.
(b)
REGIONAL GOSSIP ROUTING
1. Message delivery rate as a function of relay probability for
different values of ellipse factor. As can be seen in figure 1,
when the probability exceeds some threshold the delivery rate
jumps from near 0% to near 100%. In each figure, this thresh-
old decreases as the ellipse factor increases because the bigger
the ellipse factor is, the more vertices contribute in message
delivery, and consequently, the probability of the message to
reach the target, which is nothing but the message delivery
rate, increases. For both figures the transmission range is
fixed to 1 unit and the number of vertices is also fixed to 1000
and 2000, respectively.
From figure 1, we observe that when the graph becomes
denser (more vertices in this case), the curve jumps earlier,
and the reason is each time a vertex relays the message,
more nodes get the message (due to more neighbors in dense
graphs) so the probability that the message reaches the target
increases.
One important observation is as follows: as we increase
the ellipse factor, the message delivery rate does not increase
proportionally. Surprisingly, when the ellipse factor is around
1.8, the message delivery rate is almost as good as the one
using global gossiping (i.e., the ellipse factor constraint is
relaxed to ). The reason is where a bigger ellipse factor
is used we are actually considering the vertices that are less
helpful than the vertices which are already considered. Intu-
itively, the vertices, which are far away from the line connect-
ing the source and target, do not help improving the message
delivery rate.
We also observe that, for a fixed relay probability, when the
graph is dense, even a narrow ellipse could guarantee a good
rate of message delivery. Achieving the same delivery rate
using the same relay probability, for a sparser graph, might
not be possible, even if the ellipse factor is relaxed to infin-
ity. In other words, the ellipse factor does not compensate the
description of the graph density. For example, in figure 1(b),
(a)
Figure 2. Number of nodes involved in message delivery as a function of relay probability for different values of ellipse factor. Here transmission range is 1.
(a) Number of vertices is 1000. (b) Number sof vertices is 2000.
67
when the relay probability is 0.3 with ellipse factor of 1.4, the
delivery rate is about 95% for n = 2000, while we cannot
achieve this rate when n = 1000 (see figure 1(a)).
2. Number of nodes involved in message delivery as a func-
tion of relay probability for different values of ellipse factor.
So far, we have concentrated on the transition phenomena of
the delivery rate over the relay probability. Not only the deliv-
ery rate is important for the network performance, but also the
number of vertices involved in the message delivery is impor-
tant for the network life since the wireless devices are often
powered by the batteries only.
The challenge is to find an ellipse factor and a relay prob-
ability such that not only the delivery rate is high (close to
100%) but also the number of vertices involved in the message
delivery is as small as possible. Actually the ellipse factor
and the number of vertices involved in sending the message
from source to target, work against each other. It means that
if we choose a bigger ellipse factor, a higher delivery rate is
achieved, on the other hand, lots of vertices will be involved
in route discovery. In reverse, if we choose a small ellipse
factor then fewer vertices will be involved but it may not have
a good delivery rate.
As can be seen in figure 2, the relation between the num-
ber of vertices involved and the relay probability with respect
to ellipse factors is close to linear. The bigger the relay prob-
ability, the more number of vertices will be involved in the
message delivery. The exact relation between the number of
vertices and relay probability is not simple. Clearly, the far-
ther it is from the source, the less probability it will get the
the message to relay.
In figure 2 when the ellipse factor is infinity, we are ac-
tually flooding the network with a uniform relay probability,
and when this relay probability is 1, the network is completely
(b)
68
flooded, i.e., traditional flooding, so all nodes have the chance
to contribute in message delivery.
Assume that we want to have the delivery rate more than
99%, first consider the case in which we have 1000 nodes,
illustrated in figures 1(b) and 2(b).
We build the table 1 as follows: for each ellipse factor, we
can find the needed relay probability to guarantee the message
delivery above 99% from figure 1, and then by knowing the
values of ellipse factor and the relay probability we can find
the percentage of vertices that are involved from figure 2.
For example, to achieve this rate (above 99%) when ellipse
factor is 1.2, the relay probability must be at least 0.9 (see fig-
ure 1). Then having these two values fixed, we can find the
number of nodes that are involved from figure 2, which would
be about 15% of all vertices. Doing the same thing for differ-
ent values of ellipse factor, we get table 1.
The first column is the different ellipse factors we simu-
lated, and the second column is the corresponding relay prob-
ability in our regional gossip method to guarantee this fixed
delivery rate 99%, and the third column is the percentage of
vertices that are involved in our regional gossiping (i.e., re-
laying the message).
Table 1 shows that we could involve only 15% of vertices
to guarantee the message delivery rate above 99% when the
ellipse factor is 1.2. If we do the same calculations where
there are 2000 nodes then only 10% of vertices will be in-
Table 1
Percentage of the vertices involved in message delivery.
Ellipse factor Relay probability Vertices involved (%)
1.2 0.9
1.4 0.8
1.6 0.7
1.8 0.7
infinity 0.7
(a)
Figure 3. (a) Message delivery rate as a function of relay probability for different values of ellipse factor. Here number of vertices is 1000 and transmission
range is 2. (b) Number of nodes involved in message delivery as a function of relay probability for different values of ellipse factor. Here number of vertices
is 1000 and transmission range is 2.
LI ET AL.
volved (see figures 1 and 2) by choosing ellipse factor 1.2
and relay probability 0.5.
So far the transmission rang was fixed to 1. We were mo-
tivated to study the effect of transmission range as well. We
then tried different values of transmission range. Obviously
the larger the transmission range is, the denser the graph will
be and as mentioned before that causes the jump to occur ear-
lier.
In figure 3 the transmission range is 2. See how similar
figure 1 and figure 3 are, the only difference between these
two figures is the probability at which the jump occurs for
any fixed ellipse factor. Since in delivery rate happens earlier
and quicker when the transmission range increases, we plot
the figures using relay probability range [0, 0.3] for r = 2,
instead of [0, 1] for r = 1.
Again assume that we want to have the delivery rate more
than 99%. Consider the case in which we have 1000 nodes,
but the transmission range is 2 (figures 3(a) and (b)).
We build table 2 as we built table 1: for each ellipse factor.
We can find the relay probability that guarantees the message
delivery rate above 99% from figure 3(a), and then by know-
ing the values of ellipse factor and the relay probability we
can find the percentage of vertices involved in message deliv-
ery from figure 3(b).
For example, to achieve this rate (above 99%) when el-
lipse factor is 1.2, the relay probability must be 0.3 (see fig-
ure 3(a)). Then having these two values fixed, we can find the
number of nodes involved from figure 3(b), which would be
about 8%. Doing the same thing for different values of ellipse
15 factor, we get table 2.
22 Table 2 shows that we could involve only 8% of vertices
25
to guarantee the message delivery rate above 99% for net-
30
70 works of 1000 nodes and with transmission range equal to 2.
If we do the same calculations for networks of 2000 nodes
(b)
REGIONAL GOSSIP ROUTING
with transmission range equal to 2, then only 6% of vertices
will be involved (figures are not shown here).
3. Message delivery rate as a function of relay probability
for different values of transmission range. So far we plotted
the message delivery rate as a function of relay probability for
different values of ellipse factor. Let us replace the ellipse fac-
tor parameter with transmission range and see how the graph
behaves.
As you can see in figure 4, transmission range plays a very
important role in message delivery (see how far the curves are
from each other). As the transmission range is increased, the
delivery rate improves significantly as opposed to the situa-
tion we had earlier with ellipse factor. The reason is when the
transmission range is bigger then each node will be connected
to more nodes, in other words the graph density increases.
Thus, each time a node relays the message, more nodes will
get it and the probability that the message dies out becomes
smaller. Here in figure 4, the ellipse factor is fixed to 1.6.
We built table 3 as follows: for each transmission range,
we can find the relay probability that guarantees the mes-
sage delivery rate above 99% from figure 4(a), and then by
knowing the values of transmission range and the relay prob-
ability we can find the percentage of vertices involved from
figure 4(b).
For example, to achieve this rate (above 99%) when trans-
mission range is 1, the relay probability must be at least 0.8,
Table 2
Percentage of the vertices involved in message delivery.
Ellipse factor Relay probability Vertices involved (%)
1.2 0.3 8 1.0
1.4 0.24 11
1.6 0.22 13
1.8 0.20 14
infinity 0.20 15
(a)
Figure 4. (a) Message delivery rate as a function of relay probability for different values of transmission range. Here number of vertices is 1000 and the
ellipse factor is 1.6. (b) Number of nodes involved in message delivery as a function of relay probability for different values of transmission range. Here
number of vertices is 1000 and ellipse factor is 1.6.
69
(see figure 4(a)). Then having these two values fixed, we
can find the percentage of vertices involved from figure 4(b),
which would be about 30%. We get table 3 by doing the same
calculation for different values of transmission range.
Table 3 illustrates the number of vertices involved in the
regional gossip routing to guarantee a fixed delivery rate 99%
for networks of 1000 nodes with ellipse factor 1.6. Observe
that, all these curves intersect in a common point when the re-
lay probability is 1. Because the ellipse factor is fixed, chang-
ing the transmission range does not change the number of
nodes that are inside ellipse, which is total number of vertices
involved in message delivery when the relay probability is 1.
Actually it is possible to have a node in the ellipse which does
not contribute in message delivery even when the relay prob-
ability is 1, but that is very unlikely. It happens only when
a node in the ellipse doesnt have any neighbor inside the el-
lipse. In our simulations this scenario happened 2 times out
of 180,000,000 iterations.
Another observation is that we get different curves for dif-
ferent transmission ranges. Typically, when the transmission
range is larger, more nodes inside this ellipse will be involved
in the message delivery.
4. Message delivery rate as a function of relay probability
for different number of nodes. In our simulations we stud-
ied networks with different densities in two different ways.
Table 3
Percentage of the vertices involved in message delivery.
Transmission range Probability Vertices involved (%)
0.8 30
1.5 0.5 20
2.0 0.3 14
2.5 0.14 12
3.0 0.11 11.71
(b)
70
(a)
Figure 5. (a) Message delivery rate as a function of relay probability for different number of nodes. Here ellipse factor is 1.6 and transmission range is 1.
(b) Number of nodes involved as a function of relay probability for different number of nodes. Here ellipse factor is 1.6 and transmission range is 1.
First, as described in the previous section, we studied net-
works with fixed number of vertices and different transmis-
sion ranges. Now we study networks with fixed transmission
range and different number of vertices placed in a 15 15
square. In both cases we expect the similar results if the net-
work densities are similar.
As you can see in figure 5, the number of vertices plays
an important role in message delivery (see how far the curves
are from each other). Here we have the same reasoning as the
previous section. As the number of vertices is increased, the
delivery rate improves significantly. The reason is when there
are more vertices in the same area, the graph becomes denser.
Thus, each time a node relays the message more nodes will
get it and the probability that the message dies out becomes
smaller.
Now let us look at the percentage of nodes that are involved
in message delivery as a function of relay probability for dif-
ferent number of nodes (see figure 5). Remember that in this
case ellipse factor and transmission range are fixed. Here we
have the same ellipse with different number of vertices in-
side them. When there are more vertices in the same area
the message is delivered with higher probability since more
nodes will relay the message. Notice that, given a fixed relay
probability, when the node density exceeds some threshold
(depending on the relay probability) almost all nodes inside
the ellipse will receive the message, thus, have the chance to
relay the massage. In other words, if the relay probability is
low, high message delivery rate still can be achieved if the
graph is dense enough and if the graph is sparse, high mes-
sage delivery rate still can be achieved by increasing the relay
probability. On the other hand, larger relay probability will
involve more nodes in message delivery (the number of nodes
involved is almost linear to the relay probability as shown in
right figure of figure 5).
LI ET AL.
(b)
4.3. Message delivery rate as a function of ellipse factor
We can look at the problem from a totally different point of
view. So far we have concentrated on the transition phenom-
ena of the delivery rate over the relay probability. In other
words, in all figures the X-axis was the relay probability. Now
let us see how the network behaves if we use different ellipse
factors while some other parameters are fixed. We found that,
regardless of the network density and relay probability, in-
creasing the ellipse factor does not improve the message de-
livery rate significantly.
1. Message delivery rate as a function of ellipse factor for
different values of transmission range. First let us fix the re-
lay probability and the number of vertices. Remember that to
change the message delivery rate dramatically we can either
increase the relay probability or increase the network density.
As can be seen in figure 6 there is no jump. In other words,
increasing the ellipse factor does not improve the message de-
livery rate dramatically.
Figure 6 shows when the relay probability is fixed, regard-
less of the value of ellipse factor, the graph density must be
above some threshold to guarantee a high message delivery.
As you can see in figure 6(a) when the transmission range
is less than 1.5 then the delivery rate is always below 20%
even if the ellipse factor constraint is relaxed (the case where
ellipse factor constraint is relaxed and not shown in figure 6).
As it is expected if we set the relay probability to a higher
value then the delivery rate would be higher. This is illus-
trated in figure 6: if we increase the value of the relay prob-
ability (from figure 6(a) to figure 6(b)) all curves will be
shifted up.
2. Message delivery rate as a function of ellipse factor for
different number of vertices. As mentioned earlier, the net-
work density can be increased either by increasing the trans-
mission range or by increasing the number of vertices. Now
REGIONAL GOSSIP ROUTING 71

(a) (b)
Figure 6. Message delivery rate as a function of relay probability for different values of transmission range. Here number of vertices is 1000. Relay probability
is (a) 0.1, (b) 0.3.

(a) (b)
Figure 7. (a) Message delivery rate as a function of ellipse factor for different number of vertices. Here transmission range is 1 and relay probability is 0.3.
(b) Number of nodes involved in message delivery as a function of ellipse factor for different number of vertices. Here transmission range is 1 and relay
probability is 0.3.

we replace the transmission range of the previous section with
number of vertices and we expect similar results. In other
words, let us fix the relay probability and the transmission
range to see the delivery rate as a function of ellipse factor for
different number of vertices.
Again, as can be seen in figure 7 there is no jump. In other
words, increasing the ellipse factor does not improve the mes-
sage delivery rate dramatically.
3. Message delivery rate as a function of ellipse factor for
different values of relay probability. In the previous two sec-
tions, we studied the effect of ellipse factor in networks with
different densities, in this section instead of changing the net-
work density, we change the relay probability. Thus, in this
section, the network density is fixed. Specifically, we study
the message delivery rate (as a function of ellipse factor for
different values of relay probability) by fixing the number of
nodes and the transmission range.
In figure 8 when the relay probability is below some
threshold, a high delivery rate cannot be achieved even when
the ellipse factor constraint is relaxed. Figure 8 is similar to
figures 6 and 7 due to the fact that a high relay probabil-
ity can compensate the sparseness of the network and vice
versa.
Intuitively, all the discussions of the two previous sections
apply to this section too. For example, when the network
density is larger than some threshold, the number of vertices
involved is almost linear to the ellipse factor, see figures 7
and 8.
72 LI ET AL.

(a) (b)
Figure 8. (a) Message delivery rate as a function of ellipse factor for different values of relay probability. Here transmission range is 1 and number of vertices
is 1000. (b) Number of nodes involved in message delivery as a function of ellipse factor for different values of relay probability. Here transmission range
is 1 and number of vertices is 1000.

(a) (b)
Figure 9. (a) Message delivery rate as a function of transmission range for different values of relay probability. Here ellipse factor is 1.6 and number of
vertices is 1000. (b) Number of nodes involved in message delivery as a function of transmission range for different values of relay probability. Here ellipse
factor is 1.6 and number of vertices is 1000.

4.4. Message delivery rate as a function of transmission
range
We can look at the problem from a totally different point of
view. So far the X-axis was the relay probability or the ellipse
factor. Thus, for each curve in figures discussed in previous
sections, the network density was fixed. But if we choose the
transmission range or number of vertices as the X-axis then
the graph density changes for each curve. We first study the
case where the X-axis is the transmission range and in the
next section we study the case where the X-axis is the the
number of vertices.
1. Message delivery rate as a function of transmission range
for different values of relay probability. First let us fix the
ellipse factor and the number of vertices. We expect to see
jump because in each curve the graph density changes and
also we expect to see curves that are far from each other due
to the fact that for each curve the relay probability is fixed.
As you can see in figure 9 when the relay probability is
bigger the jump occurs earlier. This figure is similar to fig-
ure 4 due to the fact that the relay probability and transmis-
sion range both improve the message delivery rate signifi-
cantly.
2. Message delivery rate as a function of transmission range
for different values of ellipse factor. Let us fix the number
of vertices and the relay probability to see the delivery rate as
a function of transmission range for different values of ellipse
factor. As you can see in figure 10, like figure 1, as we in-
REGIONAL GOSSIP ROUTING
(a)
Figure 10. (a) Message delivery rate as a function of transmission range for different values of ellipse factor. Here number of vertices is 1000 and relay
probability is 0.2. (b) Number of nodes involved in message delivery as a function of transmission range for different values of ellipse factor. Here number of
vertices is 1000 and relay probability is 0.2.
crease the ellipse factor, the message delivery rate does not in-
crease proportionally. The only difference between figure 10
and figure 1 is: in figure 10 the network density changes in
each curve but in figure 1 the relay probability changes in
each curve. Since increasing either the relay probability or
transmission range improves the message delivery, exchang-
ing those will lead to similar results. Observe that when the
ellipse factor is 1.8, the delivery rate is almost the same as the
global gossiping.
Observe that, in figure 10, the number of vertices involved
in message delivery is almost linear after the transmission
range is larger than some threshold (almost 2). When the
transmission range is small, the number of nodes involved is
small since the message quickly dies out (the relay probability
is 0.2 here).
3. Message delivery rate as a function of transmission range
for different number of vertices. Now let us fix the ellipse
factor and the relay probability to study the message delivery
rate (as a function of transmission range for different number
of vertices). Since the transmission range and the number of
vertices are factors that affect the network density, not only
the network density changes in each curve, but also the net-
work density is different for each curve.
In figure 11, not only the jump occurs (due to the change
of graph density), but also the curves are far from each other
(again due to the change of graph density).
Observe that, the number of vertices involved in the mes-
sage delivery increases almost proportionally to the transmis-
sion range when the relay probability is set to 0.2 (see figure
11(a)). However, when the relay probability increases, say
0.7, the percentage of the number of vertices involved is al-
most constant, see figure 11(b).
73
(b)
4.5. Message delivery rate as a function of number of
vertices
The last parameter is the number of vertices. Since both trans-
mission range and number of vertices affect the network den-
sity, we expect similar results like the previous section.
1. Message delivery rate as a function of number of vertices
for different values of relay probability. Now let us fix the
ellipse factor and the transmission range to see delivery rate
as a function of number of vertices for different values of re-
lay probability. As shown in figure 12, if we use a big enough
relay probability, a high delivery rate is guaranteed. But when
the relay probability is small then we need a large number of
vertices to compensate this small relay probability to guaran-
tee a high delivery rate.
2. Message delivery rate as a function of number of vertices
for different values of ellipse factor. Now let us fix the relay
probability and the transmission range to see delivery rate as
a function of number of vertices for different values of ellipse
factor. Illustrated by figure 13, like figure 10, as we increase
the ellipse factor, the message delivery rate does not increase
proportionally.
3. Message delivery rate as a function of number of vertices
for different values of transmission range. Now let us fix the
ellipse factor and the relay probability to see delivery rate as
a function of number of vertices for different values of trans-
mission range. As you can see in figure 14, the bigger the
number of vertices is, the earlier the jump occurs.
Figures 1214 study the number of vertices that are in-
volved in the message delivery. In these figures, we found
that there are some strange jumps when the number of
74 LI ET AL.

(a) (b)
Figure 11. (a) Message delivery rate as a function of transmission range for different number of vertices. Here ellipse factor is 1.6 and relay probability is 0.2.
(b) Number of nodes involved in message delivery as a function of transmission range for different number of vertices.Here ellipse factor is 1.6 and relay
probability is 0.2.

(a) (b)
Figure 12. (a) Message delivery rate as a function of number of vertices for different values of relay probability. Here ellipse factor is 1.6 and transmission
range is 1. (b) Number of nodes involved in message delivery as a function of number of vertices for different values of relay probability. Here ellipse factor
is 1.6 and transmission range is 1.

vertices is around 1250. We are studying why this hap-
pens.
5. Fault tolerance
To study the fault tolerance of the ad-hoc networks, we simu-
lated the cases in which the target receives the message more
than once. The figure 15 shows the number of times that the
message is delivered to the target at least twice as a function
of relay probability for different values of ellipse factor. If
target has h neighbors inside the ellipse in the best case (i.e.,
all neighbors of the target receive the message) we expect the
message to be delivered p h times. Note that if the target
has only one neighbor inside the ellipse, then the target has no
chance to receive the message more than once. Observe that
figure 15 is a little bit misleading. It shows that with a narrow
ellipse and the replay probability fixed to 1 the probability
that the target receives the message more than once is below
95%. The reason is in our simulations, the sourcetarget pairs
are chosen randomly, so in some cases the target is only one
hop away from the source, thus the target gets the message for
sure but at the same time, due to the closeness of source and
target, there might not be another neighbor inside the ellipse
for target. Thus the target has no chance to receive the mes-
sage more than once. In other words, in some cases, although
the message delivery rate is 100%, the chance that the target
receives the message more than once is 0%.
REGIONAL GOSSIP ROUTING 75

(a) (b)
Figure 13. (a) Message delivery rate as a function of number of vertices for different values of ellipse factor. Here relay probability is 0.4 and transmission
range is 1. (b) Number of nodes involved in message delivery as a function of number of vertices for different values of ellipse factor. Here relay probability
is 0.4 and transmission range is 1.

(a) (b)
Figure 14. (a) Message delivery rate as a function of number of vertices for different values of transmission range. Here ellipse factor is 1.6 and relay
probability is 0.2. (b) Number of nodes involved in message delivery as a function of number of vertices for different values of transmission range. Here
ellipse factor is 1.6 and relay probability is 0.2.

6. Conclusion and future work
We proposed a regional gossip approach, where only the
nodes within some region forward the routing message with
some probability, to reduce the overhead of the routing pro-
tocol imposed on the network. We showed how to set the for-
warding probability based on the region and the estimated
network density both by theoretical analysis and by exten-
sive simulations. Our simulations showed that the number of
messages generated using this approach is less than the sim-
ple global flooding (up to 94%), which already saves many
messages compared with global flooding.
Hass et al. [24] expected that the global gossiping com-
bined with the cluster-based routing can further improve the
performance. We doubt this due to two reasons: (1) the back-
bone formed by clusterheads are already very sparse, and to
guarantee that all nodes receive the messages, the gossiping
probability is very high; and (2) the communication cost to
maintain the backbone will also offset the benefit gained by
global gossiping, if there is any. We will conduct simulations
to study this.
One of the main questions remaining to be studied is to use
non-uniform ellipse factors. In our simulations, the ellipse
factor is uniform regardless of the distance between source
and target. We believe that using a bigger ellipse factor, when
the source and target are close, will get better results.
Another question is studying networks with different den-
sities, meaning that instead of trying different transmission
76
Figure 15. The number of times that the message receives the target more
than once as a function of relay probability for different values of ellipse
factor. Here transmission range is 1 and number of vertices is 1000.
ranges and different number of nodes, networks with differ-
ent densities can be studied. To generate a network with
a given density with respect to transmission range, we can
keep adding nodes to the network until the desired density is
reached.
We had assumed that two nodes can always communicate
if their distance is no more than the transmission range. How-
ever, this is not totally true practically. Some pair of nodes
cannot communicate at all even if they are close. We can
model this by assigning another link probability pl : a link ex-
ist with probability pl . Here probability pl could be uniform
or dependent on the distance between the pair of nodes.
References
[1] S. Capkun, M. Hamdi and J.P. Hubaux, GPS-free positioning in mobile
ad-hoc networks, in: Proc. Hawaii Internat. Conf. on System Sciences
(2001).
[2] C. Bettstetter, On the minimum node degree and connectivity of a wire-
less multihop network, in: 3rd ACM Internat. Symposium on Mobile Ad
Hoc Networking and Computing (MobiHoc02) (June 2002).
[3] G. Calinescu, I. Mandoiu and A. Zelikovsky, Symmetric connectivity
with minimum power consumption in radio networks, in: IFIPTCS
(2002) to appear.
[4] A. Clementi, P. Penna and R. Silvestri, The power range assignment
problem in radio networks on the plane, in: XVII Symposium on The-
oretical Aspects of Computer Science (STACS00), Lecture Notes in
Computer Science, Vol. 1770 (Springer, New York, 2000) pp. 651660.
[5] P. Gupta and P. R. Kumar, Critical power for asymptotic connectivity in
wireless networks, in: Stochastic Analysis, Control, Optimization and
Applications: A Volume in Honor of W.H. Fleming, eds. W.M. McE-
neaney, G. Yin and Q. Zhang (1998).
[6] J. Monks, V. Bharghavan and W.-M Hwu, A power controlled multi-
ple access protocol for wireless packet networks, in: IEEE INFOCOM
(2001).
[7] M. Sanchez, P. Manzoni and Z. Haas, Determination of critical trans-
mission range in ad-hoc networks, in: Multiaccess, Mobility and Tele-
traffic for Wireless Communications (MMT99) (1999).
[8] M. Bahramgiri, M.T. Hajiaghayi and V.S. Mirrokni, Fault-tolerant
and 3-dimensional distributed topology control algorithms in wireless
LI ET AL.
multi-hop networks, in: Proc. of the 11th Annual IEEE Internat. Conf.
on Computer Communications and Networks (ICCCN) (2002) pp. 392
397.
[9] L. Hu, Topology control for multihop packet radio networks, IEEE
Transactions on Communications 41(10) (1993).
[10] X.-Y. Li, P.-J. Wan, Y. Wang and O. Frieder, Sparse power efficient
topology for wireless networks, Journal of Parallel and Distributed
Computing (2002) to appear; preliminary version appeared in: ICCCN
(2001).
[11] X.-Y. Li, G. Calinescu, P.-J. Wan and Y. Wang, Localized delaunay tri-
angulation with application in wireless ad hoc networks, IEEE Trans-
actions on Parallel and Distributed Systems (2003); short version ap-
peared in: IEEE INFOCOM (2002).
[12] X.-Y. Li, G. Calinescu and P.-J. Wan, Distributed construction of pla-
nar spanner and routing for ad hoc wireless networks, in: 21st Annual
Joint Conf. of the IEEE Computer and Communications Societies (IN-
FOCOM), Vol. 3 (2002).
[13] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
wireless networks using transmit power adjustment, in: IEEE INFO-
COM (2000).
[14] R. Wattenhofer, L. Li, P. Bahl and Y.-M. Wang, Distributed topology
control for wireless multihop ad-hoc networks, in: IEEE INFOCOM01
(2001).
[15] S. Singh, M. Woo and C. Raghavendra, Power-aware routing in mobile
ad hoc networks, in: Proc. of MOBICOM (1998) pp. 181190.
[16] I. Stojmenovic and X. Lin, Power aware localized routing in ad hoc
wireless networks, IEEE Transactions on Parallel and Distributed Sys-
tems 12(11) (2001) 11221133.
[17] J. Wu, F. Dai, M. Gao and I. Stojmenovic, On calculating power-
aware connected dominating sets for efficient routing in ad hoc wireless
networks, IEEE/KICS Journal of Communication and Networks 4(1)
(2002) 5970.
[18] P. Bose, P. Morin, I. Stojmenovic and J. Urrutia, Routing with guar-
anteed delivery in ad hoc wireless networks, ACM/Kluwer Wireless
Networks 7(6) (2001) 609616, also in: 3rd Internat. Workshop on
Discrete Algorithms and Methods for Mobile Computing and Commu-
nications (1999) pp. 4855.
[19] B. Das and V. Bharghavan, Routing in ad-hoc networks using minimum
connected dominating sets, in: 1997 IEEE Internat. Conf. on Commu-
nications (ICC97), Vol. 1 (1997) pp. 376380.
[20] C.E. Perkins and P. Bhagwat, Highly dynamic destination-sequenced
distance-vector routing (DSDV) for mobile computers, Computer
Communications Review (October 1994) 234244.
[21] D.B. Johnson and D.A. Maltz, Dynamic source routing in ad hoc
wireless networks, in: Mobile Computing, eds. Imielinski and Korth,
Vol. 353 (Kluwer Academic, Dordrecht, 1996).
[22] B. Karp and H.T. Kung, Gpsr: Greedy perimeter stateless routing for
wireless networks, in: ACM/IEEE Internat. Conf. on Mobile Comput-
ing and Networking (2000).
[23] Y.-B. Ko and N.H. Vaidya, Using location information to improve rout-
ing in ad hoc networks, Technical Report, Department of Computer
Science, Texas A&M University (1997).
[24] Z. Haas, J. Halpern and L. Li, Gossip-based ad hoc routing, in: IEEE
INFOCOM (2002).
[25] S. Murthy and J. Garcia-Luna-Aceves, An efficient routing protocol for
wireless networks, ACM Mobile Networks and Applications (Special
Issue on Routing in Mobile Communication Networks) 1(2) (1996).
[26] V. Park and M. Corson, A highly adaptive distributed routing algorithm
for mobile wireless networks, in: IEEE INFOCOM (1997).
[27] C. Perkins, Ad-hoc on-demand distance vector routing, in: MILCOM
97 (November 1997).
[28] C. Perkins and P. Bhagwat, Highly dynamic destination-sequenced
distance-vector routing, in: Proc. of the ACM SIGCOMM (October
1994).
[29] P. Sinha, R. Sivakumar and V. Bharghavan, CEDAR: Core extraction
distributed ad hoc routing algorithm, IEEE Journal on Selected Areas
in Communications 17(8) (1999) 14541465.
REGIONAL GOSSIP ROUTING
[30] P. Sinha, R. Sivakumar and V. Bharghavan, Enhancing ad hoc rout-
ing with dynamic virtual infrastructures, in: Proc. of IEEE INFOCOM
2001, Vol. 3 (2001) pp. 17631772.
[31] J. Wu and H. Li, A dominating-set-based routing scheme in ad
hoc wireless networks, Telecommunication Systems (Special Issue on
Wireless Networks) 3 (2001) 6384.
[32] M. Joa-Ng and I.-T. Lu, A peer-to-peer zone-based two-level link state
routing for mobile ad hoc networks, IEEE Journal on Selected Areas in
Communications 17(8) (1999) 14151425.
[33] E. Royer and C. Toh, A review of current routing protocols for ad-
hoc mobile wireless networks, IEEE Personal Communications (April
1999).
[34] S. Ramanathan and M. Steenstrup, A survey of routing techniques for
mobile communication networks, Mobile Networks and Applications
(1996) 89104.
[35] M. Mauve, J. Widmer and H. Harenstein, A survey on position-based
routing in mobile ad hoc networks, IEEE Network Magazine 15(6)
(2001) 3039.
[36] P. Hall, On continuum percolation, The Annals of Probability 13(4)
(1985).
[37] R. Meester and R. Roy, Continuum Percolation (Cambridge Univ.
Press, Cambridge, 1996).
[38] S. Basagni, I. Chlamtac, V.R. Syrotiuk and B.A. Woodward, A dis-
tance routing effect algorithm for mobility (DREAM), in: Proc. of
ACM/IEEE MOBICOM98 (1998).
[39] Z. Haas and B. Liang, Ad-hoc mobility management with uniform
quorum systems, IEEE/ACM Transactions on Networking 7(2) (1999)
228240.
[40] I. Stojmenovic, A routing strategy and quorum based location update
scheme for ad hoc wireless networks, Technical Report TR-99-09,
Computer Science, SITE, University of Ottawa (1999).
[41] I. Chlamtac and A. Farago, A new approach to design and analysis of
peer to peer mobile networks, Wireless Networks 5 (1999) 149156.
[42] P.-J. Wan, K.M. Alzoubi and O. Frieder, Distributed construction of
connected dominating set in wireless ad hoc networks, in: INFOCOM
(2002).
[43] K.M. Alzoubi, P.-J. Wan and O. Frieder, New distributed algorithm
for connected dominating set in wireless ad hoc networks, in: HICSS,
Hawaii (2002).
[44] K. Alzoubi, P.-J. Wan and O. Frieder, Message-optimal connected-
dominating-set construction for routing in mobile ad hoc networks,
in: 3rd ACM Internat. Symposium on Mobile Ad Hoc Networking and
Computing (MobiHoc02) (2002).
[45] K.N. Amouris, S. Papavassiliou and M. Li, A position based multi-zone
routing protocol for wide area mobile ad-hoc networks, in: Proc. of the
49th IEEE Vehicular Technology Conf. (1999) pp. 13651369.
[46] G. Calinescu, I. Mandoiu, P.-J. Wan and A. Zelikovsky, Selecting for-
warding neighbors in wireless ad hoc networks, in: ACM DialM (2001).
[47] M. Penrose, The longest edge of the random minimal spanning tree,
Annals of Applied Probability 7 (1997) 340361.
[48] X.-Y. Li, Y. Wang, C.-W. Yi, P.-J. Wan and O. Frieder, Robust wireless
ad hoc networks, in: IEEE ICC (2003) accepted for publication.
[49] P.-J. Wan, C.-W. Yi, X.-Y. Li, Y. Wang and O. Frieder, Asymptotic
distribution of critical transmission range for k-connectivity in wireless
ad hoc networks (2002) submitted for publication.
[50] B. Bollobs, Random Graphs (Cambridge Univ. Press, Cambrige,
2001).
[51] H. Dette and N. Henze, Some peculiar boundary phenomena for ex-
tremes of rth nearest neighbor links, Statistics & Probability Letters 10
(1990) 381390.
77
[52] M. Penrose, Extremes for the minimal spanning tree on normally dis-
tributed points, Advances in Applied Probability 30 (1998) 628639.
[53] M. Penrose, On k-connectivity for a geometric random graph, Random
Structures and Algorithms 15 (1999) 145164.
[54] M. Penrose, A strong law for the longest edge of the minimal spanning
tree, Annals of Probability 27 (1999) 246260.
[55] D.M. Blough, M. Leoncini, G. Resta and P. Santi, On the symmetric
range assignment problem in wireless ad hoc networks, in: Proc. of the
2nd IFIP Internat. Conf. on Theoretical Computer Science (2002).
[56] C. Cooper and A. Frieze, On the connectivity of random kth nearest
neighbour graphs, Combinatorics, Probability and Computing 4 (1995)
343362.
[57] M. Grossglauser and D. Tse, Mobility increases the capacity of ad-hoc
wireless networks, in: INFOCOMM, Vol. 3 (2001) pp. 13601369.
[58] P. Gupta and P. Kumar, Capacity of wireless networks, Technical Re-
port University of Illinois, Urbana-Champaign (1999).
[59] O.D. Patrick, Connectivity in ad-hoc and hybrid networks, in: IEEE
INFOCOM (2002).
[60] P. Santi and D.M. Blough, An evaluation of connectivity in mobile
wireless ad hoc networks, in: Proc. of IEEE DSN (2002) pp. 8998.
[61] F. Xue and P.R. Kumar, The number of neighbors needed for connec-
tivity of wireless networks, Wireless Networks (2002) submitted.
[62] T. Lukovszki, New results of fault tolerant geometric spanners, in:
Workshop on Algorithms and Data Structures (1999) pp. 193204.
Xiang-Yang Li has been an Assistant Professor of
Computer Science at the Illinois Institute of Tech-
nology since 2000. He hold M.S. (2000) and Ph.D.
(2001) degree in computer science from University
of Illinois at Urbana-Champaign. He received his
Bachelor degree in computer science and Bachelor
degree in business management from Tsinghua Uni-
versity, P.R. China in 1995. His research interests
span the computational geometry, wireless ad hoc
networks, optical networks, and cryptography. He
is a member of the ACM and IEEE.
E-mail: xli@cs.iit.edu
Kousha Moaveninejad received the Bachelor in
computer soft engineering from Sharif University of
Technology, Tehran, Iran, in 1997. He joined the
Department of Computer Science of Illinois Insti-
tute of Technology in 2000, as a M.S. student and
received the Masters degree in 2002. He then contin-
ued his study as a Ph.D. student at Illinois Institute of
Technology and his current research interests include
wireless ad-hoc networks, computational geometry,
algorithm design, and mobile computing.
E-mail: moavkoo@iit.edu
Ophir Frieder is the IITRI Professor of Computer
Science at the Illinois Institute of Technology. His
research interests span the general area of distributed
information systems. He is a member of ACM and a
fellow of the IEEE.
E-mail: ophir@cs.iit.edu
 Mobile Networks and Applications 10, 7988, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Comparison and Evaluation of Multiple Objective Genetic

Algorithms for the Antenna Placement Problem
LARRY RAISANEN and ROGER M. WHITAKER
Centre for Mobile Communications, Department of Computer Science, Cardiff University, Queens Buildings, The Parade,
P.O. Box 916, Cardiff CF24 3XF, UK

Abstract. The antenna placement problem, or cell planning problem, involves locating and configuring infrastructure for cellular wireless
networks. From candidate site locations, a set needs to be selected against objectives relating to issues such as financial cost and service
provision. This is an NP-hard optimization problem and consequently heuristic approaches are necessary for large problem instances. In this
study, we use a greedy algorithm to select and configure base station locations. The performance of this greedy approach is dependent on
the order in which the candidate sites are considered. We compare the ability of four state-of-the-art multiple objective genetic algorithms
to find an optimal ordering of potential base stations. Results and discussion on the performance of the algorithms are provided.
Keywords: genetic algorithms, antenna placement

1. Introduction multiple and conflicting objectives. In this regard, the APP

The proliferation of cellular wireless services for mobile com-
munication has led to the antenna placement problem (APP).
For cellular wireless systems, mobile communication is fa-
cilitated by base stations which have an appropriate spatial
distribution. The area of service coverage from a single an-
tenna at a base station constitutes a cell, which is a region
where the radiated signal power from the serving antenna is
of sufficient strength to be received by subscribers. As the
power of transmitted signals must be restricted, multiple cells
are required to provide wide area coverage. The collection of
all cells across the network constitutes a cell plan.
The APP involves selecting base station site locations from
a set of candidates, which are normally located irregularly.
Selected sites must be configured to provide adequate service
coverage and capacity while adhering to constraints involv-
ing regions which can be served by more than one antenna.
Such constraints are imposed to ensure that the potential for
interference is controlled while providing regions in the net-
work for call handover, which is necessary for seamless call
transfer between cells. Areas covered by more than one cell
must be carefully controlled both to maintain network opera-
tion and minimize the total commitment to infrastructure. The
primary configuration variable at a site is transmission power.
This directly affects the cell size, the required number of cells
(and therefore financial cost), and handover regions.
The number of factors involved in solving the APP means
that automatic software for designing cell plans has become
increasingly common [50]. We classify software as auto-
matic if the associated computer program has autonomy in
the selection of base station locations and the configuration
of antennae. The underlying algorithm in the software is re-
quired to tackle an NP-hard [33] optimization problem with
Supported by a Doctoral Scholarship from the EPSRC.
Corresponding author.
is an extension of the well-studied facilities location prob-
lem [11], namely, the capacitated facilities location prob-
lem with unsplittable demands. Consequently, heuristic and
meta-heuristic techniques have become increasingly popular
for solving the APP.
In this paper, we focus on resolving the two fundamental
aspects of cell planning: providing the required service cov-
erage at the lowest possible financial cost. These two con-
flicting objectives always exist when setting up cellular net-
work services, as adding base stations to improve coverage
inherently increases the cost of the network. In this study, we
produce cell plans in which base station locations are selected
and allocated a transmission power, assuming an isotropic ra-
diation pattern (i.e., power radiates in all directions with equal
strength). As base stations are expensive to commission and
manage, we optimize the total cost and location of base sta-
tions commissioned. Despite the importance of finding an
optimal trade off between these objectives, we are not aware
of any studies in the literature addressing this issue.
The cell plans we produce are the first step in establishing
a cost effective operational network, and only factors which
have the largest impact on financial cost and service cover-
age are considered. Although not the focus of our work, we
note that once the tension between cost and coverage is re-
solved, the resultant cell plan is then ready to undergo de-
tailed dimensioning of individual cells [15]. This second stage
may involve adjusting additional variables at the selected base
stations such as tilt and direction (i.e., azimuth). Addition-
ally, multiple directed co-sited antenna may be invoked at
this stage to increase the capacity for multiplexing, given
knowledge of anticipated traffic patterns. Known as sector-
ization, this is common operational practice in mobile tele-
phony, whereby using multiple co-sited antenna is generally
far cheaper than commissioning a new site.
80
To optimally resolve the competition between service cov-
erage and financial cost, we introduce a multiple objective
optimization framework that does not require a priori knowl-
edge of the relative importance of service coverage versus
cost. This is achieved by providing a range of alternative
site selections which approximate the best possible trade-offs
(i.e., Pareto front) between cost and coverage. This means
that unlike the current convention for cell planning, which
generally seeks to generate a single cell plan given informa-
tion on the relative importance of objectives, a radio engineer
will be able to choose from a range of alternative cell plans,
given visual and detailed information regarding each. This is
particularly beneficial when there is a nonlinear relationship
or unknown dependency between the objective functions as
for the general APP.
As far as we are aware [50], this method has only been con-
sidered for cell planning in [36], where a genetic algorithm
was developed specifically for the APP; however, the total
network cost was not considered. Unlike [36], the framework
we propose considers financial cost and is flexible because
it is possible to plug-in any multiple objective optimiza-
tion algorithm (MOA) which seeks to approximate a Pareto
front. This flexibility is achieved by making the cell plan rep-
resentation independent from the task of the MOA, which, in
this case, is to find optimal orderings of candidate site loca-
tions which optimize the two objective functions. Exploring
a search space in this manner is common practice in many
discrete optimization problems (e.g., the knapsack problem)
and has also been successfully applied to the frequency as-
signment problem [27].
Using this approach, we explore the relative trade-offs be-
tween financial cost and service coverage using four state-of-
the-art genetic MOAs, which were selected as they produce
many alternative solutions in parallel. The performance of
SEAMO, SPEA2, NSGA-II, and PESA is considered using
a range of synthesized test problems. We argue that the ap-
proach we take is more likely to lead to an efficient opera-
tional cellular network, as the two fundamental aspects are
considered in isolation from other cell planning tasks.
2. Solving the APP
The first published paper on optimizing antenna placement
dates back to 1994 [6]. Since then a large number of ap-
proaches and scenarios have appeared in the literature. Al-
though exact approaches are only feasible for relatively small
test problems, they have been applied in a range of papers
[3739,49]. However, in some papers, they have been re-
laxed or selectively applied. Sequential, or greedy, algo-
rithms have been less well used and then predominantly for
comparison purposes (see [5,37]). Deterministic heuristic
algorithms have also been proposed by a range of authors
[9,17,18,28,34,37,38,44,45,53]. Frequently, these approaches
exploit observations (e.g., density of base station locations)
about the APP and incorporate them to enhance the perfor-
mance of the cell plans obtained. However, meta-heuristic
RAISANEN AND WHITAKER
algorithms based on simulated annealing [1], tabu search
[19,20] and genetic algorithms [23] are far more popular.
Simulated annealing has been adopted for the APP in
[2,3,6,26,33], and tabu search for the APP in [4,21,32,47].
Both these techniques operate by ranking solutions using a
cost function. Given a solution, small changes are made to
create a neighborhood of solutions from the current solution.
The meta-heuristic then guides the acceptance of new solu-
tions available in the neighborhood. The advantage of this
approach is that it has the ability to escape from local minima
in the search space (regarding the cost function), thereby im-
proving performance. Differences in the application of these
approaches involve how the cell planning problem is mod-
elled, the formulation of rules to create neighborhoods, and
the cost function used to rank solutions.
Genetic algorithms have also become increasingly popu-
lar for the APP [8,21,25,31,32,36,37,40]. With the excep-
tion of [36], these approaches predominantly seek to opti-
mize a single function (or a linear combination of multiple
objective functions) to create a population of high quality so-
lutions. These algorithms mimic evolution and natural se-
lection through fitness assignment, selection, recombination,
and mutation on a population of solutions. For a genetic al-
gorithm to succeed, a suitable representation of the problem
needs to be used. The most popular representation for the
APP in the literature is a binary string with crossover opera-
tions defined using geographic information between individ-
ual base stations. However, as we show in this paper, genetic
algorithms need not be restricted to this representation. Our
study is unique in that it combines an integer string represen-
tation with a multiple objective approach.
2.1. Resolving conflicting multiple objectives
Regardless which optimization technique is adopted to solve
the APP, it is necessary to resolve the conflict between com-
peting multiple objectives, such as service coverage and fi-
nancial cost. The following definition is useful in this context.
Definition 1 (Pareto optimality). Let o1 , o2 , . . . , on be ob-
jective functions which are to be maximized. Let S be the set
of all possible solutions. s S is dominated by t S (de-
noted t  s) if j , j {1, . . . , n}, such that oj (t) > oj (s)
and i, 1
i
n, oi (t)  oi (s). A non-dominated solution
is said to be Pareto optimal.
Pareto optimal cell plans are non-dominated in the sense
that it is not possible to improve the value of any objective
without simultaneously degrading the quality of one or more
of the other objectives. The set of all possible Pareto optimal
solutions in the entire search space is called the Pareto front.
In figure 1, a hypothetical Pareto front is indicated for the ob-
jectives of cost and coverage. The most desirable cell plan in
the Pareto front depends on which objective is most impor-
tant. However, in the absence of such a relative ranking of
objectives, solutions from the Pareto front must be regarded
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS
Figure 1. Progress towards Pareto front of cost and coverage.
as equivalent. Our approach is to generate a set of alternative
solutions (i.e., cell plans) which approximate the Pareto front.
Despite the potential strength of using Pareto optimality
within the context of cell planning in this way, it has not been
addressed adequately in the literature. This may be partially
due to the fact that there are a number of alternative strategies
available, which can also handle multiple objective problems.
These strategies are:
1. Combine all objectives into a single scalar value, typically
as a weighted sum, and optimize the scalar value.
2. Solve for the objectives in a hierarchical fashion, optimiz-
ing for a first objective then, if there is more than one solu-
tion, optimize these solution(s) for a second objective, and
repeat.
3. Obtain a set of alternative, non-dominated solutions, each
of which must be considered equivalent in the absence of
further information regarding the relative importance of
each of these objectives.
Each approach involves exploring the search space of all
possible cell plans to find one or more suitable solutions.
Approach one is by far the most popular approach in the
literature (e.g., [46,8,17,21,24,34,35,42,45,47,51,52]). The
biggest problem with this approach is that setting the rela-
tive weights of different components in the cost function may
lead to inappropriate favoring or penalizing of different ob-
jectives. Approach two may be combined with approach one,
as in [26,32,40,53], which may involve changing the objec-
tive function at different points in the search in a phased or
staged manner. This approach effectively prioritizes differ-
ent single optimization objectives a priori and therefore has
similar problems to the first approach. Only in [36] has a
multi-objective search been implemented using approach 3,
where the Pareto front, in this case, was approximated using
a problem specific genetic algorithm which did not consider
the financial cost of the cell plan.
2.2. Cell plan model
Before discussing the genetic algorithms used to generate a
set of non-dominated solutions, we first turn our attention to
81
the cell plan model and representation used. Firstly, we define
a working area as the region over which transmission is con-
sidered. This is characterized by discretized test points. The
following sets form the input to our formulation of the APP:
A set of candidate sites for locating base stations, denoted
L = {L1 , . . . , LnBS }.
A list of possible transmission powers p0 , p1 , p2 , . . . , pk
in ascending order of magnitude. Zero power is denoted
by p0 .
A set of service test points (STP), {s1 , . . . , snstp }, where a
signal must be received above a minimum specified ser-
vice threshold Sq to ensure a required quality of service.
A maximum handover percentage used to consider the vi-
ability of the handover region when commissioning a new
cell.
For purposes of candidate sites, we assume that each base
station is operating a single omni-directional antenna with an
isotropic radiation pattern. The antenna height is assumed to
be fixed at the maximum permitted at the site to enhance po-
tential transmission range. Finally, each base station location
Li has a cost $(Li ) associated with commissioning it. The
cost of each base station was set to a fixed uniform random
value between 1 and 100 for each test problem. We assume
that service test points are regularly spaced every 300 me-
ters and the maximum handover parameter has been set at
30% throughout. The service threshold has been taken as
90 dBm, which is a realistic value for GSM services and
equipment.
2.2.1. Propagation
A service test point r is said to be covered by antenna A if
the received signal strength from A, denoted PrA , is greater
than Sq . We assume
PrA = P A PL L + G
where P A is the power at which A is transmitting, PL is the
path loss experienced between A and r, L is the aggregation
of other losses experienced, and G is the aggregation of gains
experienced. For experimental purposes, we assume G = L.
For each combination of A and r, PL may be recorded in the
field or estimated using a free space path loss or empirical
model. In the absence of data from the field, we adopted the
empirical model proposed by Hata [22]:
PL = 69.55 + 26.16 log(f ) 13.82 log(hb ) a(hm )


K + 44.9 6.55 log(hb ) log(R),
given particular values for variables such as frequency (f ),
base station height (hb ), and mobile receiver height (hm ). For
this investigation, these values are set as f = 800 MHz,
hb = 31 meters, and hm = 1.5 meters. Additional envi-
ronmental correction factors include a(hm ), K and the prop-
agation distance R. As the mobile receiver height was set to
1.5 meters, a(hm ) is 0. As the standard urban version of the
model was used, K is 0. R is the distance in kilometers from
82
each base station to each STP. As well as Hatas model, many
other propagation models would have been equally suitable.
2.2.2. Handover regions and objectives
The subset of service test points covered by a particular an-
tenna A is the cell served by A, denoted cA . Note that cells
served by different antennae are not necessarily disjoint since
an STP can potentially be covered by more than one antenna.
Such an STP is referred to as a handover STP. A handover
STP which is contained in more than two distinct cells is re-
ferred to as a soft handover STP. For a cell cA , the subset of
handover STPs is denoted hA . For cA , the handover percent-
age is defined as
|hA |
100.
|cA |
Controlling the size and distribution of handover regions is
crucial for both operational and financial reasons. Handover
regions are a prerequisite for seamless call transfer between
cells for mobile users. However, if very large handover re-
gions are permitted, there is a greater potential for inter-
ference due to strong signals being received from multiple
sources. In frequency division multiple access systems, large
handover regions increase the need for large channel separa-
tion between adjacent cells in the frequency assignment prob-
lem. Large handover regions may also adversely affect the
cost of the network by increasing the total number of base
stations required to cover a given area.
The objectives we are concerned with relate to financial
cost and area coverage of service test points. The cost of a
cell plan L , denoted costL , is defined as:
 plans) through a process of selection and recombination. This
costL = $(Li ).
Li L
The coverage of a cell plan is expressed as the proportion of
STPs which are covered. Handover is not considered as an
objective in our problem formulation, but imposed as a con-
straint controlled by via the decoder.
2.3. Cell plan representation
The potential base station location Li is referred to as the ith
base station. We use a permutation of the potential base
station locations to represent cell plans. Each permutation
orders the potential base station locations. Under the permu-
tation , the ith base station location listed is denoted (i).
We introduce a decoder which translates a permutation into
a cell plan. This approach mimics the way in which the prob-
lem might be attempted manually. The decoder is effectively
a greedy, sequential algorithm for creating a cell plan, which
is dependent on the order of inspection for commissioning po-
tential sites occurring in . The decoder adds cells iteratively
to create a cell plan L as follows:
Initially L = .
Potential sites (1), (2), . . . , (n) are inspected (in the each solution which is the sum of two parts. The first part is a
order induced by ) for possible selection.
RAISANEN AND WHITAKER
At iteration j (1
j
n), (j ) is considered for addi-
tion to the set L .
Handover between c(j ) and L is feasible if the han-
dover percentage for c(j ) is less than the maximum
permitted.
The largest power setting, denoted pmax , is identified
from the list p0 , p1 , p2 , . . . , pk such that handover is
feasible between c(j ) and L .
If pmax = p0 , then (j ) is added to L , and the trans-
mission power of (j ) is recorded as pmax . Otherwise
(j ) is not added to L .
A number of observations can be made regarding this ap-
proach. Firstly, the approach is greedy in the sense that once
a base station location is added to L at power pmax , the base
station cannot be removed from the cell plan L nor can its
transmission power be adjusted. Secondly, for a particular list
of potential site locations, characteristics (e.g., cost and cov-
erage) of the resultant cell plan L is entirely dependent on
the order (i.e., permutation ) in which the base stations are
considered for selection. It is our aim to find the best permu-
tations, which lead to Pareto optimal cell plans, using genetic
algorithms.
3. Genetic algorithms
Only over the last decade have genetic algorithms (GAs) been
successfully adapted to solve multiple objective problems. An
excellent overview of this area is given in [12]. The general
principle is to breed a new population of solutions (i.e., cell
occurs over a number of generations to try to improve the per-
formance of the population, as shown in figure 1. The expec-
tation is that desirable characteristics in solutions from one
generation will combine to produce better solutions for the
next generation. Introduced by Holland [23], GAs are sup-
ported by theory which identifies the conditions under which
solutions converge to a high performing set of solutions. GAs
which approximate the Pareto front seek to find, ideally, a di-
verse set of solutions spread evenly over the entire range of
the Pareto optimal front. In this study, we consider the abil-
ity of four state-of-the-art GAs, namely, SPEA2, NSGA-II,
PESA, and SEAMO to perform this function. Each algorithm
is briefly described below.
3.1. Brief description of each GA
The Strength Pareto Evolutionary Algorithm version II
(SPEA2) is an enhancement of that originally proposed in
[54], and is described in detail in [55]. SPEA2 has been used
in numerous studies (e.g., [7]) where good performance, in
comparison to other MOAs, has been reported. In SPEA2,
the most fit individuals from the union of archive and child
populations are determined by computing a fitness value for
raw fitness value based on how many solutions it dominates,
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS
and the second is a density estimate based on the its proximity
to other solutions in objective space. At each generation, the
most fit n solutions are saved to the archive, and genetic oper-
ators are applied to form a new child population. This process
is repeated until termination.
NSGA-II is a fast elitist non-dominated sorting genetic al-
gorithm (see [13] for a full description), which has been well
studied (e.g., [14,29]). NSGA-II is similar to SPEA2, but uses
slightly different mechanisms. For example, in NSGA-II the
most fit individuals from the union of archive and child popu-
lations are determined by a ranking mechanism (or crowded
comparison operator) composed of two parts. The first part
peels away layers of non-dominated fronts, and ranks solu-
tions in earlier fronts as better. The second part computes a
dispersion measure, the crowding distance, to determine how
close a solutions nearest neighbors are, with larger distances
being better. At each generation, the best n solutions with
regard to these two measures are saved to the archive, and ge-
netic operators applied to form a new child population. This
process is repeated until termination.
The Pareto Envelope-based Selection Algorithm, PESA, is
described in [10]. It uses different mechanisms than SPEA2
and NSGA-II. The main differences are that its archive pop-
ulation is not of fixed size and only allows non-dominated
solutions to be members, which is a more limited set than
the previous two GAs allowed. If the archive ever exceeds
n solutions, a squeeze factor is calculated for all members of
the archive. The squeeze factor is the total number of mem-
bers in the same subregion of a hyper-grid (which partitions
the search space into subregions (see [30])). The higher the
squeeze factor, the more local neighbors a solution has. Ran-
dom members from the grid region with the highest squeeze
factor are then removed until the size of the archive is reduced
to n. Genetic operators are then applied to archive members
to form a new child population. This process is repeated until
termination.
Finally, the Simple Evolutionary Algorithm for Multi-
objective Optimization, known as SEAMO, has performed
particularly well on the benchmark test knapsack optimiza-
tion problem [46]. The main difference between SEAMO and
the other algorithms, is that it is steady-state and has only one
population (of constant size n) to maintain. The main advan-
tage of SEAMO is the simple approach it uses to dispose of
all selection mechanisms based on fitness or rank. Instead,
the search progresses based on three simple rules:
(1) Parents can only be replaced by their own offspring.
(2) Duplicates in the population are deleted.
(3) Offspring can only replace parents if superior elitism.
Genetic operators are applied to each parent in turn to form
a new child, which is considered for substitution into the par-
ent population based on the three rules. This process is re-
peated until termination.
83
3.2. Recombination and mutation
Each of the algorithms considered has a specific method for
selecting parents. SPEA2 bases selection on fitness, NSGA-
II on the crowded comparison operator, PESA on non-
dominated members of its archive set, and SEAMO uni-
formly. However, common recombination and mutation op-
erators have been used to maintain a fair comparison between
the algorithms. The well-known cycle crossover [43] has been
used as the recombination operator and the mutation operator
involves the simple transposition of candidate base station lo-
cations in a randomly selected pair of positions. This was
governed by a mutation rate (set to 1%) to restrict the fre-
quency of mutation.
3.3. Measuring the relative performance of GAs
Comparing the performance of multiple objective algorithms
is problematic because a set of solutions rather than a single
solution is obtained. Although several alternatives have been
proposed (e.g., [16,48,56,57] ) no single approach seems most
prevalent. We adopt the approach first given in [57] to calcu-
late a set coverage metric. This involves the concept of weak
domination. Solution A weakly dominates solution B if A
and B have the same performance across all objectives or A
dominates B. For two sets of solutions SA and SB , the set
coverage metric of set SA with respect to SB is the percent-
age of solutions in SB which are weakly dominated by at least
one solution from SA . The higher the set coverage metric ob-
tained, the greater the superiority of SA over SB .
4. Results
The performance of the algorithms have been compared using
a wide range of synthesized test problems, each of which has
been randomly generated. Each test problem gives the loca-
tion and cost of candidate sites. Test problems are classified
in two ways: the size of area in which they are positioned and
the density of candidate sites, as documented in figure 2.
Combining the size of regions and the density of sites leads
to a total of nine test problem classes, as indicated in figure 2.
For each test problem class, we produce five incidences on
which each algorithm is tested. This means that average algo-
rithm performance is estimated and compared using five prob-
lem instances from each of nine classes, with four different al-
gorithms, leading to a total of 180 experiments. All problem
instances are available at: http://www.cs.cf.ac.uk/
user/L.Raisanen/downloads.html.
Region size km2 Density of sites per km2
0.03 0.06 0.12
15 15 7 14 28
30 30 27 54 108
45 45 61 122 244
Figure 2. Number of candidate sites in nine problem classes defined by region
size and density.
84 RAISANEN AND WHITAKER

Table 1
The average set coverage values obtained in each problem class, for all pairwise comparisons of algorithms.

Set coverage metrics

Algorithm Problem instances (km2 number of candidate sites)

SA SB 15-7 15-14 15-18 30-27 30-54 30-108 45-61 45-122 45-244 Average

SEAMO SPEA2 100.00 92.00 86.83 91.34 43.23 19.03 39.17 7.85 18.00 55.27
NSGA-II 100.00 92.00 84.33 92.80 39.81 13.71 37.15 9.20 11.41 53.38
PESA 100.00 96.00 97.50 96.80 60.29 47.74 63.32 38.05 29.90 69.96
Average 100.00 93.33 85.56 93.64 47.78 26.83 46.55 18.37 19.77 59.53
SPEA2 SEAMO 100.00 100.00 100.00 100.00 95.38 94.55 93.74 92.11 82.64 95.38
NSGA-II 100.00 100.00 97.50 100.00 74.80 69.67 86.56 52.04 49.97 81.17
PESA 100.00 100.00 100.00 100.00 86.47 86.75 97.65 83.98 77.21 92.45
Average 100.00 100.00 99.17 100.00 85.55 83.65 92.65 76.04 69.94 89.67
NSGA-II SEAMO 100.00 100.00 100.00 100.00 94.83 98.18 93.74 84.33 81.79 94.76
SPEA2 100.00 100.00 100.00 96.36 75.34 80.96 92.36 65.91 47.14 84.23
PESA 100.00 100.00 100.00 100.00 91.67 88.99 100.00 87.22 82.22 94.46
Average 100.00 100.00 100.00 98.79 87.28 89.38 95.38 79.16 70.38 91.15
PESA SEAMO 100.00 90.00 70.89 98.33 45.43 80.93 50.57 53.22 65.21 72.73
SPEA2 100.00 86.00 67.33 93.01 31.29 75.53 35.80 31.28 22.28 60.28
NSGA-II 100.00 86.00 64.83 94.46 35.35 76.13 35.12 26.91 12.31 59.01
Average 100.00 87.33 67.69 95.27 37.36 77.53 40.50 37.14 33.27 64.01

Power setting dBW
p1 30
p2 27
p3 24
p4 21
p5 18
Figure 3. Power settings used in tests.
To maintain a fair comparison between algorithms, com-
mon parameter settings have been adopted for each experi-
ment. Five nonzero power settings (displayed in units of dBW
and Watts) have been used, as specified in figure 3. Unless
otherwise specified, a population size of 100 is adopted us-
ing 500 generations. Additionally, the same random starting
populations have been used for each problem class.
We consider the performance of each GA in four ways:
(1) the average performance (in terms of the objective values
of members in the final population) compared to other
GAs across all test problems using the set coverage met-
ric, with diagrams to show obtained Pareto fronts and cell
plans,
(2) the average measure of population spread,
(3) the speed of convergence to solutions in the final popula-
tion, and
(4) the average speed of execution.
4.1. Average performance across test problems
Watts erage and cost, closely followed by SPEA2 (89.67%), then
1000 PESA (64.01%) and, finally, SEAMO (59.53%). See table 1
501 for details.
251 In figure 4 we plot the Pareto fronts (i.e., non-dominated
125
solutions from the final population) achieved by each algo-
63
rithm on the large region problem at each density. Despite
the differences in relative algorithm performance, the Pareto
fronts obtained are closely clustered in real terms. Generally,
the plots show that that as candidate site density increases, so-
lutions with a higher level of coverage are achievable. Also, in
the most dense problem instances, lower cost solutions with
higher coverage are achievable due to more freedom in site
selection and cost.
In figure 5, we display an example of cell plans with the
highest coverage for the large region problem at each density,
with density increasing left to right.
4.2. Measure of solution distribution
To measure the distribution, or spread, of solutions along the
Pareto front, a metric proposed in [41] has been implemented.
The spacing measure is based on the range of values for di ,
which is the distance (in terms of solution space) between the
ith element of the solution set and its nearest neighbour. The
average of di values for a solution set of size n is denoted d.
Then the measure of spread is defined as:


 1  n

2
S=  d di (1)
n1
i=1

In terms of the average performance of each GA compared
to other GAs across all test problems using the set coverage
metric, it was found that NSGA-II achieves the best perfor-
mance by weakly dominating an average of 91.15% of solu-
tions obtained by other algorithms in terms of service cov-
for the n members in the final population. Note that S = 0
indicates all members of the Pareto front are spaced equidis-
tantly in the solution space.
It was found that PESA performed the best on this mea-
sure, with an average spacing value of 19.75, followed by
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS 85

Figure 4. Pareto fronts (coverage versus cost) for large problem size instances (v4: 45 45) with 61, 122, and 244 candidate sites.

Figure 5. Example cell plans with the highest coverage at each density for the large region problem.

Algorithm Problem instances (km2 number of candidate sites)

15-7 15-14 15-28 30-27 30-54 30-108 45-61 45-122 45-244 Average
SEAMO 3.94 19.63 13.07 14.32 39.92 53.68 25.24 38.07 33.71 26.84
SPEA2 3.94 21.31 13.90 14.94 13.83 15.37 29.35 33.36 42.17 20.91
NSGA-II 3.94 21.31 15.02 18.05 25.63 14.84 27.34 30.16 36.83 21.46
PESA 3.94 19.29 8.50 14.09 21.53 29.78 18.57 23.81 38.26 19.75

Figure 6. Average spacing values by algorithm for each test problem class.

Algorithm Problem instances (km2 number of candidate sites)

15-7 15-14 15-28 30-27 30-54 30-108 45-61 45-122 45-244
SEAMO 7.2 17.4 37.8 104.0 242.8 542.2 526.6 1272.8 2854.8
SPEA2 7.8 18.4 40.2 107.0 250.8 553.0 543.8 1305.6 3089.0
NSGA-II 7.2 17.6 39.0 106.0 247.2 548.6 539.6 1300.0 3119.6
PESA 7.4 17.2 38.8 111.0 246.6 555.2 578.4 1293.8 2963.2
Average 7.4 17.7 39.0 107.0 246.9 549.8 547.1 1293.1 3006.7

Figure 7. Average speed of execution in seconds.

SPEA2 at 20.91, NSGA-II at 21.46, and SEAMO at 26.84
(see figure 6 for details). It is little surprise that the algorithm
which is specifically designed to encourage spacing in objec-
tive space, PESA, performed the best, and that the algorithm
with no direct measure to control dispersion, SEAMO, per-
formed the worst.
4.3. Convergence
The ability of an algorithm to rapidly converge to the final
solution set is desirable. This has been investigated for each
algorithm using the largest problem class with the highest site
density. Each algorithm has been applied for 1500 genera-
tions, and intermediate populations (produced every 250 gen-
erations) have been compared against each other, using the
set coverage metric (see section 3.3). The results indicate
that PESA and SEAMO converge quickly, by generation 1000
dominating 88.88% and 77.77% of solutions in the final gen-
eration (1500), respectively, whereas SPEA2 and NSGA-II
converge more slowly dominating 22.22% and 33.33% re-
spectively (see figure 8 for details). However, considering
that SPEA2 and NSGA-II outperform PESA and SEAMO in
terms of final solutions (see section 4.1), this may also indi-
cate that these algorithms are better at improving solutions
over time than the other two.
86 RAISANEN AND WHITAKER

Algorithm Generation Generation forming SB conceptually elegant, easy to implement, and fast to run, but
forming SA 250 500 750 1000 1250 1500
this simplicity appears to impede the overall quality and dis-
SEAMO 250 12.50 0.00 0.00 0.00 0.00
500 100.00 28.57 0.00 0.00 0.00 tribution of the solutions obtained, with SEAMO obtaining
750 100.00 100.00 71.43 16.67 14.29 the lowest performance measures in these areas. On balance,
1000 100.00 100.00 100.00 33.33 28.57 we consider NGSA-II to be the strongest performing algo-
1250 100.00 100.00 100.00 100.00 71.43 rithm for purposes of cell planning when using the general
1500 100.00 100.00 100.00 100.00 100.00
framework proposed. This is mainly based on the consistent
SPEA2 250 0.00 0.00 0.00 0.00 0.00
500 100.00 27.27 11.11 9.09 9.09 comparative quality of the solutions obtained.
750 100.00 100.00 22.22 18.18 9.09
1000 100.00 100.00 100.00 54.55 36.36
1250 100.00 100.00 100.00 100.00 45.45 References
1500 100.00 100.00 100.00 100.00 100.00
NSGA-II 250 0.00 0.00 0.00 0.00 0.00 [1] E.H.L. Aarts and J.H.M. Korst, Simulated Annealing and Boltzmann
500 100.00 9.09 8.33 33.33 33.33 Machines (Wiley, New York, 1989).
750 100.00 100.00 33.33 33.33 33.33 [2] R.G. Akl, M.V. Hedge, M. Naraghi-Pour and P.S. Min, Multicell
1000 100.00 100.00 100.00 75.00 58.33 CDMA network design, IEEE Transactions on Vehicular Technology
1250 100.00 100.00 100.00 100.00 58.33 50 (2001) 711722.
1500 100.00 100.00 100.00 100.00 100.00 [3] S.M. Allen, S. Hurley, R.K. Taplin and R.M. Whitaker, Automatic cell
PESA 250 0.00 0.00 0.00 0.00 0.00 planning of broadband fixed wireless networks, in: Proc. of the IEEE
500 100.00 55.55 55.55 44.44 44.44 VTC Conf. (Spring), Rhodes, Greece (May 2001) pp. 28082812.
750 100.00 100.00 88.88 77.77 66.67 [4] E. Amaldi, A. Capone and F. Malucelli, Improved models and algo-
1000 100.00 100.00 100.00 77.77 66.67 rithms for UMTS radio planning, in: Proc. of the 54th IEEE Conf. on
1250 100.00 100.00 100.00 100.00 88.89 Vehicular Technology, Vol. 2 (2001) pp. 920924.
1500 100.00 100.00 100.00 100.00 100.00 [5] E. Amaldi, A. Capone and F. Malucelli, Optimizing base station sit-
ing in UMTS networks, in: Proc. of the 53th IEEE Conf. on Vehicular
Figure 8. Comparison of intermediate populations for each algorithm, using
Technology, Vol. 4 (2001) pp. 28282832.
the set coverage metric for a total of 1500 generations.
[6] H.R. Anderson and J.P. McGeehan, Optimizing microcell base station
locations using simulated annealing techniques, in: Proc. of the 44th
IEEE Conf. on Vehicular Technology (1994) pp. 858862.
4.4. Speed of execution [7] S. Bleuler, M. Brack, L. Thiele and E. Zitzler, Multiobjective genetic
programming: Reducing bloat using SPEA2, in: Proc. of the Congress
The average execution times varied from an average of 7.4 on Evolutionary Computation 2001, Vol. 1 (2001) pp. 536543.
seconds to complete the smallest problem, to 3006.7 seconds [8] P. Calegari, F. Guidec, P. Kuonen and D. Wagner, Genetic approach to
(roughly 50 minutes) to complete the largest, with SEAMO radio network optimizations for mobile systems, in: Proc. of the 47th
IEEE Conf. on Vehicular Technology, Vol. 2 (1997) pp. 755759.
performing the fastest marginally (see figure 7 for details).
[9] B. Chamaret, S. Josselin, P. Kuonen, M. Pizarroso, B. Salas-
All recorded times were obtained using a Pentium IV 1.8 GHz Manzanedo, S. Ubeda and D. Wagner, Radio network optimization with
processor, 256 megabytes of RAM, Java JDK 1.3.1-04, and maximum independent set search, in: Proc. of the IEEE VTC97 Conf.,
Windows XP Professional. It is suspected that obtaining only Phoenix, AZ (May 1997) pp. 770774.
marginal differences in execution time was due to a bottleneck [10] D.W. Corne, J.D. Knowles and M.J. Oates, The Pareto envelope-based
selection algorithm for multiobjective optimization, in: Proc. of the
incurred by the computationally intensive decoder.
Sixth Internat. Conf. on Parallel Problem Solving from Nature (2000)
pp. 839848.
[11] M.S. Daskin, Network and Discrete Location (Wiley, New York, 1995).
5. Conclusions [12] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms
(Wiley, New York, 2001).
In this paper we have have introduced a general framework for [13] K. Deb, S. Agrawal, A. Pratap and T. Meyarivan, A fast elitist non-
applying multiple objective genetic algorithms to the antenna dominated sorting genetic algorithm for multi-objective optimization:
placement problem. The key aspect of this framework is a NSGA-II, in: Lecture Notes in Computer Science, Vol. 1917 (Springer,
New York, 2000) pp. 848849.
decoder which uses an ordering of candidate site locations to
[14] K. Deb, L. Thiele, M. Laumanns and E. Zitzler, Scalable test prob-
construct a cell plan. Subsequently, the performance of four lems for evolutinary multi-objective optimization, Kangal Report
GAs to find an optimal ordering of potential site locations No. 2001001 (2001) 127.
were compared and evaluated using a range of test problems [15] E. Ekici and C. Ersoy, Multi-tier cellular network dimensioning, Wire-
classified by size and density. less Networks 7 (2001) 401411.
[16] C.M. Fonseca and P.J. Fleming, On the performance assessment and
We found that all the algorithms considered find closely
comparison of stochastic multiobjective optimizers, in: Fourth Internat.
comparable solutions, in real terms. However, there are differ- Conf. on Parallel Problem Solving from Nature (1996) pp. 584593.
ences. NGSA-II and SPEA2 have very similar performance [17] M. Galota, C. Glasser, S. Reith and H. Vollmer, A polynomial-time ap-
throughout, confirming the findings in [55] concerning the proximation scheme for base station positioning in UMTS networks, in:
performance of these algorithms. Meanwhile, PESA gener- Proc. of the 5th Internat. Workshop on Discrete Algorithms and Meth-
ods for Mobile Computing and Communications, Rome, Italy (July
ally obtains slightly lower quality sets of solutions, but has
2001) pp. 5259.
the best performance in terms of distribution of solutions and [18] A. Ganz, C.M. Krishna, D. Tang and Z.J. Haas, On optimal design
speed of convergence. Both the advantage and disadvantage of multiter wireless cellular systems, IEEE Communications Magazine
of SEAMO lie in its simplicity. This makes the algorithm (February 1997) 8893.
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS
[19] F. Glover, Heuristics for integer programming using surrogate con-
straints, Decision Sciences 8 (1977) 156166.
[20] F. Glover, E. Taillard and D. de Werra, A users guide to tabu search,
Annals of Operations Research 41 (1993) 328.
[21] J.K. Han, B.S. Park, Y.S. Choi and H.K. Park, Genetic approach with a
new representation base station placement in mobile communications,
in: Proc. of the 54th IEEE Conf. on Vehicular Technology, Vol. 4 (2001)
pp. 27032707.
[22] M. Hata, Empirical formula for propogation loss in land mobile ra-
dio services, IEEE Transactions on Vehicular Technology 29(3) (1980)
317325.
[23] J.H. Holland, ed., Adaptation in Natural and Artificial Systems (Uni-
versity of Michigan Press, Ann Arbor, 1975).
[24] I. Howitt and S.-Y. Ham, Base station location optimization, in: Proc.
of the IEEE VTC99 Conf., Vol. 4 (1999) pp. 20672071.
[25] X. Huang, U. Behr and W. Wiesbeck, Automatic cell planning for a
low-cost and spectrum efficient wireless network, in: Proc. of Global
Telecommunications Conf. (GLOBECOM), Vol. 1 (2000) pp. 276
282.
[26] S. Hurley, Planning effective cellular mobile radio networks, IEEE
Transactions on Vehicular Technology 51(2) (2002) 243253.
[27] S. Hurley, D.H. Smith and C.L. Valenzuela, A permutation-based ge-
netic algorithm for minimum span frequency assignment, in: Lecture
Notes in Computer Science, Vol. 1498 (Springer, New York, 1998)
pp. 907916.
[28] L.J. Ibbetson and L.B. Lopes, An automatic base station placement al-
gorithm, in: Proc. of the IEEE VTC97 Conf., Phoenix, AZ (May 1997)
pp. 770774.
[29] V. Khare, X. Yao and K. Deb, Performance scaling of multi-objective
evolutionary algorithms, KanGAL Report No. 2002009 (2002) pp. 1
15.
[30] J.D. Knowles, Local-search and hybrid evolutionary algorithms for
Pareto optimization, Ph.D. thesis, The University of Reading (2002).
[31] I. Laki, L. Farkas and L. Nagy, Cell planning in mobile communication
systems using SGA optimization, in: Proc. of Internat. Conf. on Trends
in Communications, Vol. 1 (2001) pp. 124127.
[32] C.Y. Lee and H.G. Kang, Cell planning with capacity expansion in mo-
bile communications: A tabu search approach, IEEE Transactions on
Vehicular Technology 49(5) (2000) 16781691.
[33] R.M. Mathar and T. Niessen, Optimum positioning of base stations for
cellular radio networks, Wireless Networks 6 (2000) 421428.
[34] R.M. Mathar and M. Schmeink, Optimal base station positioning and
channel assignment for 3G mobile networks by integer programming,
Annals of Operations Research 107 (2001) 225236.
[35] R.M. Mathar and M. Schmeink, Integrated optimal cell site selection
and frequency allocation for cellular radio networks, Telecommunica-
tion Systems 21 (2002) 339347.
[36] H. Meunier, E. Talbi and P. Reininger, A multiobjective genetic algo-
rithm for radio network optimization, in: Proc. of the 2000 Congress
on Evolutionary Computation, Vol. 1 (2000) pp. 317324.
[37] A. Molina, G.E. Athanasiadou and A.R. Nix, The automatic location
of base-stations for optimised cellular coverage: A new combinatorial
approach, in: Proc. of the IEEE VTC99 Conf. (1999) pp. 606610.
[38] A. Molina, G.E. Athanasiadou and A.R. Nix, Optimised base-station
location algorithm for next generation microcellular networks, Elec-
tronics Letters 36(7) (2000) 668669.
[39] A. Molina, G.E. Nix and A.R. Athanasiadou, The effects of delay
spread for cellular network planning using the combination algorithm
for total optimisation, in: Proc. of the 1st Internat. Conf. on 3G Mobile
Communication Technologies (2000) pp. 171175.
[40] P. Reininger, S. Iksal, A. Caminada and J.J. Korczak, Multi-stage op-
timization for mobile radio network planning, in: Proc. of the IEEE
VTC99 Conf., Vol. 3 (1999) pp. 20342038.
[41] J. Schott, Fault tolerant design using simple multicriteria genetic algo-
rithms, Ph.D. thesis, M.S. thesis, Department of Aeronautics and As-
tronautics, MIT, Cambridge, MA (1995).
[42] H.D. Sherali, C.H. Pendyala and T.S. Rappaport, Optimal location
of transmitters for micro-cellular radio communication system design,
87
IEEE Journal on Selected Areas in Communications 14(4) (1996) 662
673.
[43] T. Starkweather, S. McDaniel, K. Mathias, D. Whitley and C. Whitley,
A comparison of genetic sequencing operators, in: Proc. of the Fourth
Internat. Conf. on Genetic Algorithms, eds. R. Belew and L. Booker
(Morgan Kaufman, San Mateo, CA, 1991) pp. 6976.
[44] D.-W. Tcha and Y.-S. Myung, Base station location in a cellular CDMA
system, Telecommunication Systems 14 (2000) 163173.
[45] K. Tutschku, Interference minimization using automatic design of cel-
lular communication networks, in: Proc. of the IEEE VTC98 Conf.
(1998) pp. 634638.
[46] C.L. Valenzuela, A simple evolutionary algorithm for multi-objective
optimisation(SEAMO), in: IEEE Congress on Evolutionary Computa-
tion (2002) pp. 717722.
[47] M. Vasquez and J.-K. Hao, A heuristic approach for antenna positioning
in cellular networks, Journal of Heuristics 7 (2001) 443472.
[48] D.A. Veldhuizen and G.B. Lamont, On measuring multiobjective evo-
lutionary algorithm performance, in: Congress on Evolutionary Com-
putation (2000) pp. 204211.
[49] R.M. Whitaker and S. Hurley, Omni-directional cell planning, in:
Telecommunications Network Design and Management, eds. G. Anan-
dalingam and S. Raghavan (Kluwer Academic, Dordrecht, 2002) chap-
ter 2, pp. 2541.
[50] R.M. Whitaker and S. Hurley, The state-of-the-art in automatic cell
planning, submitted for publication.
[51] J.K.L. Wong, M.J. Neve and K.W. Sowerby, Optimisation strategy for
wireless communications system planning using linear programming,
IEE Electronics Letters 37(17) (2001) 10861087.
[52] M.H. Wright, Optimization methods for base station placement in wire-
less systems, in: Proc. of the IEEE VTC98 Conf. (1998) pp. 387
391.
[53] J. Zimmermann, R. Hons and H. Muhlenbein, ENCON: Evolutionary
algorithm for the antenna placement problem, Computers and Industrial
Engineering 44 (2003) 209226.
[54] E. Zitzler, Evolutionary algorithms for multiobjective optimization:
Methods and applications, Ph.D. thesis, Swiss Federal Institute of Tech-
nology, Zurich, Switzerland, TIK-Schriftenreihe Nr. 30 (1999).
[55] E. Zitzler, M. Laumanns and L. Thiele, SPEA2: Improving the strength
Pareto evolutionary algorithm, Technical Report 103, Computer En-
gineering and Networks Laboratory (TIK), ETH Zurich, Switzerland
(2001).
[56] E. Zitzler, M. Laumanns, L. Thiele, C. Fonseca and G. da Fonseca, Per-
formance assessment of multiobjective optimizers: An analysis and re-
view, Technical Report 139, Computer Engineering and Networks Lab-
oratory (TIK), ETH Zurich, Switzerland (2002).
[57] E. Zitzler and L. Thiele, Multiobjective optimization using evolutionary
algorithms a comparative case study, in: Parallel Problem Solving
from Nature (1998) pp. 292301.
Larry Raisanen is a second year Ph.D. student
studying at Cardiff University at the Centre for Mo-
bile Communications based in the School of Com-
puter Science. The focus of his work is on the devel-
opment, testing, and analysis of algorithms to resolve
the base station placement problem in wireless mo-
bile communication systems using a multi-objective
optimization framework. He graduated in 1993 with
a BA magna cum laude from Marquette Univer-
sity (WI, USA), and in 2003 with a M.Sc. in com-
puting distinction from Cardiff University (Wales, UK).
E-mail: L.Raisanen@cs.cf.ac.uk
88 RAISANEN AND WHITAKER

Roger M. Whitaker holds a Ph.D. degree in discrete systems. He is currently leading a number of externally supported research
mathematics (1999) and a B.Sc. degree in mathe- projects in this area.
matics and managment science. He is a lecturer and E-mail: R.M.Whitaker@cs.cf.ac.uk
a co-director of the Centre for Mobile Communica-
tions, School of Computer Science, Cardiff Univer-
sity, UK. Prior to this position, Roger carried out
research for the UK Radiocommunications Agency
into spectrum efficiency. His research addresses the
application of computer science to the design, co-
ordination and optimization of wireless networks and
 Mobile Networks and Applications 10, 8998, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

A Characterisation of Optimal Channel Assignments for Cellular

and Square Grid Wireless Networks
M.V.S. SHASHANKA
Boston University, Boston, MA, USA

AMRITA PATI
Virginia Polytechnic Institute and State University, Blacksburg, VA, USA

ANIL M. SHENDE
221 College Lane, Roanoke College, Salem, VA 24153, USA

Abstract. In this paper we first present a uniformity property that characterises optimal channel assignments for networks arranged as
cellular or square grids. Then, we present optimal channel assignments for cellular and square grids; these assignments exhibit a high value
for 1 the separation between channels assigned to adjacent stations. We prove an upper bound on 1 for such optimal channel assignments.
This upper bound is greater than the value of 1 exhibited by our assignments. Based on empirical evidence, we conjecture that the value
our assignments exhibit is a tight upper bound on 1 .
Keywords: wireless computing, channel assignment, cellular and square grids

1. Introduction  = (1 , 2 , . . . , 1 ) is said to place the channel separation

The enormous growth of wireless networks has made the ef-
ficient use of the scarce radio spectrum important. A Fre-
quency Assignment Problem (FAP) models the task of as-
signing frequencies (channels) from a radio spectrum to
a set of transmitters and receivers, satisfying certain con-
straints [9]. The main difficulty in an efficient use of the
radio spectrum is the interference caused by unconstrained
simultaneous transmissions. Interferences can be eliminated
(or at least reduced) by means of suitable channel assignment
techniques, which partition the given radio spectrum into a
set of disjoint channels that can be used simultaneously by
the stations while maintaining acceptable radio signals. Since
radio signals get attenuated over distance, two stations in a
network can use the same channel without interferences pro-
vided the stations are spaced sufficiently apart. The minimum
distance at which channels can be reused with no interfer-
ences is called the co-channel reuse distance (or simply reuse
distance) and is denoted by .
In a dense network a network where there are a large
number of transmitters and receivers in a small area inter-
ference is more likely. Thus, reuse distance needs to be high
in such networks. Moreover, channels assigned to nearby sta-
tions must be separated in value by at least a gap which is
inversely proportional to the distance between the two sta-
tions. A minimum channel separation i is required be-
tween channels assigned to stations at distance i, with i < ,
such that i decreases when i increases [8]. is said to
place the co-channel reuse distance constraint, and the vector
This work was partially funded by NSF grant 0200823.
Corresponding author.
constraint on the channel assignment problem.
The purpose of channel assignment algorithms is to assign
channels to transmitters in such a way that (1) the co-channel
reuse distance and the channel separation constraints are sat-
isfied, and (2) the span of the assignment, defined to be the
difference between the highest and the lowest channels as-
signed, is as small as possible [2].
This paper has two significant contributions:
1. A characterisation of optimal channel assignments for cel-
lular and square grids. We essentially show a nice unifor-
mity across the grid that every optimal assignment must
satisfy. (See section 3.)
2. Optimal channel assignments for cellular and square grids
where the channel separation between adjacent stations is
large. We prove an upper bound on i for such optimal
channel assignments. This upper bound is greater than the
value of i exhibited by our assignments. Based on empiri-
cal evidence, we conjecture that the value our assignments
exhibit is a tight upper bound on 1 . (See section 4.)
In section 2 we formally define the problem of channel
assignments and its formulation as a colouring problem, and
provide a brief literature survey. We also outline the general
strategy we use for our optimal colourings discussed in sec-
tion 4. In section 3 we first define the cellular (section 3.1) and
square (section 3.2) grids, and point out some useful proper-
ties of these grids. Then, in section 3.3, we prove a charac-
terisation of optimal colourings for cellular and square grids.
In section 4 we present our colourings and prove that they are
optimal. Then, in section 5, we present an upper bound on
90
the value of the channel separation among adjacent stations
as witnessed by optimal colourings.
2. Preliminaries
Formally, the Channel Assignment Problem with Separation
(CAPS) can be modelled as an appropriate colouring prob-
lem on an undirected graph G = (V , E) representing the net-
work topology, whose vertices in V correspond to stations,
and edges in E correspond to pairs of stations that can hear
each others transmission [2]. The colour assigned to a par-
ticular vertex corresponds to the frequency channel assigned
to the corresponding station. For a graph G, we will denote
the distance between any two vertices in the graph, i.e., the
number of edges in a shortest path between the two vertices,
by dG (, ). (When the context is clear, we will denote the
distance as simply d(, ).) CAPS is then defined as:
CAPS (G, , ). Given an undirected graph G, an integer
> 1, and a vector of positive integers  = (1 , 2 , . . . ,
1 ), find an integer g > 0 so that there is a function
f : V {0, . . . , g}, such that for all u, v G, for each i,
1
i
1, if d(u, v) = i, then |f (u) f (v)|  i .
This assignment is referred to as a g-L(1 , 2 , . . . ,
1 ) colouring of the graph G [7], and CAPS (G, , )  is

sometimes referred to as the L() colouring problem for G.
Note that a g-L(1 , 2 , . . . , 1 ) uses only the (g + 1)
colours in the set {0, . . . , g}, but does not necessarily use all
the (g + 1) colours. A g-L(1 , 2 , . . . , 1 ) colouring of G
is optimal iff g is the smallest number witnessing a solution
for CAPS (G, , ).  The lattice Zd is the set of ordered d-tuples of integers, and
Finding the optimal colouring for general graphs has
been shown to be NP-complete. The problem remains
NP-complete even if the input graphs are restricted to planar
graphs, bipartite graphs, chordal graphs, and split graphs [4].
Most of the work on this problem has dealt with specific
graphs such as grids and rings, for small reuse distance ( )
values, and for small channel separation (i ) values, e.g., op-
timal L(1, 1) colourings for rings and bidimensional grids [1],
optimal L(2, 1) and L(2, 1, 1) colourings for hexagonal, bidi-
mensional, and cellular grids [2], etc. Recently, Bertossi
et al. [3] exhibited optimal L(1 , 1, . . . , 1) colourings, for
1
/2, for bidimensional grids and rings. (See [3] for a
succinct literature survey of this problem.) Below, we refer to
L(, 1, . . . , 1) colourings by L(, 1k ) colourings.
As pointed out in [2], a lower bound for the L(1, 1 k )
colouring problem is also a lower bound for the L(, 1 k ),
> 1. Given an instance of CAPS, consider the augmented
graph obtained from G by adding edges between all those
pairs of vertices that are at a distance of at most 1. Clearly,
then, the size (number of vertices) of any clique in this aug-
mented graph places a lower bound on an L(1, 1 1 ) colour-
ing for G; the best such lower bound is given by the size of a
maximum clique in the augmented graph.
In each graph G, for each , we identify a canonical sub-
graph, T (G, ), of the graph so that the vertices of T (G, )
SHASHANKA ET AL.
induce a clique in the augmented graph of the graph. We will
refer to T (G, ) as a tile. When the context is clear, we will
refer to the size of T (G, ) simply as c( ).
Most (but not all) of the assignment schemes described in
this paper follow the pattern: for a given graph G, and for a
given ,
(1) identify T (G, ),
(2) find the number of vertices in T (G, ), and hence a lower
bound for the given assignment problem,
(3) describe a colouring scheme to colour all the vertices of
T (G, ),
(4) demonstrate a tiling of the entire graph made up of
T (G, ) to show that the colouring scheme described
colours the entire graph, and
(5) show that the colouring scheme satisfies the given reuse
distance and channel separation constraints.
3. A characterisation of optimal colourings
We first introduce the conventions we follow to represent
square grids and cellular grids. We explain tilings in both
grids, and define some notation. Then we present our charac-
terisation of optimal colourings in cellular and square grids.
For any d-dimensional lattice L, the minimal distance in
the lattice is denoted by (L). The infinite graph, denoted
G(L), corresponding to the lattice L consists of the set of lat-
tice points as vertices; each pair of lattice points that are at a
distance (L) constitute the edges of G(L).
Ad is the hyperplane that is a subset of Zd+1 , and is charac-
terised as the set of points in Zd+1 such that the coordinates
of each point add up to zero. (Zd ) = 1, and the minimal
length vectors in Zd are the unit vectors in each dimension.
For each d > 0, for each i, j, 0
i, j
d, i = j , define
dij = (x0 , . . . , xd ), where xi = 1, xj = 1, and for each
k, 0
k
d, k = i, j, xk = 0. Then, (Ad ) = 2, and
the set of minimal length vectors in Ad is {dij | i, j, 0
i,
j
d, i = j }. (See [5,10] for more on these lattices.)
The infinite 2-dimensional square grid is, then, G(Z2 ), and
the infinite 2-dimensional cellular grid is G(A2 ).
3.1. Cellular grids
For a given value of , two kinds of tiles can be identified:
triangular and hexagonal. The tiles are shown in figure 1(a).
It can be easily shown that:
Lemma 1. 1. The number of vertices in a triangular tile cor-
responding to reuse distance , denoted by cT ( ), is given by
cT ( ) = ( + 1)/2.
2. The number of vertices in a hexagonal tile corre-
sponding to reuse distance , denoted by cH ( ), is given by
cH ( ) = (3 2 +( mod 2))/4. Thus, when = 2k +1(odd),
c( ) = 3k 2 + 3k + 1.
A CHARACTERISATION OF OPTIMAL CHANNEL ASSIGNMENTS 91

Figure 1. Cliques in cellular and square grids.

Figure 2. Basis vectors in A2 . Figure 4. Tiling of A2 and Z2 .

= 0 mod 4 = 2 mod 4 = 1, 3 mod 4 six neighbouring tiles. For any tile H , we will refer to the
(0, 0) (0, 0) (0, 0) neighbouring tiles as H0 , H1 , . . . , H5 as shown in figure 4.



  
1, 0 ,0 ,0 If the coordinates of the origin of H are (0, 0), then the ori-

2 2 2  


gins of H0 , H1 , . . . , H5 will have coordinates (/2, ),
1, 1, 1 1,
2 2 2 ( /2
, /2), (, /2
), (/2, ), ( /2
, /2)
( 1, 1) ( 1, 1) (a 1, 1)



   and (, /2
), respectively. These points are marked in
1, 1 , 1 , 1 figure 4. The edge of tile H which is adjacent to the tile Hi
2

2
2  
  will be denoted by ti .
0, 0, 1 0,
2 2 2

Figure 3. Coordinates of the corners of a hexagonal tile.

From the above lemma, we observe that for a given value
of , the size of a hexagonal tile is greater than the size of a
triangular tile. As mentioned in the previous section, the size
of the maximum clique, i.e. cH ( ) places a lower bound on
the colouring of G(A2 ). Henceforth, for G(A2 ), we consider
only hexagonal tiles and the word tile refers to hexagonal tile
unless otherwise mentioned. Also, we refer to cH ( ) simply
as c( ).
For a particular , hexagons are regular with sides of
/2
vertices if is odd. In case a is even, alternate sides As mentioned in section 2, the size of the maximum clique
of the hexagon are equal and consecutive sides have /2 and
( + 1)/2
vertices, respectively. G(Z2 ). The following lemma gives a formula for the size of
Figure 2 shows the coordinate system we use for represent-
ing vertices in A2 where (0, 1, 1) and (1, 1, 0) indicate the
basis vectors i and j . In the table in figure 3, we list the co-
ordinates of the corners of a hexagon in clockwise order, for
.
various values of . We start with the left-most vertex, which
we refer to as the origin and assign (0, 0) as its coordinates.
Consider the arrangement of tiles as shown in figure 4. It is
clear that such an arrangement will tile all of A2 . Note that
any translation of this tiling will also tile A2 .
vertices. They tile the entire grid Z2 . In
In such a tiling, we will refer to two tiles as neighbours if
there is an edge e1 of one and an edge e2 of the other such
that at least two points on e1 have neighbours on e2 and vice
versa. In such a tiling of A2 , every tile is surrounded by
Definition 1. In a cellular grid tile, we define a diagonal to be
a line formed by all vertices having the same ith coordinate.
In a tile with origin (io , jo ), diagonal corresponding to the
coordinate ic is represented as Licio and (ic io ) is called
the diagonal number.
In a tile corresponding to reuse distance , there are
diagonals. Figure 5 shows the diagonals in a cellular grid.
3.2. Square grids
for a particular places a lower bound on the colouring of
such a clique which is also referred to as a tile.
Lemma 2. The number of vertices in a tile corresponding to
reuse distance , denoted by c( ) is given by c( ) = 2 /2
Thus, when = 2k + 1 (odd), c( ) = 2k 2 + 2k + 1.
For a particular , tiles are diamonds with their diagonals
along the X and Y axes (as shown in figure 1(b)) and every
side contains /2
the case of odd , every corner of the tile corresponds to
a vertex on the grid. We use the vectors (1, 0) and (0, 1)
as the basis vectors i and j for representing points in Z2 .
Then, the coordinates of the vertices of the corners of a tile
92 SHASHANKA ET AL.

Figure 5. Verticals and diagonals.

Figure 7. Bounding box B(p) with edges marked.

Figure 6. Possible tiling of Z2 and A2 for odd .

(in clockwise order, starting with the left-most vertex) are Definition 3. In a square grid tile, we define a diagonal to
(0, 0), (/2, /2), ( 1, 0) and (/2, /2). In be a line of the form i j = c, where c is a constant. It
the case of even , only opposite corners of the tile along the
is represented as Di , where i, called the diagonal number is
X direction correspond to vertices on the grid, their coordi-
given by (i j ) mod .
nates being (0, 0) and ( 1, 0).
Consider the arrangement of tiles as shown in figure 4. It
is clear that such an arrangement will tile all of Z2 . Note that In a tile corresponding to reuse distance , there are di-
any translation of this tiling will also tile Z2 . agonals/verticals as the case may be. Figure 5 shows verticals
In such a tiling, we will refer to two tiles as neighbours if and diagonals of a square grid tile.
there is an edge e1 of one and an edge e2 of the other such
that at least two points on e1 have neighbours on e2 and vice Definition 4. 1. Consider a point p in a square/cellular grid
versa. In such a tiling of Z2 , every tile is surrounded by four and consider all points which are at a distance from p,
neighbouring tiles. For any tile H , we will refer to the neigh- where is the reuse distance. In the case of square grids,
bouring tiles as H0 , H1 , H2 and H3 . If the coordinates of the all these points form a diamond centered at p and in the case
left-most vertex of H (which we refer to as the origin) are of cellular grids, they form a hexagon centered at p. This di-
(0, 0), then the origins of H0 , H1 , H2 and H3 will have coor- amond/hexagon will be called the bounding box surrounding
dinates (/2, /2
), ( /2
, /2), (/2, /2
) point p and will be denoted by B(p). The edges, considered
and ( /2
, /2), respectively. These points are shown in a clockwise direction, are denoted by d0 , d1 , . . . , d3 in case
in figure 4. The edge of tile H which is adjacent to the tile Hi
of square grids and d0 , d1 , . . . , d5 in case of cellular grids, as
will be denoted by ti .
shown in figure 7.
There is another kind of tiling possible in both cellular and
2. Consider the bounding box for point p. Every edge
square grids for odd reuse distances, as shown in figure 6 for
contains 1 vertices apart from the two corners. Each cor-
= 5. We shall refer to the tiling shown in figure 4 as tiling A
ner, which belongs to two edges di and di+1 , is taken to be
and the one in figure 6 as tiling B.
a part of the second edge di+1 , where i refers to (i mod s), s
Definition 2. In a square grid tile, we define a vertical to be being 4 in case of square grids and 6 in case of cellular grids.
a line formed by all vertices having the same X-coordinate. For each edge, we number the vertices consecutively, clock-
In a tile with origin (io , jo ), a vertical corresponding to the wise, starting with 0 being assigned to the left-corner vertex.
coordinate ic is represented as Vic io and (ic io ) is called These numbers are called position numbers. This is shown in
the vertical number. figures 811.
A CHARACTERISATION OF OPTIMAL CHANNEL ASSIGNMENTS
Figure 8. Cellular grid bounding box for = 4 with position numbers.
Figure 9. Cellular grid bounding box for = 5 with position numbers.
Figure 10. Square grid bounding box for = 4 with position numbers.
1 or
.
3.3. Optimal colouring schemes
A colouring scheme is optimal if it uses the smallest possible
number of colours. In other words, a colouring which uses
colours from the set {0, 1, . . . , g} will be optimal if it uses the
smallest possible value for g. From lemmas 1 and 2, we know
that c( ) is a lower bound on the number of colours used. We
are concerned only with such colouring schemes which use
exactly c( ) different colours.
We already know that is the minimum distance at which
93
Figure 11. Square grid bounding box for = 5 with position numbers.
Lemma 3. Consider an optimal colouring scheme for a wire-
less network modelled as an infinite square or cellular grid
with reuse distance . For a given point p, there exists at least
one point at distance from p which has the same colour
as p.
Proof. Let us assume that, on the contrary, there is no point
at distance from p which has the same colour as p. Thus,
no point inside, or on the boundary of, B(p) is assigned the
same colour as that of p.
Now, consider one of the edges of B(p), say d0 and a tile
inside B(p) such that one of its edges t0 is completely con-
tained in this edge of B(p). Clearly, p is not in this tile. Since
we have an optimal colouring, one of the points in the tile
must be assigned the same colour as the colour assigned to p.
This is a contradiction, and hence the result. 
We now present a theorem using which we will be able to
establish an important property of optimal colouring schemes.
Theorem 1. Consider an optimal colouring scheme for a
wireless network modelled as an infinite square or cellular
grid with reuse distance . For every point p, there is a po-
sition number n, such that each point corresponding to this
position number on each edge of the bounding box surround-
ing p has the same colour as p. Moreover, n = /2
n = /2
Proof. Consider the edge d0 of the bounding box around
p, B(p). Consider the k different tiles, each of whose edge
t0 is a part of the edge d0 of B(p), where = 2k + 1 for odd
and = 2k for even . Refer to figures 811.
Let P (i) denote the sequence of position numbers on d0 of
B(p) that are on the edge t0 of the ith of these k tiles. In the
case of odd , P (i) are given by:
channels can be reused. In other words, the same colour can P (1) = 1, . . . , k + 1,
be used for vertices which are at distance or greater. The P (2) = 2, 3, . . . , k + 2,
following lemma establishes that in an optimal colouring the .. (1)
nearest vertex where a colour is reused is no more than dis- .  
tance away. P (k) = 2 + (k 2), 2 + (k 1), . . . , 2k .
94 SHASHANKA ET AL.

In case is even, P (i) are given by: The following characterisation of optimal colourings of
cellular and square grids is an immediate consequence of the-
P (1) = 1, 2, . . . , k, orem 1.
P (2) = 2, 3, . . . , k + 1,
(2)
..
. square grid by tiles (for ), a colouring with reuse distance
 
P (k) = 2 + (k 2), 2 + (k 1), . . . , 2k 1 .
Since the colouring is optimal, the colour c, that the
point p is coloured in, must appear somewhere on each of
these tiles. Except for the edge t0 , each of these tiles is com-
pletely contained within B(p). Thus, the colour c must appear
on the edge t0 of each of these tiles, otherwise, the reuse con-
straint is violated. Since no pair of vertices with position num-
bers 1, 2, . . . , 2k on the edge d0 of B(p) are at a distance ,
it must be the case that the colour c is assigned to some vertex
that is common to all the above tiles. In case of odd , as seen
from equation (1), the only two common vertices are the ones
with position numbers k and k + 1, and hence, one of these
two vertices must be assigned the colour c. In case of even ,
we see from equation (2) that the only common vertex is the
one corresponding to position number k, and hence, it has to
be assigned the colour c. A similar argument establishes that
on each edge of B(p), the vertices corresponding to position
numbers k and k + 1 in case of odd and k in case of even
are the only possible candidates for being assigned colour c.
Now, in case of odd , let q be the vertex, corresponding
to position number k on the edge d0 of B(p), that is assigned
colour c (see figures 9 and 11). Suppose, by way of contra-
diction, the vertex with position number k + 1 on the edge d1
of B(p) is assigned colour c. Let us name this vertex x. We
will now consider cellular and square grids separately in two
different cases.
Case 1 (Cellular grids). Consider the bounding box B(q).
The edge d2 of B(q) passes through the vertex with position
number k on the edge d1 of B(p) as shown in figure 9. By
the above argument, one of the two vertices with position
numbers k or k + 1 on this edge d2 of B(q) must be as-
signed colour c. But both these vertices are at a distance less
than from the vertex x. Therefore, x cannot be assigned
colour c, implying that the vertex with position number k on
the edge d1 must be assigned colour c.
Case 2 (Square grids). Consider the bounding boxes B(q)
and B(x). Let r be the point of inter-section of the edges d1
of B(q) and d0 of B(x) (see figure 11). If both q and x are
coloured c, it follows that r should be assigned the colour c.
This is not possible because r lies within the bounding box
B(p) of point p which is also coloured c. This implies that
the vertex with position number k on the edge d1 must be
assigned colour c.
Similar arguments in both cases above establish that, if the
vertex with position number k on any one edge of B(p) is
coloured the same as the colour of p, then on each edge of
B(p), the vertex with position number k is also coloured the
same as p.  same colour as (p, q).
Theorem 2. Given , and given a tiling of a cellular or
is optimal iff all the tiles in the tiling are identical in their
colour assignment.
Recall the definition of tilings A and B from section 3.2
(see figures 4 and 6). From the proof of theorem 1, we make
the following observation.
Corollary 1. Suppose = 2k + 1, and we have an optimal
colouring of the cellular (square) grid. If for any point p in
the grid, the vertex corresponding to position number k on an
edge of the bounding box of p has the same colour as that as-
signed to p, then the tiling of the grid, by identically coloured
tiles, corresponds to tiling B; if the position number is k + 1,
then the tiling of the grid corresponds to tiling A.
4. Optimal L(1 , 1 2 ) colourings for G(A2 ) and G(Z2 )
In this section, we deal with optimal frequency assignment
schemes for wireless networks modelled as cellular grids
and square grids. We first present an L(1 , 1 2 ) colouring
scheme of G(A2 ) for the case where reuse distance is odd
i.e., = 2k + 1, k {1, 2, . . .}. This is followed by an
L(1 , 1 2 ) colouring scheme of G(Z2 ) for all values of .
The colouring schemes presented here correspond to tiling A.
4.1. Cellular grids
We present a colouring scheme where 1 varies as the square
of , for  5, odd. We note that the colouring of the
entire cellular grid is achieved by colouring one tile and re-
producing the same colouring in all the tiles present in the
grid. Recall that the number of vertices c( ), in a tile cor-
responding to an odd reuse distance = 2k + 1 is equal to
3k 2 + 3k + 1 (see lemma 1). From the above fact and from
theorem 2, we make the following observations (refer to fig-
ure 12).
Lemma 4.
1. Colouring c( ) points starting from the vertex of a tile
along the direction j is equivalent to colouring all the di-
agonals of a tile in the following order:
L0 , Lk+1 , L1 , Lk+2 , . . . , Lk1 , L2k , Lk .
2. Along a line i = m, where m is a constant, any pair of
points which are at a distance c( ) apart will have the same
colour assigned to them.
3. Consider a point (p, q) on the line i = p. The point
(p + 1, q 3k 1) on the line i = p + 1 will have the
A CHARACTERISATION OF OPTIMAL CHANNEL ASSIGNMENTS 95

(i, j 1) ( (i, j ) + k 2 + k)
(i + 1, j ) ( (i, j ) + 2k 2 + 3k + 1)
(i + 1, j + 1) ( (i, j ) + k 2 + 2k + 1)
(i, j + 1) ( (i, j ) + 2k 2 + 2k + 1)
(i 1, j ) ( (i, j ) + k 2 )
(i 1, j 1) ( (i, j ) + 2k 2 + k)

Figure 13. Colours (mod c( )) assigned to neighbours of (i, j ) by the

scheme .

The colour assigned to (i, j ) according to the above

(j + i(3k + 1)).
Figure 12. L(1 , 1 2 ) colouring for = 7.
From lemma 4, we see that a colouring for c( ) points
along a line i = m for some arbitrary m describes the colour-
ing for the entire grid.
The following colouring scheme is shown in figure 12. To
colour along the line i = 0, we proceed as follows: Starting
with the point (0, 0) which is assigned the colour 0, we assign
consecutive colours to every third vertex, and wrap around
after the c( )th vertex. This will colour all the c( ) points in
three passes uniquely. This can be easily seen because c( ) =
c
(k) = 3k 2 + 3k + 1 = 1 (mod 3). Consecutive sets of c( )
vertices along this line follow the same colouring pattern.
Formally, this colouring scheme can be expressed as fol-
lows. Let (i, j ) represent the colour assigned to the vertex
(i, j ) and
(j ) represent the colour assigned to the vertex table in figure 13 twice proves the value of 2 . The value of
(0, j ), i.e.
(j ) = (0, j ). We first give a formula for
(j ) 3 is 1 by construction.
and then derive an expression for (i, j ).
, j = 0 (mod 3),
(j ) = + 2k 2 + 2k + 1, j = 1 (mod 3), We present colouring schemes where 1 varies as the square
+ k 2 + k + 1, j = 2 (mod 3),
where j = j mod c( ) and = j/3. Now, from lem-
ma 4.3, we can easily derive that
scheme will be (i, j ) =
The table in figure 13 shows the colours assigned to the
neighbours of (i, j ). (All the colour expressions are modulo
c( ).)
From the table in figure 13, we see that the least difference
between the colours assigned to neighbouring points is k 2 .
Hence, 1 = k 2 .
From the formula for (i, j ), it can be easily seen that
given any arbitrary point (i, j ) in the grid, the colour assigned
to (i, j ) can be computed in constant time. 
Lemma 5 notes the values of 2 and 3 for the above
colouring.
Lemma 5. For all = 2k + 1, k = {1, 2, . . .}, the colouring
scheme described above has the properties that 2 = k and
3 = 1.
Proof. Similar to the proof of theorem 3 above, using the

4.2. Square grids
of , for  4. There are two different schemes, one for
the case where is odd and one for even . We note that the
colouring of the entire square grid is achieved by colouring
one tile and reproducing the same colouring in all the tiles
 
(i, j ) =
j + i(3k + 1) .
Theorem 3. For all = 2k +1, k = {1, 2, . . .}, the colouring
scheme described above is an optimal L(1 , 1 2 ) colouring
for G(A2 ), with i = k 2 . Moreover, this is a constant time
colouring scheme.
Proof. From lemma 1, c( ) is a lower bound. We can easily
see that each vertex in the tile is assigned a unique colour from
the set {0, 1, . . . , c( ) 1}. This implies that the optimality
condition is satisfied.
Again, the above scheme ensures that corresponding
points in neighbouring tiles have the same colour and are ex-
actly distance apart. Thus, the re-use constraint is satisfied.
To derive the value of 1 , we proceed as follows. Consider
a point (i, j ) in the grid. Its six neighbours are (i, j 1), (i +
1, j ), (i + 1, j + 1), (i, j + 1), (i 1, j ) and (i 1, j 1).
present in the grid.
Odd . Recall that the number of vertices c( ), in a tile
corresponding to an odd reuse distance = 2k + 1 is equal
to 2k 2 + 2k + 1 (see lemma 2). From the above fact and
from theorem 2, we make the following observations (refer to
figure 14):
Lemma 6.
1. Colouring c( ) points starting from the vertex of a tile
along the direction j is equivalent to colouring all the di-
agonals of a tile in the following order:
V0 , Vk , V2k , Vk1 , . . . , Vk+2 , V1 , Vk+1 .
2. Along a line i = m, where m is a constant, any pair of
points which are at a distance c( ) apart will have the same
colour assigned to them.
96 SHASHANKA ET AL.

(i 1, j ) ((i, j ) + k 2 + 2k + 1)
(i, j + 1) ((i, j ) + k 2 + k + 1)
(i + 1, j ) ((i, j ) + k 2 )
(i, j 1) ((i, j ) k 2 k 1)

Figure 15. Colours (mod c( )) assigned to the neighbours of (i, j ) by the

scheme  when is odd.

Figure 14. L(1 , 1 2 ) colouring for = 7.

3. Consider a point (p, q) on the line i = p. The point

(p + 1, q + 2k + 1) on the line i = p + 1 will have the
same colour as (p, q).
Figure 16. L(1 , 1 2 ) colouring for = 8.
From lemma 6, we see that a colouring for c( ) points
along a line i = m for some arbitrary m describes the colour- Again, the above scheme ensures that corresponding
ing for the entire grid. points in neighbouring tiles have the same colour and are ex-
The following colouring scheme is shown in figure 14. To actly distance apart. Thus, the re-use constraint is satisfied.
colour along the line i = 0, we proceed as follows: Starting To derive the value of 1 , we proceed as follows. Con-
with the point (0, 0) which is assigned the colour 0, we assign sider a point (i, j ) in the grid. Its four neighbours are
consecutive colours to every second vertex, and wrap around (i 1, j ), (i, j + 1), (i + 1, j ), and (i, j 1).
after the c( )th vertex. This will colour all the c( ) points in The colour assigned to (i, j ) according to the above
two passes uniquely. This can be easily seen because c( ) = scheme will be (i, j ) = 
(j i(2k + 1)).
c
(k) = 2k 2 + 2k + 1 is odd and hence, points coloured in the The table in figure 15 shows the colours (modulo c( ))
first pass will not be repeated again. Consecutive sets of c( ) assigned to the neighbours of (i, j ).
vertices along this line follow the same colouring pattern. From the table in figure 15, we see that the least difference
Mathematically, this colouring scheme can be expressed between the colours assigned to neighbouring points is k 2 .
as follows. Let (i, j ) represent the colour assigned to the Hence, 1 = k 2 .
vertex (i, j ) and 
(j ) represent the colour assigned to the From the formula for (i, j ), it can be easily seen that
vertex (0, j ), i.e. 
(j ) = (0, j ). We first give a formula given any arbitrary point (i, j ) in the grid, the colour assigned
for 
(j ) and then derive an expression for (i, j ). to (i, j ) can be computed in constant time. 



, j even, Even . We now present a colouring scheme for even ,
 (j ) =  4, i.e. = 2k, k {2, 3, . . .}. We first note that the total
+ k 2 + k + 1, j odd,
number of points in a tile in terms of k will be equal to 2k 2 .
Since colouring of the entire grid is achieved by colouring one
where j = j mod c( ) and = j/2. Now, from lem-
tile and reproducing the same colouring in all tiles of the grid,
ma 6.3, we can easily derive that
description of the colouring for a single tile is sufficient.
  The colouring scheme is shown in figure 16. Alternate di-
(i, j ) = 
j i(2k + 1) .
agonals are coloured consecutively starting with D0 , i.e. the
following diagonals D0 , D2 , . . . , D 2 , D1 , D3 , . . . , D 1
Theorem 4. For all = 2k +1, k = {1, 2, . . .}, the colouring are coloured in order. Starting with the origin of the tile which
scheme described above is an optimal L(1 , 1 2 ) colouring is assigned colour 0, points are coloured consecutively within
for G(Z2 ), with 1 = k 2 . Moreover, this is a constant time each diagonal.
colouring scheme. Let (i, j ) be the colour assigned to the point (i, j ) in the
grid. It can be mathematically expressed as follows:
Proof. From lemma 2, c( ) is a lower bound. We can easily    
(i j ) mod 2k (i + j ) mod 2k
see that each vertex in the tile is assigned a unique colour from (i, j ) = k+
the set {0, 1, . . . , c( ) 1}. This implies that the optimality 2 2
  2
condition is satisfied. + (i + j ) mod 2 k .
A CHARACTERISATION OF OPTIMAL CHANNEL ASSIGNMENTS
(i 1, j ) (i, j ) k 2
(i, j + 1) (i, j ) k 2 k + 1,
if (i + j ) mod 2k = 2k 1
(i, j ) k 2 + 1, otherwise
(i + 1, j ) (i, j ) k 2 + 1,
if (i + j ) mod 2k = 2k 1
(i, j ) k 2 + k + 1, otherwise
(i, j 1) (i, j ) 2k 2 + k,
if (i j ) mod 2k = 2k 1
(i, j ) k 2 + k, otherwise
Figure 17. Colours (mod c( )) assigned to the neighbours of (i, j ) by the
scheme  when is even.
Theorem 5. For all = 2k, k = {2, 3, . . .}, the colouring
scheme described above is an optimal L(1 , 1 2 ) colouring
for G(Z2 ), with 1 = k 2 k 1. Moreover, this is a constant
time colouring scheme.
Proof. From lemma 2, c( ) is a lower bound. We can easily
see that each vertex in the tile is assigned a unique colour from
the set {0, 1, . . . , c( ) 1}. This implies that the optimality
condition is satisfied.
Again, from the formula, we see that corresponding points
in neighbouring tiles have the same colour and are exactly
distance apart. Thus, the re-use constraint is satisfied.
To derive the value of 1 , we proceed as follows. Consider
the neighbours of an arbitrary point (i, j ) in the grid. They are
(i 1, j ), (i, j + 1), (i + 1, j ) and (i, j 1). We will find the
differences between the colours assigned to (i, j ) and each of
its neighbours. The least difference will be equal to 1 .
There are two cases to consider: (1) (i + j ) is even, and
(2) (i + j ) is odd. Note that (i + j ) value for alternate points
in both X and Y directions will be of the same parity. If we
consider a point for which (i + j ) is odd, (i + j ) for all its
neighbours will be even and vice versa. It follows that we
need to consider only one case, as considering the other case
will yield the same expressions for the differences.
Consider a point (i, j ) and suppose (i + j ) is odd. Let the
colour assigned to (i, j ) be (i, j ). The table in figure 17
shows the colours assigned to the neighbours of (i, j ).
Clearly, from the table in figure 17, the least difference
between the colours assigned to neighbouring points is k 2
k 1. Hence, 1 = k 2 k 1.
From the formula for (i, j ), it can be easily seen that
given any arbitrary point (i, j ) in the grid, the colour assigned
to (i, j ) can be computed in constant time.
5. Upper bound on 1
The previous subsections presented colouring schemes for
odd reuse distances where 1 , the channel separation con-
straint, has a value of k 2 , where = 2k + 1. Lemma 7
provides an upper bound on 1 .
Lemma 7. For all = 2k + 1, k = {1, 2, . . .}, for any op-
timal L(1 , 1 2 colouring for G(A2 ) (G(Z2 ), respectively)
1 < k 2 + k.
97
Proof. Suppose C is an optimal L(1 , 1 2 ) colouring for
G(A2 ). Since there are 3k 2 + 3k + 1 vertices in a tile of
G(A2 ), C assigns each number in {0, . . . , 32 + 3k} to some
vertex in G(A2 ). Let 1 = be the separation between the
colours assigned by C to any two adjacent vertices.
Consider the vertex v assigned the colour 1. Each of its
six neighbours must be assigned a colour that is at least 21.
Since (1) the neighbours of v form a cycle of length 6, and (2)
each adjacent pair of vertices in the cycle must be assigned
colours differing by at least , it follows that at least three of
these vertices must each be assigned a colour that is at least
3 1. Thus, at least one neighbour of v must be assigned a
colour that is at least 3 + 1. Then,
3 + 1
3k 2 + 3k < k 2 + k.
A similar argument can be used to show that for an optimal
L(1 , 1 2 ) colouring of G(Z2 ), < k 2 + k. 
If = 2k is even, the size of the tile in G(A2 ) and in G(Z 2 )
is 3k 2 and 2k 2 , respectively. Then, an argument similar to the
one in the proof of lemma 7 can be used to show that
Lemma 8. For all = 2k, k = {1, 2, . . .}, for any opti-
mal L(1 , 1 2 ) colouring for G(A2 ) (G(Z2 ), respectively)
1 < k 2 .
Based on experimental verification by means of an exhaus-
tive search for all values of k
4, we conjecture that:
Conjecture 1. For all , for any optimal L(1 , 1 2 )-col-
ouring for G(A2 ) (G(Z2 ), respectively) 1
(/2)2.
The conjecture implies a tighter upper bound, for odd ,
than the one presented in lemma 7 above. Note that, for
odd , our assignments presented in sections 4.1 and 4.2 do
realise this value for 1 .
6. Conclusions and open problems
We characterised optimal channel assignment schemes for
cellular and square grids, and hence showed that any such
scheme must be uniform across the entire grid. More specif-
ically, in an optimal colouring, the colouring of a tile (for a
 given ) will be identically repeated in all the tiles through-
out the grid. We also presented optimal L(1 , 1 2 ) colour-
ing schemes, with a high value for 1 , for square grids for all
 4 and for cellular grids for the case where reuse distance
is odd, i.e., = 2k + 1, k {1, 2, . . .}. The previous best
known results have been restricted to 1  3k 2 /8 [6], in case
of cellular grids and 1
( 1)/2 [3] in case of square
grids. We prove an upper bound for 1 for optimal colourings
of cellular and square grids. In the case of being odd, we
conjecture that our value of 1 is a tight upper bound on 1 for
optimal colouring schemes for these grids.
Several interesting open questions arise from the work pre-
sented here. We list a few of them here:
98 SHASHANKA ET AL.

(1) Find optimal colouring schemes for cellular grids with
high 1 values for the case when is even.
(2) Find and prove the existence of tight upper bounds for
1 , 1 2, . . . for a general .
M.V.S. Shashanka obtained a B.E. (Hons.) from the
Birla Institute of Technology & Science, Pilani, In-
dia. He is currently a graduate student in the De-
partment of Cognitive and Neural Systems at Boston
University, Boston, MA, USA.
E-mail: mvss@bu.edu

References

[1] R. Battiti, A.A. Bertossi and M.A. Bonuccelli, Assigning codes in wire-
less networks: Bounds and scaling properties, Wireless Networks 5
(1999) 195209.
Amrita Pati obtained a B.E. (Hons.) from the Birla
[2] A.A. Bertossi, C.M. Pinotti and R.B. Tan, Efficient use of radio spec-
Institute of Technology & Science, Pilani, India. She
trum in wireless networks with channel separation between close sta-
is currently a graduate student in the Department of
tions, in: Proc. of the DIAL M Workshop (2000) pp. 1827.
Computer Science at the Virgina Polytechnic Insti-
[3] A.A. Bertossi, C.M. Pinotti and R.B. Tan, Channel assignment with
tute and State University, Blacksburg, VA, USA.
separation for special classes of wireless networks: Grids and rings, in:
E-mail: apati@vt.edu
Proc. of IPDPS (2002).
[4] H.L. Boedlander, T. Kloks, R.B. Tan and J. van Leeuwen, -coloring
of graphs, in: Proc. of STAGS (2000) pp. 395406.
[5] J. Conway and N. Sloane, Sphere Packings, Lattices and Groups, 2nd
ed. (Springer, Berlin, 1993).
[6] A. Dubhashi, A. Pati, M.V.S. Shashanka, R. Shashank and
A.M. Shende, Channel assignment in wireless networks modelled as
cellular and d-dimensional square grids, http://citeseer.nj.
nec.com/ (2002). Anil M. Shende is an Associate Professor of Com-
[7] J.R. Griggs and R.K. Yeh, Labeling graphs with a condition at dis- puter Science at Roanoke College, Salem, VA, USA.
tance 2, SIAM J. Discrete Math. (1992) 586595. Part of this work was done while he was at the Birla
[8] W.K. Hale, Frequency assignment: Theory and application, Proc. IEEE Institute of Technology & Science, Pilani, India on
68 (1980) 14971514. sabbatical leave.
[9] R.A. Murphey, P.M. Pardalos and M.G.C. Resende, Frequency assign- E-mail: shende@roanoke.edu
ment problems, in: Handbook of Combinatorial Optimization (Kluwer
Academic, Dordrecht, 1999).
[10] D.S. Rajan and A.M. Shende, A characterization of root lattices, Dis-
crete Math. 161 (1996) 309314.
 Mobile Networks and Applications 10, 99113, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

CARD: A Contact-based Architecture for Resource Discovery in

Wireless Ad Hoc Networks
AHMED HELMY , SAURABH GARG and NITIN NAHATA
3740 McClintock Avenue, EEB 232, Electrical Engineering Department, University of Southern California, Los Angeles, CA 90089-2562, USA

PRIYATHAM PAMU
Computer Science Department, University of Southern California, Los Angeles, CA 90089-2562, USA

Abstract. Traditional protocols for routing in ad hoc networks attempt to obtain optimal or shortest paths, and in doing so may incur
significant route discovery overhead. Such approaches may be appropriate for routing long-lived transfers where the initial cost of route
discovery may be amortized over the life of the connection. For short-lived connections, however, such as resource discovery and small
transfers, traditional shortest path approaches may be quite inefficient. In this paper we propose a novel architecture, CARD, for resource
discovery in large-scale wireless ad hoc networks. Our mechanism is suitable for resource discovery as well as routing very small data
transfers or transactions in which the cost of data transfer is much smaller than the cost of route discovery. Our architecture avoids expensive
mechanisms such as global flooding and complex hierarchy formation and does not require any location information. In CARD resources
within the vicinity of a node, up to a limited number of hops, are discovered using a proactive scheme. For resources beyond the vicinity,
each node maintains a few distant nodes called contacts. Contacts help in creating a small world in the network and provide an efficient way
to query for distant resources. Using contacts, the network view (or reachability) of the nodes increases, reducing the discovery overhead and
increasing the success rate. On the other hand, increasing the number of contacts also increases control overhead. We study such trade-off
in depth and present mechanisms for contact selection and maintenance that attempt to increase reachability with reduced overhead. Our
schemes adapt gracefully to network dynamics and mobility using soft-state periodic mechanisms to validate and recover paths to contacts.
Our simulation results show that CARD is scalable and can be configured to provide desirable performance for various network sizes.
Comparisons with other schemes show overhead savings reaching over 93% (vs. flooding) and 80% (vs. bordercasting or zone routing) for
high query rates in large-scale networks.
Keywords: energy efficient, sensor networks, routing

1. Introduction eral attempt to discover optimal (shortest path) routes. In

Ad hoc networks are wireless networks composed of mobile
devices with limited power and transmission range. These
networks are rapidly deployable as they neither require a
wired infrastructure nor centralized control. Because of the
lack of fixed infrastructure, each node also acts as a relay
to provide communication throughout the network. Appli-
cations of ad hoc networks include coordination between
various units (e.g., in a battlefield), search and rescue mis-
sions, rapidly deployable networks, and vehicular networks,
among others. Although research on mobile ad hoc networks
(MANets) has attracted a lot of attention lately, little attention
has been given to resource discovery in large-scale MANets.
In addition, a very important mode of communication that
has been largely ignored in the ad hoc networks literature is
that of short flows and small transactions, where the com-
munication cost of discovering shortest routes is usually the
dominant factor (not the data transfer as in long flows). For
such short flows reducing overhead (not route optimization)
is the main design goal. Current routing protocols in gen-
A. Helmy was supported by NSF CAREER Award 0134650, and research
grants from Intel Corp. and Pratt & Whitney Institute for Collaborative
Engineering.
our study, instead of obtaining shortest paths, we focus on
reducing the overhead of resource (or route) discovery for
short flows. Examples of resource discovery and small trans-
fers in ad hoc networks include discovering servers, objects
and capabilities (e.g., GPS capable nodes), instant and text
messaging, short transactions, DNS-like queries, paging, and
dissemination of sensory data in sensor and vehicular net-
works.
In ad hoc networks, lack of infrastructure renders resource
discovery a challenging problem. In addition, mobility in-
duces frequent route changes. Traditional protocols proposed
for resource discovery employ either global flooding or com-
plex hierarchy formation schemes. While flooding is ineffi-
cient and does not scale well, hierarchy formation involves
complex coordination between nodes and therefore may suf-
fer significant performance degradation due to frequent, mo-
bility induced, changes in network connectivity.
To overcome these limitations we propose a new architec-
ture for efficient resource discovery in large-scale ad hoc net-
works, called CARD. Our study targets resource discovery
and routing for short flows. CARD is not a general routing
protocol, as we make a design decision to trade-off shortest
paths for drastic reduction in discovery overhead. CARD,
100
however, may be integrated easily with zone routing proto-
cols to compose a general routing solution.
Nodes in ad hoc networks are usually portable devices with
limited battery power. Therefore to save power the resource
discovery mechanism should be efficient in terms of commu-
nication overhead. Our architecture is designed to meet re-
quirements for power-efficient resource discovery and small
transfers in large-scale ad hoc networks with (potentially)
thousands of wireless devices. Scalability is one of our main
design goals.
Our architecture is based on the concept of small worlds
[8,26,27] where the addition of a small number of short cuts
in highly clustered networks results in significant reduction
in the average path length (or degrees of separation) to ap-
proach that of random networks. In our architecture we adopt
a hybrid approach in which a node uses periodic updates to
reach its vicinity within a limited number of hops, R, and re-
active querying beyond the vicinity via contacts. Contacts
act as short cuts that attempt to transform the network into a
small world by reducing the degrees of separation between the
source and destination of the transfer. They help in providing
a view of the network beyond the vicinity during resource dis-
covery. Each node maintains state for a few contacts beyond
its vicinity. Contacts are polled periodically to validate their
presence and routes. For discovering resources efficiently,
queries are sent to the contacts that leverage the knowledge
of their vicinity. As the number of contacts increases, the net-
work view (or reachability) increases. However, at the same
time the overhead involved in contact selection and mainte-
nance also increases. Our results show this trade-off. We
introduce and study alternative mechanisms for contact selec-
tion and identify a novel scheme (called the edge method for
contact selection) that is able to achieve a balanced trade-off
and good performance in terms of increased reachability and
reduced overhead.
Once the contacts are selected by a node they are used in
the resolution of resource discovery queries. Only the contact
nodes are queried without the need for flooding, resulting in
drastic reduction in per-query communication overhead. The
total overhead, however, is the resultant of
(i) the query overhead, which is a function of the per-query
overhead and the query rate,
(ii) the vicinity establishment and maintenance overhead,
which is a function of the node mobility, and
(iii) the contact selection and maintenance overhead.
Our study elaborates on the interplay between these var-
ious overhead components, the query rate, and the mobility
rate using the call-to-mobility-ratio (CMR) metric.
Extensive simulation-based comparisons with flooding
and bordercasting [5,20] show our architecture to be more ef-
ficient, especially for high query rates. Simulation results also
show that our protocol is scalable and can be configured to
provide good performance for various network sizes. Over-
head savings are function of the query rate, reaching 93%
(vs. flooding) and 80% (vs. bordercasting) in communication
HELMY ET AL.
savings for high query rates in large-scale networks; a drastic
improvement in performance.
The rest of this document is organized as follows. Sec-
tion 2 discusses related work. Section 3 describes our design
goals and provides an overview of our architecture, CARD,
and introduces the contact selection, maintenance and query
algorithms. Section 4 presents analysis of CARD, and com-
pares it to flooding, smart flooding and bordercasting. We
conclude in section 5.
2. Related work
Related research lies in the areas of routing and resource dis-
covery in ad hoc networks. Due to lack of infrastructure in ad
hoc networks, resource (and route) discovery is a challeng-
ing problem. Most of the routing protocols can be broadly
classified as: proactive (table-driven), reactive (on-demand),
hybrid, or hierarchical.
Proactive schemes such as DSDV [21], WRP [18] and
GSR [2] flood periodic updates throughout the network. This
is resource consuming, especially for large-scale networks.
Reactive schemes such as AODV [22] and DSR [13] attempt
to reduce the overhead due to periodic updates by maintaining
state only for the active resources and using route caching. In
these schemes a search is initiated for new discovery requests.
However, the search procedure generally involves flooding (or
expanding ring search), which also incurs significant over-
head. Furthermore, the performance of on-demand routing
with caching has been shown to degrade significantly with
small transfers in large-scale mobile networks.
Hybrid schemes such as the zone routing protocol (ZRP)
[5,20] try to combine the benefits of both the proactive and re-
active schemes. ZRP limits the overhead of periodic updates
to a limited number of hops (called the zone radius). Re-
sources beyond the zone are discovered in a reactive manner
by sending queries through nodes at the edges of the zones
(bordercasting). The zone concept is similar to the vicin-
ity concept in our study. However, instead of bordercasting
we use contact queries. The design principles upon which
our CARD architecture was designed employing contacts
as short cuts to create a small world, and trading off optimal
paths for energy efficiency are fundamentally different from
those used for ZRP bordercast. In our study, through detailed
comparison we show that the contact-based approach is much
more efficient than bordercasting for our purposes. Further-
more, CARD maybe easily integrated with ZRP to provide a
complete routing protocol in which ZRP is used to discover
routes for long-lived flows and CARD is used for resource
discovery and small transfers.
Hierarchical schemes, such as CGSR [3,15], tend to have
good scalability, but involve election of cluster-heads, that
have greater responsibilities than other nodes. A cluster-head
is responsible for routing traffic in and out of the cluster.
Cluster-based hierarchies rely on complex coordination and
thus are susceptible to major re-configuration due to mobility
and node failure, leading to serious performance degradation
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY
in highly dynamic networks. Also, a cluster head may be a
single point of failure and a potential bottleneck. In our ar-
chitecture each node has its own view of the network, and
hence there is very little coordination between various nodes.
This enables our architecture to adapt gracefully to network
dynamics. GLS [14] provides a location-discovery service for
geographic routing. GLS requires nodes to know of a network
grid map and assumes knowledge of node locations (via GPS
or other). CARD does not require location information.
Related work on smart or efficient flooding has been pro-
posed in [4,6,16,19]. These techniques attempt to reduce the
redundancy inherent in flooding, and may be integrated in our
work to provide more efficient vicinity establishment instead
of regular link state protocol. One major difference between
smart flooding and CARD is that smart flooding reduces the
redundant messages in querying every node in the network,
whereas CARD attempts to create a small world and only
queries a small number of nodes on the order of the degrees of
separation from source to target. In relatively sparse networks
(some of which we include in our study) smart flooding will
not be very effective since there is no significant redundancy
in flooding anyway. Section 4.3 discusses this issue further.
In [8] we have shown the relationship between small
worlds and wireless networks. In this paper, we build
upon that relationship by introducing the contacts to act as
short cuts in the highly clustered multi-hop wireless network,
proposing and evaluating in details two proactive contact-
selection mechanisms. We first introduced the high level idea
of using contacts in [7]. The initial work on the CARD archi-
tecture was presented in [11]. This work extends the analysis
of the CARD architecture and explores the important inter-
play between the query rate and mobility rate. The MARQ
architecture [9] provides a mobility-assisted contact selec-
tion mechanism, the efficiency of which increases with mo-
bility. In cases of static networks (e.g., sensor networks), or
when mobility is low, CARD may be used in conjunction with
MARQ for efficient query resolution. TRANSFER [10] pro-
vides a reactive (on-the-fly) contact selection mechanism to
reduce node-contact vicinity overlaps, but does not explicitly
reduce the contact-contact vicinity overlap because contacts
are selected in parallel. CARD, by virtue of selecting contacts
proactively and using the edge method for contact selection
in serial is able to guarantee non-overlapping node-contact
vicinities and reduce the contact-contact vicinity overlap, but
may incur more overhead for periodic contact maintenance.
ACQUIRE [23,24] is an architecture for multi-variable query
resolution in sensor networks, that uses the look-ahead tech-
nique to optimize overhead. The query is forwarded from one
querying node to another d hops away, randomly. The work
provides an analytical framework to get optimal d for given
level of network and event dynamics. A variant of CARDs
contact selection may be used to reduce the overlap between
the look ahead zones for successive querying nodes to im-
prove the performance of ACQUIRE.
101
3. CARD architectural overview
In this section we provide an overview of the CARD archi-
tecture. In particular, we describe the design requirements for
our architecture, present definitions and terminology used in
this document, and introduce and investigate alternative con-
tact selection, maintenance and query mechanisms.
3.1. Design requirements
The design requirements of our CARD resource discovery ar-
chitecture for large-scale Ad hoc networks include (a) scala-
bility, (b) power-efficiency, (c) robustness, (d) decentralized
self-organization, and (e) independence of location informa-
tion.
(a) Scalability. Applications of large-scale ad hoc networks
include military and sensor network environments that may
include thousands of nodes. Therefore the resource discovery
mechanism should be scalable in terms of control overhead
with increase in network size. We shall show that CARD may
be configured to perform very well over a wide array of net-
work sizes and conditions.
(b) Power and communication efficiency. Ad hoc networks
include portable devices with limited battery power. There-
fore, resource discovery mechanisms should be power-effi-
cient. CARD achieves dramatic reduction in communica-
tion overhead (in terms of transmitted and received messages)
over the several existing schemes considered in our study.
(c) Robustness. The mechanism should be robust in the face
of network dynamics. A periodic soft-state mechanism is pro-
vided to handle node failures and frequent link failures due to
mobility.
(d) Decentralized operation. For the network to be rapidly
deployable, it should not require any centralized control.
CARD does not require or assume any centralized entity or
special infrastructure.
(e) Independence of location information. GPS (or other lo-
cation information) may not be available in many context
(e.g., indoors, or in simple devices and sensors). Hence, as-
suming availability of location information limits the applica-
bility of the proposed scheme. We avoid such limitation in our
design and do not assume or require any location information.
3.2. Definitions
An overview of the CARD architecture is shown in figure 1.
Following are some terminology definitions we use through-
out this document.
Vicinity (of a node). All nodes within a particular number
of hops (R) from the node. R is the radius of the vicinity.
Edge nodes (of a nodes vicinity). All nodes at a distance
of exactly R hops away from the node.
Maximum contact distance (r). The maximum distance (in
hops) from the source within which a contact is selected.
102
Figure 1. Architectural overview for CARD: Node S (potentially any source) keeps track of nodes and resources in its vicinity, up to R hops away. S also
elects and maintains routes to a small number of contacts (NoC) (in this case NoC = 3 contacts: C1, C2, and C3). Contacts are selected within r hops away
from S. Nodes exactly R hops away from S are called the edge nodes (Ei).
Overlap. Overlap between nodes represents number of
common nodes between their vicinities.
Number of Contacts (NoC). NoC specifies the value of
the maximum number of contacts to be selected by each
source node. The actual number of contacts chosen is usu-
ally less than this value. This is due to the fact that for
a particular value of R and r, there is only a limited re-
gion available for choosing contacts. Once this region has
been covered by vicinities of the chosen contacts, choos-
ing more contacts in the same region is not possible, as
their vicinities would overlap with the vicinities of the al-
ready chosen contacts. This is according to our contact
selection policy to minimize overlap.
Depth of search (D). D specifies the levels of contacts
(i.e., contacts of contacts) queried by a source.
Reachability. The reachability of a source node refers to
the number of nodes that can be reached by the source
node. This includes the nodes within the vicinity that can
be reached directly and the nodes that lie in the contacts
vicinities, and their contacts vicinities, and so on, up to D
levels of contacts. This is also considered a measure of the
discovery success rate.
3.3. Establishing and maintaining vicinity information
Our architecture employs a hybrid of proactive and reactive
approaches for resource discovery. As shown in figure 1, all
nodes within R hops from a node form the nodes vicinity.
HELMY ET AL.
Each node proactively (e.g., using a link state protocol) main-
tains state for resources within its vicinity. Alternatively, a
smart flooding scheme (e.g., based on dominating sets) may
be used to reduce the vicinity establishment and maintenance
overhead. For comparison reasons, however, in this study
we shall use a link state protocol similar to that used in ZRP
to maintain vicinity information. The overhead of such link
state protocol increases with node mobility and the number of
nodes within the vicinity. Such overhead under mobility sce-
narios will be thoroughly studied later in section 4.3, and will
be factored into the overall overhead of CARD. As we shall
see, when the vicinity overhead is amortized over a reasonable
number of queries the overall gain is still quite significant.
Each node also maintains state for (a few) nodes that lie
outside the vicinity. These nodes serve as contacts for ac-
cessing resources beyond the vicinity. Contacts are selected,
maintained and queried using the mechanisms described be-
low.
3.4. Contact selection, maintenance and query mechanisms
Contacts are key to the efficient resolution of resource discov-
ery queries. In CARD, the contacts are selected proactively in
anticipation of queries, and paths to these contacts are main-
tained using a periodic soft state mechanism to capture net-
work dynamics and mobility effects. Since the contacts are
selected proactively, the contact selection delays become less
of a concern (than they would otherwise in a reactive, on-the-
fly, scheme, for example). Hence, contacts are selected in
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY
a serial fashion, one after the other, with information about
previously selected contacts being utilized to effectively se-
lect new contacts. When a resource discovery query is issued,
and the resource is not found in the vicinity, the source node
queries its contacts first, and the contacts may query their con-
tacts, and so on, until the query is resolved. Below, we in-
troduce the details of the contact selection, maintenance and
query mechanisms.
3.4.1. Contact selection mechanism
Any potential source of query or small transfer may choose
to select contacts. The procedure starts when a node s sends
a Contact Selection (CS) message through each of its edge
nodes (Ei), one at a time, until NoC number of contacts are
selected or until all edge nodes have been attempted. An edge
node receiving a CS forwards it to a randomly chosen neigh-
bor (X).
A node receiving a CS decides whether or not to be a con-
tact for s based on a contact selection method. This deci-
sion is made using either a probabilistic method (PM) or edge
method (EM). These methods are described later in this sec-
tion. After using either procedure PM or EM for deciding
whether (or not) to be a contact, if the node receiving a CS
does not choose to be the contact, it forwards the CS to one of
its randomly chosen neighbor (excluding the one from which
the CS was received).
The CS traverses in a depth-first manner until a contact
is chosen or the distance traversed by the CS from s reaches
r hops. If a contact is still not chosen (due to overlap), CS
backtracks to the previous node, which forwards it to another
randomly chosen neighbor. When a contact is selected, the
path to the contact is returned and stored at s.
3.4.2. Contact selection methods
We introduce and compare two different methods for contact
selection: (a) the probabilistic method (PM), and (b) the edge
method (EM).
(a) Probabilistic Method (PM). Contacts increase a nodes
view (reachability) of the network beyond its own vicinity.
To increase the reachability of a node, the vicinities of that
node, call it s, and its contacts should be disjoint, i.e., there
should be reduced (or no) overlap between the vicinity of s
and the vicinity of any of its contacts. The vicinities of differ-
ent contacts of the same node should also be non-overlapping,
to achieve good increase in reachability. To achieve this, the
CS contains the following information: (i) ID of node s, (ii) a
list of already-selected-contacts of s (Contact_List; typically
small of 5 IDs), and (iii) the hop count d.
This information is used as follows. When a node X
receives a CS, it first checks if s lies within its vicinity.
This check is easily performed since each node has complete
knowledge of its vicinity. So a node knows the IDs of all the
other nodes in its vicinity. X also checks if its vicinity con-
tains any of the node IDs contained in the Contact_List.
If neither s nor any of its already-selected-contacts lie in
the vicinity of X, then X probabilistically chooses itself as
103
(a) (b)
Figure 2. Overlap in (a) due to the use of P . (a) Heavy overlap; (b) no
overlap.
the contact. This probability (P ) of choosing to be a contact
is defined as follows:
d R
P = , (1)
rR
where d is the number of hops traversed from s to X. The
value of d is included in the CS as hop count. From the above
equation, when d = R, P = 0, and when d = r, P = 1. This
aims to select contacts between R and r hops away from s,
and is formulated to provide an increase in reachablility with
the addition of new contacts outside the vicinity of s, i.e., with
distance >R hops from s. The probability P increases with
the number of hops traversed, d. However, there are cases
where equation (1) does not provide the maximum benefit of
adding a contact. An example case is shown in figure 2(a)
where c is the contact for node s and the contact route (route 1
in the figure) is R + 2 hops.
In this figure although the distance between s and its con-
tact c is greater than R hops, there is still heavy overlap be-
tween the two vicinities. Such situations will arise whenever a
node within R hops from the edge node becomes the contact.
To alleviate this effect, equation (1) is modified to:
d 2R
P = . (2)
r 2R
In this equation P = 0 when d = 2R and P = 1 when
d = r. Hence, contacts are chosen after traversing between
2R and r hops from the source s.
Figure 3 explains the contact selection procedure with an
example. In the figure R = 3 and r = 6. Nodes a, b, c
and d are the edge nodes for node s. Node s sends a Con-
tact Selection (CS) message through its edge node a. Node
a randomly chooses one of its neighbors, e, and forwards the
CS to that node. Node e calculates the probability P , say ac-
cording to equation (1). If the probability of being the contact
fails at e, it forwards the CS to one of its neighbors f (chosen
randomly). Node f again forwards the CS to g. As g is at
r hops from s, the probability P at g is 1. However, g still
104 HELMY ET AL.

Figure 4. Reachability for (1) PM and (2) EM. 1

Figure 3. Selecting contacts.

cannot become a contact for s as there already exists another

contact h (which was selected through a previous selection
via another edge node d) in the vicinity of g. So g returns the
CS to f (backtracking). Node f then forwards CS to another
neighbor, and so on.

(b) Edge Method (EM). Even with equation (2) the proba-
bilistic method can result in a situation where there is some
overlap between the vicinity of the contact and the vicinity
of s. This is possible due to the fact that the nodes do not
have a sense of direction once the CS message is forwarded
out of the vicinity (i.e., d > R). Therefore, it is possible
that a contact may be selected at a location where the CS has
traversed more than 2R hops, but the contact may in fact be
closer than 2R hops from the source, as shown in figure 2(a)
route 2, leading to heavy overlap.
More seriously, the probabilistic method for contact selec-
tion can be expensive in terms of the amount of traffic gener-
ated by the CS. This is due to the extra traffic generated due
to backtracking, and lost opportunities when the probability
fails, even when there is no overlap. To reduce the possibil-
ity of such a situation, the probability equations (1) and (2)
are not used. The probability equations were formulated to
have a higher possibility of choosing the contact that lies ei-
ther between R and r hops (equation (1)) or between 2R and
r hops (equation (2)). To maintain this non-overlapping prop-
erty without the probability equations, the contact selection
procedure is modified as follows.
The list of all edge nodes (Edge_List) of s is added to
the CS. Also, the query and source IDs are included to pre-
vent looping. Note that the Edge_List is readily available
through the vicinity information and obtaining it does not re-
quire any extra overhead. Upon receiving a CS, in addition
to checking for overlap with ss vicinity and the vicinities of
all the already-selected-contacts (Contact_List), the receiving
node also checks for overlap with the vicinities of any of the
nodes on the Edge_List as well. It can be easily proven that
Figure 5. Overhead for (1) PM and (2) EM.1
this scheme guarantees non-overlap between the node and the
contacts vicinities, as follows. Any node that lies at a dis-
tance of R hops or less from the edge will have an overlap-
ping vicinity with the ss vicinity, and hence will have at least
one of ss edge nodes in its vicinity. Thus, checking for non-
overlap with the edge nodes ensures that a contact is chosen at
least 2R +1 hops away from s. This eliminates the possibility
of an overlap due to the lack of direction. The Edge_List may
be added to the CS in a communication-efficient manner by
using bloom filters [17] to represent membership in the edge-
list. Figures 4 and 5 show a comparison of the probabilistic
and edge methods. As can be seen from figure 4 the reach-
ability saturates in both PM and EM. However the saturation
occurs much earlier in the case of probabilistic method. Also
as compared to EM, the reachability achieved is less for PM,
for the same values of NoC. Figure 5 shows the backtrack-
ing overhead for PM and EM. Due to the reasons explained
earlier, overhead is significantly reduced for EM.
3.4.3. Contact maintenance mechanism
Node mobility may cause the path to a contact to change.
Therefore a node needs to keep track of its contacts and their
paths. This is done using soft-state periodic polling of the
contacts as follows.
1 Shown: 500 nodes, 710 m 710 m, Tx range = 50 m, R = 3, r = 20,
D = 1. Similar trends were obtained for other simulation scenarios.
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY
(1) Each node periodically sends a validation message to-
wards each of its contacts. These validation messages
contain the path from a node s to the contact.
(2) Each node on the path that receives the validation mes-
sage checks if the next hop in the path is a directly con-
nected neighbor. If so, it forwards the validation message
to the next hop node. If the next hop is missing, the node
tries to salvage the path using local recovery, discussed
later in this subsection.
(3) If a path cannot be salvaged using local recovery, the con-
tact is considered to be lost.
(4) If the path to a contact is validated but the number of hops
to the contact does not lie between 2R and r, the contact
is considered to be lost.
(5) After validating all the contacts, if the number of contacts
left is less than the specified NoC, then a new contact se-
lection procedure is initiated.
The local recovery mechanism is illustrated using an ex-
ample of a contact path (a b c d e). Assuming
reasonable values of node velocities and validation frequency
(section 1 in our study), there is a high probability that if a
node (say c) has moved out of a contact path (i.e., moved out
of transmission range of b), that it is still within the vicinity
of the previous hop (b) in the path. Even in the case when
the moving node (c) is completely lost (because it has moved
out of the vicinity of the previous hop, b), some other node
further down the path (say dor e) might have moved into the
vicinity of the previous node (b). Local recovery takes advan-
tage of these cases to recover from changes in the path when
Figure 6. Contact maintenance using local recovery: (A) Path to the contact node e goes through a b c d e. Node c is moving away from bs
transmission range, and node d is moving away from e. (B) During validation, node b loses contact with node c but finds node d in its range. Also, node d
loses direct contact with e but finds a path in its vicinity to node e through node f . The updated part of the contact path is thus a b d f e.
105
possible, without having to initiate new searches from s. Thus
local recovery provides an efficient mechanism for validating
contacts and recovering from changes in the contact paths. If
the next hop on the path (node c) is missing, the node that
received the validation message (node b) looks for the next
hops (c, d and e) in its vicinity routing table. If any of the
next hops (c, d or e) is found the vicinity, the path is updated
and the validation message is forwarded to that next hop. If
the lookups for all next hops fail, an error message is returned
to the source s, and another contact selection is initiated. Fig-
ure 6 further illustrates an example of local recovery when
two nodes along the path to the contact (nodes c and d in this
case) move.
3.4.4. Query mechanism
When a source node s (potentially any node), needs to reach
a destination or target resource T it first checks its vicinity
table to see if T exists in its own vicinity. If T is not found
in the vicinity, s sends a Destination Search Query (DSQ) to
its contacts. The DSQ contains the following information:
(1) depth of search (D), and (2) target resource ID (T ). Upon
receiving a DSQ, each contact checks the value of D. If D
is equal to 1, the contact performs a lookup for T in its own
vicinity. If T exists, then the path to T is returned to s, and
the query is considered successful. Otherwise, if D > 1, the
contact receiving the DSQ decrements D by 1 and forwards
the DSQ to its contacts. In this way the DSQ travels through
multiple levels of contacts until D reduces to 1.
The source node s first sends a DSQ with D = 1 to its
contacts. So only the first level contacts are queried with this
DSQ. After querying all its contacts if the source does not
106
Table 1
Description of the various scenarios used for simulating CARD.
Scenario Nodes Area Transmission range No. of links
1 250 500 500 50 837
2 250 710 710 50 632
3 250 1000 1000 50 284
4 500 710 710 30 702
5 500 710 710 50 1854
6 500 710 710 70 3564
7 1000 710 710 50 8019
8 1000 1000 1000 50 4062
receive a path to the target within a specified time, it creates
a new DSQ with D = 2 and sends it again to its contacts.
Each contact observes that D = 2 and recognizes that this
query is not meant for itself. So it reduces the value of D in
the DSQ by 1 and forwards it to its contacts. These contacts
serve as second level contacts for the source. Upon receiving
the DSQ, a second level contact observes that D = 1 and it
does a lookup for the target T in its own vicinity and returns
the path to T , if found. In this way the value of D is used
to query multiple levels of contacts in a manner similar to the
expanding ring search. However, querying in CARD is much
more efficient than the expanding ring search as the queries
are not flooded with different TTLs but are directed to indi-
viual nodes (the contacts). Contacts leverage knowledge of
their vicinity (gained through the proactive scheme operating
within the vicinity) to provide an efficient querying mecha-
nism.
4. Evaluation and analysis
In this section we present detailed simulation based evalua-
tion and analysis of our architecture. NS-2 [1] along with our
CARD extensions and other utilities were used to generate
various scenarios of ad hoc networks. The mobility model
used for these simulations was the random way-point model.
Our simulations so far did not consider MAC-layer issues. In
random way point model a node is assigned a random velocity
from [0, Vmax ] and assigned a destination location randomly.
Once the node reaches its destination it is assigned a random
velocity and random destination again, so on. In the reacha-
bility analysis experiments the mobility was set to 0 to un-
derstand the basic effects of the various architectural parame-
ters on reachability characteristics. For the maintenance over-
head, total overhead and comparison experiments, continuous
mobility was used (with no pauses) with Vmax = 20 m/s.
First we try to understand the effect of various parameters
such as vicinity radius (R), maximum contact distance (r),
the number of contacts (NoC), the depth of search (D) and
network size (N) on reachability and overhead. Reachability
here is defined as the percentage of nodes that are reachable
from a source node. For overhead we consider the number
of control messages; the contact selection (CS) messages and
the periodic contact maintenance validation messages. Hav-
ing developed an understanding of the various parameters in
our architecture, we then compare it to other schemes such
HELMY ET AL.
Aver. node degree Network diameter Aver. path length (hops)
6.75 23 9.378
5.223 25 9.614
2.57 13 3.76
4.32 20 5.8744
7.416 29 11.641
14.184 17 7.06
16.038 24 8.75
8.156 37 14.33
as flooding and bordercasting in terms of query overhead and
query success rate.
Table 1 shows the scenarios used in our simulations. These
scenarios vary in number of nodes, network size, node den-
sity and transmission range. The variation is considered to
capture the effect of these factors on CARD. As was shown
in figures 4 and 5, the edge method outperforms the proba-
bilistic method. Therefore, we use the edge method (EM) for
contact selection in the rest of our study.
4.1. Analysis of reachability
Analysis of the reachability, or query success rate, was con-
ducted to understand how contacts help in increasing the view
of the network. Here we present results for a topology of 500
nodes spread over area of 710 m 710 m. The details can
be seen from table 1, scenario number 5. Similar trends were
observed for other scenarios.
4.1.1. Varying vicinity radius (R)
Figure 7 shows the effect of increasing the vicinity radius (R)
on reachability. As R increases, the reachability increases
and the reachability distribution in figure 7(a) shifts to the
right; i.e., more nodes achieve higher percentage of reacha-
bility. This increase in reachability with the increase in R is
due to increase in the number of nodes within the vicinity.
As the value 2R approaches the maximum contact distance r
(r = 16 in this experiment), the region available for contact
selection (between 2R and r) is reduced. This results in less
number of contacts being chosen. In figure 5, when R = 7,
contacts can only be selected between 2R = 14 and r = 16
hops from the source. This small region for contact selec-
tion significantly reduces the number of contact and hence
the reachability reduces as seen in figure 7(b). At this point
most reachability is due to the vicinity of the source.
4.1.2. Varying maximum contact distance (r)
Figure 8 shows the effect of increasing r on reachability.
Since contacts are selected between 2R and r hops from the
source, higher values of r provide a wider region for contact
selection. The mechanisms for the edge method for contact
selection described earlier provide selection of contacts that
have vicinities with reduced overlaps. This implies that as r
increases a larger number of contacts can be selected with-
out having vicinity overlaps. Therefore reachability increases
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 107

(a) (b)
Figure 7. Effect of vicinity radius (R) on reachability. N = 500, area = 710 m710 m, propagation range = 50 m, r = 16, NoC = 10, D = 1. (a) Histogram
of reachability for different values of R; (b) average reachability with R.

(a) (b)
Figure 8. Effect of maximum contact distance (r) on reachability. N = 500, area = 710 m 710 m, propagation range = 50 m, R = 3, NoC = 10, D = 1.
(a) Histogram of reachability for different values of r; (b) average reachability with r.

(a) (b)
Figure 9. Effect of number of contacts (NoC) on reachability. N = 500, area = 710 m 710 m, propagation range = 50 m, R = 3, r = 10, D = 1.
(a) Histogram of reachability for different values of NoC; (b) average reachability with NoC.

with increase in r. Larger values of r also mean that the av-
erage contact path length would increase (as more contacts
are chosen at larger distances from the source). However,
once the vicinities of the contacts and the source become non-
overlapping, for r > (2R + 8), we see no significant increase
in reachability with further increase in r.
4.1.3. Varying number of contacts (NoC)
NoC specifies the maximum number of contacts to be selected
for each node. The actual number of contacts chosen may
be less than this value. This is because of the limited re-
gion available for choosing contacts for given R and r accord-
ing to the contact selection mechanism. Once this region has
108
been covered by vicinities of chosen contacts, choosing more
contacts in the same region is not possible as their vicinities
would overlap with the vicinities of the already chosen con-
tacts. Therefore the contact selection mechanism prevents the
selection of more contacts. This can be seen in figure 9, in
which the reachability initially increases sharply as more and
more contacts are chosen. However, the increase in reacha-
bility saturates beyond NoC = 6 as the actual number of con-
tacts chosen saturates due to the effect of overlapping vicini-
ties.
4.1.4. Varying depth of search (D)
D specifies the levels of contacts that are queried in a breadth
first manner. When D = 1, a source node looking for a re-
source beyond its vicinity, queries its first level contacts only.
When D = 2, if none of the first level contacts contain the re-
source in its vicinity, second level contacts (contacts of the
first level contacts) are queried through the first level con-
tacts. As can be seen from the figure 10, reachability increases
sharply as the depth of search, D is increased. The depth of
search, D, results in a tree-like structure of contacts, improv-
ing the reachability and success rate of CARD.
Figure 10. Effect of depth of search (D) on reachability. N = 500, area =
710 m 710 m, Tx range = 50 m, R = 3, NoC = 10, r = 10.
(a)
Figure 12. Effect of number of contacts (NoC) on contact selection overhead. N = 500, area = 710 m 710 m, Tx range = 50 m, R = 3, r = 10, D = 1.
(a) Overhead over time for different values of NoC; (b) average (per sec) overhead for different values of NoC.
HELMY ET AL.
4.1.5. Varying network size
Figure 11 shows the reachability distribution for three differ-
ent network sizes, N = 250, 500 and 1000 nodes. The area of
the three networks has been chosen so that the node density
is almost same across the three networks. Figure 11 shows
that for a given network (specified by the values of N and the
area), the values of R and r can be configured to provide a
desirable reachability distribution in which most of the nodes
have a high value of reachability.
4.2. Contact selection and maintenance overhead analysis
The overhead analysis measures the number of control mes-
sages required for contact selection and maintenance. The
query overhead is considered in the next section. The over-
head considered in this section includes:
1. Contact selection overhead: This is the amount of CS traf-
fic generated for selecting new contacts. This includes
overhead due to backtracking as described earlier.
2. Contact maintenance overhead: This is the traffic gener-
ated by the contact path validation messages. Local recov-
Figure 11. Reachability for different network sizes (D = 1).
(b)
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY
ery, as described earlier, helps in reducing this part of the
total overhead.
Results are shown for scenario number 5 in table 1 (N =
500, area = 710 m 710 m). Similar trends were observed
for other scenarios.
4.2.1. Varying number of contacts (NoC)
As shown in figure 12, as the number of contacts increases
the maintenance overhead increases sharply as more nodes
are periodically maintained through the validation scheme.
4.2.2. Varying maximum contact distance (r)
As r increases the number of selected contacts increases. The
increase in the number of contacts is due to the availabil-
ity of a wider area for choosing contacts. Moreover, with
higher values of r, contacts may lie at greater distances from
the source. That is, the contact path length is expected to be
higher for larger values of r. This suggests that the mainte-
nance overhead should increase with increase in r. However,
as shown in figure 13, the overhead actually decreases with in-
crease in r. Figure 14 explains this decrease in maintenance
overhead. Figure 14 shows that as the value of r increases
the backtracking overhead decreases significantly. Recall that
Figure 13. Effect of maximum contact distance (r) on contact selection over-
head. N = 500, area = 710 m 710 m, Tx range = 50 m, NoC = 5, R = 3,
D = 1.
(a)
Figure 14. Effect of maximum contact distance (r) on backtracking overhead. N = 500, area = 710 m 710 m, propagation range = 50 m, NoC = 5,
R = 3, D = 1. (a) Backtracking over time for different values of r; (b) contact selection and backtracking overheads (per sec).
109
backtracking occurs when a node receiving a CS cannot be-
come a contact due to overlap with already existing contacts.
As r increases, the possibility of this overlap decreases due
to availability of a wider area for contact selection. This de-
crease in back-tracking overhead is significantly more than
the increase in overhead due to increased number of contacts
and contact path length. Therefore, the total contact selection
and maintenance overhead decreases.
4.3. Maintenance overhead over time
Figure 15 shows the maintenance overhead per node over a
20 s period for Vmax = 20 m/s. The maintenance overhead
decreases steadily with time. However, the number of con-
tacts increases slightly. This suggests that the source nodes
find more stable contacts over time. Stable contacts may be
defined as those nodes that have low velocity relative to the
source node. For example, a node moving in the same direc-
tion as source node with similar velocity could prove to be
a stable contact. Hence, over time, CARD leads to source
nodes finding more stable contacts.
4.4. Comparison with related schemes (query overhead and
total overhead)
We compare the performance of CARD to that of flooding,
smart flooding [19] and bordercasting [20], in terms of aver-
age query overhead and overall overhead. Simulations were
repeated several times with various random seeds to filter out
the noise.
Figure 16 shows the average traffic generated per query
for the three protocols. We select random source-destination
pairs in the network (the same pairs were used for all the three
protocols). The graph shows the average overhead for random
queries with different network sizes, for each protocol. The
overhead includes number of transmissions as well as number
of receptions. Therefore the overhead for flooding is about
twice the number of links (as expected). Bordercasting is
implemented as described in [20]. We implemented query
detection (QD1 and QD2) and early termination (ET) as de-
scribed in [20] to improve the performance. For smart flood-
(b)
110
Figure 15. Variation of overhead with time. N = 250, area = 710 m710 m,
Tx range = 50 m, NoC = 6, R = 4, r = 16, D = 1.
ing we investigated several techniques (probabilistic flood-
ing, minimum dominating set, counter based methods) and
we show the results for those settings that achieved success
rate of 90%. This was equivalent to probabilistic flooding as
in [19] with p = 0.65. For CARD the values of R and r used
were chosen as the values that gave maximum reachability for
that particular network size. This information was obtained
from previous results shown under the analysis of CARD with
respect to various parameters (see figure 11. Reachability for
different network sizes). Flooding and bordercasting result in
100% success in queries, smart flooding achieved 90% suc-
cess rate, and CARD showed a 95% success rate with D = 3.
CARDs success rate can be increased by increasing D, or
with resource replication. No replication is assumed in our
study. As can be seen from figure 16, CARD leads to sig-
nificant savings in communication overhead over the other
two approaches. CARD incurs, on average, around 5% of
the query overhead for flooding, and around 10% or more of
the query overhead of bordercasting or smart flooding. We
note that smart flooding achieves the least success rate. To
increase the success rate for smart flooding the overhead ap-
proaches that of flooding.
What is not shown in figure 16, however, is the effect
of contact and vicinity maintenance. For that we show the
following total overhead comparison results. Maintenance
overhead (for contacts and vicinity) is a function of mobil-
ity and simulation time. Its cost is amortized over the num-
ber of queries performed during that period. Hence, we
present our results as function of the query rate per mobil-
ity per node (i.e., query/sec/(m/s) or query/m); this is re-
ferred to as call-to-mobility ratio (CMR) or q query/m per
node. We show results for simulations with Vmax = 1 m/s
and 20 m/s, for various query rates q for 20 seconds of sim-
ulated time. These results take into consideration the con-
tact selection and maintenance overhead, the vicinity estab-
lishment and maintenance overhead and the query overhead.
As can be seen from figures 17 and 18, the advantage of us-
ing contacts becomes clearer for higher query rates, where
the cost of maintenance is amortized over a large number of
queries. For low mobility, in figures 17(a) and (b), the main-
tenance overhead is low and the advantages of using con-
tacts are the clearest (4685% savings for low query rates
HELMY ET AL.
Figure 16. Query overhead for CARD, flooding and bordercasting.
q = 0.005 query/m, and 8694% savings for high query rates
q = 0.05 to 0.5 query/m).
For high mobility, in figures 18(a), (b) the savings are less
than low mobility scenarios, nonetheless they are still signif-
icant for moderate to high query rates (2275% savings for
q = 0.05 query/m, 7993% savings for q = 0.5 query/m
over flooding or bordercast). For low query rates and high
mobility however, e.g., for 20 m/s and q = 0.005 query/m,
CARD and bordercasting perform worse than flooding, where
maintenance overhead dominates and only very few queries
are triggered (an unlikely scenario in mobile ad hoc net-
works). For high mobility, large-scale, high query rates (1000
nodes, 20 m/s, 0.5 query/m), we get savings between 79%
(vs. bordercasting) and 87% (vs. flooding).
To further understand the effect of query rate and mobility
on the total overhead we investigate the overhead ratio (OR)
metric for CARD over the total overhead of bordercast and
flooding. This metric enables us to have a more comprehen-
sive view of the operating conditions under which CARD is
favorable. Let OR(C/B) be the overhead ratio for CARD over
bordercast, and OR(C/F) and OR(C/S) be the overhead ratio
of CARD over flooding and smart flooding. Let CSM be the
contact selection and maintenance overhead, and let ZO be
the zone (or vicinity) maintenance overhead, both in packets
per node per m/s. Also, let CQO be the CARD query over-
head in packets per query, hence q CQO is the overhead in
packets per node per m/s. Define BQO as the query overhead
for bordercast. Hence, we get
CSM + ZO + q CQO
OR(C/B) = .
ZO + q BQO
Similarly, we have
CSM + ZO + q CQO
OR(C/F) = and
q FQO
CSM + ZO + q CQO
OR(C/S) = ,
q SQO
where FQO and SQO is the flooding and smart flooding over-
head in packets per query, respectively.
OR(C/B) and OR(C/F) were evaluated for q = 0.01 to
100 query/sec/(m/s) per node. Figure 19 shows results for
OR(C/B) and figure 20 shows results for OR(C/F). From the
figures we note that, in general, when q is quite small (e.g.,
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 111

(a) (b)
Figure 17. Total overhead for low mobility and different query rates. (a) Vmax = 1 m/s, CMR q = 0.005 query/m. (b) Vmax = 1 m/s, CMR q = 0.05 to
0.5 query/m.

(a) (b)
Figure 18. Total overhead for high mobility and different query rates. (a) Vmax = 20 m/s, CMR q = 0.05 query/m. (b) Vmax = 20 m/s, CMR
q = 0.5 query/m.

the call-to-mobility ratio during active periods. Hence, it is

Figure 19. OR(C/B): the overhead ratio for CARD over bordercast for vari-
ous values of q.
q < 0.01) then CARD incurs more overhead than flooding
and bordercasting. This is due to the fact that CARD ex-
pends communication overhead to select and maintain con-
tacts, as well as vicinities. If the nodes are relatively idle,
resulting in very small q, then there is not enough query to
amortize the cost of the maintenance overhead. This sce-
nario is unlikely though, as we expect idle nodes to transit
into sleep mode (to conserve energy) and not participate in
periodic activities (such as vicinity and contact maintenance)
while idle. From that perspective, one may consider q to be
unlikely that q will become too small for most practical pur-
poses. As q becomes moderate (around q = 0.01 query/m)
we start noticing the advantage of CARD in overhead sav-
ings. In figure 18 we see that OR(C/B) becomes less than 1
(the cross over point) for q 0.010.025 query/m. Also,
OR(C/B) becomes less than 0.2 (i.e., 80% overhead savings)
for q 0.2950.315 query/m for 500 and 1000 nodes and
q = 0.810 query/m for 250 nodes. For q 10 query/m
OR(C/B) approaches 0.11 for 500 and 1000 nodes and 0.18
for 250 nodes; i.e., over 80% saving in overhead.
In figure 20 we observe that OR(C/F) becomes less than 1
for q 0.0150.02 query/m. Furthermore, OR(C/F) < 0.2
for q 0.140.155 query/m, and OR(C/F) < 0.1 for
q 0.430.51 query/m. For q 10 query/m, the over-
head ratio OR(C/F) approaches 0.066; i.e., over 93% saving
in overhead. In figure 21 the overhead ratio with respect to
smart flooding is shown. In addition to achieving better suc-
cess rate than smart flooding, CARD also achieves less to-
tal overhead for all values of q > 0.035 query/m. The ratio
OR(C/S) goes below 0.2 for q 0.220.3 query/m, and goes
below 0.1 for q 0.929.8 query/m, approaching 6.69.9%
(i.e. more than 90% in overhead savings) when q approaches
10 query/m.
112
Figure 20. OR(C/F): the overhead ratio for CARD over flooding for various
values of q.
Figure 21. OR(C/S): the overhead ratio for CARD over smart flooding for
various values of q.
5. Conclusions
In this paper we presented the CARD architecture for re-
source discovery and small transfers in large-scale ad hoc
networks. The main contributions of this paper include the
introduction of a contact-based architecture that explicitly
trades-off route optimality (as in shortest path routes) for
communication and energy efficiency, along with a proactive
contact selection scheme to reduce vicinity overlap. Unlike
existing routing protocols, instead of expending significant
overhead to discover shortest path routes, CARD explicitly
focuses on route discovery or query delivery with the least
overhead, even if the routes used are suboptimal. We be-
lieve such trade-off is appropriate for our target applications,
mainly resource discovery and small transfers. Salient fea-
tures of our architecture include its ability to operate without
requiring any location information or any complex coordina-
tion. In our architecture, each node proactively discovers re-
sources within its vicinity. Based on small world concepts,
we have utilized the notion of contacts to serve as short cuts
that increase reachability beyond the vicinity. Two protocols
for contact selection were introduced and evaluated: (a) prob-
abilistic method and (b) edge method. The edge method was
found to result in more reachability and less overhead during
HELMY ET AL.
selection due to reduced backtracking, and was thoroughly
analyzed over the various dimensions of the parameter space
(including R, r, D, NoC, and network size). We further com-
pared our approach to flooding and bordercasting. The overall
overhead experienced by CARD was found to be significantly
lower than the other approaches. Overhead savings are func-
tion of the query rate, reaching over 93% (vs. flooding and
smart flooding) and over 80% (vs. bordercasting) in commu-
nication saving for high query rates; a drastic improvement in
performance.
These results show a lot of promise for the contact-based
approach to support short transfers in many applications of ad
hoc networks. One possible future research direction to inves-
tigate is to integrate CARD with other routing protocols (e.g.,
ZRP), where CARD may be used as the resource discovery
(and transaction routing) protocol. Similarly, we plan to in-
vestigate the integration of CARD in other data dissemination
protocols for sensor networks, such as directed diffusion [12].
Instead of using flooding, CARD maybe use for efficient re-
source discovery.
References
[1] L. Breslau, D. Estrin, K. Fall, S. Floyd, J. Heidemann, A. Helmy,
P. Huang, S. McCanne, K. Varadhan, Y. Xu and H. Yu, Advances in
network simulation, IEEE Computer (May 2000).
[2] T.-W. Chen and M. Gerla, Global state routing: A new routing scheme
for ad-hoc wireless networks, in: Proc. of the IEEE Internat. Conf. on
Communications (ICC) (1998).
[3] C.-C. Chiang, Routing in clustered multihop, mobile wireless networks
with fading channel, in: Proc. of IEEE SICON97 (April 1997).
[4] T. Clausen, P. Jacquet, A. Laouiti, P. Muhlethaler, A. Qayyum and
L. Viennot, Optimized link state routing protocol, in: Proc. of IEEE
INMIC (2001).
[5] Z. Haas and M. Pearlman, The zone routing protocol (ZRP) for ad hoc
networks, IETF Internet draft for the Manet group (June 1999).
[6] W. Heinzelman, J. Kulik and H. Balakrishnan, Adaptive protocols for
information dissemination in wireless sensor networks, in: The ACM
MOBICOM Conf., Seattle, WA (August 1999).
[7] A. Helmy, Architectural framework for large-scale multicast in mo-
bile ad hoc networks, IEEE Internat. Conf. on Communications (ICC),
Vol. 4, New York (April 2002) pp. 20362042.
[8] A. Helmy, Small worlds in wireless networks, IEEE Communications
Letters 7(10) (2003) 490492.
[9] A. Helmy, Mobility-assisted resolution of queries in large-scale mo-
bile sensor networks (MARQ), Computer Networks (Special Issue on
Wireless Sensor Networks) 43(4) (2003) 437458.
[10] A. Helmy, TRANSFER: Transactions routing for ad-hoc networks with
efficient energy, in: IEEE Global Communications Conf. (GLOBE-
COM) (December 2003).
[11] A. Helmy, S. Garg, P. Pamu and N. Nahata, Contact-based architecture
for resource discovery (CARD) in large scale MANets, in: IEEE/ACM
IPDPS Internat. Workshop on Wireless, Mobile and Ad Hoc Networks
(WMAN) (April 2003) pp. 219227.
[12] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion:
A scalable and robust communication paradigm for sensor networks,
in: ACM MobiCOM Conf. (August 2000).
[13] D.B. Johnson and D.A. Maltz, The dynamic source routing protocol for
mobile ad hoc networks, IETF Internet draft (October 1999).
[14] J. Li, J. Jannotti, D. Couto, D. Karger and R. Morris, A scalable location
service for geographic ad hoc routing, in: The ACM MOBICOM Conf.
(2000).
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY
[15] J. Liu, Q. Zhang, W. Zhu, J. Zhang and B. Li, A novel framework for
QoS-aware resource discovery in MANets, in: IEEE Internat. Conf. on
Communications (ICC) (May 2002).
[16] W. Lou and J. Wu, On reducing broadcast redundancy in ad hoc wire-
less networks, IEEE Transactions on Mobile Computing 1(2) (2002).
[17] M. Mitzenmacher, Compressed bloom filters, in: The Twentieth ACM
Symposium on Principles of Distributed Computing (PODC) (August
2001).
[18] S. Murthy and J.J. Garcia-Luna-Aceves, An efficient routing protocol
for wireless networks, Mobile Networks and Applications (Special Is-
sue on Routing in Mobile Communication Networks) (October 1996).
[19] S. Ni, Y. Tseng, Y. Chen and J. Sheu, The broadcast Storm problem
in a mobile ad hoc network, in: Proc. of the ACM MOBICOM Conf.
(August 1999) pp. 151162.
[20] M. Pearlman and Z. Haas, Determining the optimal configuration for
the zone routing protocol, IEEE Journal on Selected Areas in Commu-
nications 8 (1999) 13951414.
[21] C.E. Perkins and P. Bhagwat, Highly dynamic destination-sequenced
distance-vector routing (DSDV) for mobile computers, ACM Computer
Communications Review (October 1994) 234244.
[22] C.E. Perkins, E.M. Royer and S.R. Das, Ad hoc on-demand distance
vector routing, IETF Internet draft (October 1999).
[23] N. Sadagopan, B. Krishnamachari and A. Helmy, Active query for-
warding in sensor networks (ACQUIRE), Ad Hoc Networks Journal
(2004) to appear.
[24] N. Sadagopan, B. Krishnamachari and A. Helmy, The ACQUIRE
mechanism for efficient querying in sensor networks, in: First IEEE
Internat. Workshop on Sensor Network Protocols and Applications
(SNPA), in conjunction with IEEE ICC, Anchorage (May 2003)
pp. 149155.
[25] S. Wang and A. Helmy, Effects of small transfers and traffic patterns
on performance and cache efficacy of ad hoc routing, (poster), in: The
ACM MOBICOM Conf. (The Ninth Annual Internat. Conf. on Mobile
Computing and Networking), San Diego, CA (September 2003).
[26] D.J. Watts, The dynamics of networks between order and randomness,
in: Small Worlds (Princeton Univ. Press, Princeton, 1999).
[27] D. Watts and S. Strogatz, Collective dynamics of small-world net-
works, Nature 393 (4 June 1998).
113
Ahmed Helmy received his Ph.D. in computer sci-
ence (1999), M.S. in electrical engineering (1995)
from the University of Southern California, M.S.
Eng. Math. (1994) and B.S. in electronics and com-
munications engineering (1992) from Cairo Univer-
sity, Egypt. Since 1999, he has been an Assistant
Professor of Electrical Engineering at the University
of Southern California. In 2002, he received the Na-
tional Science Foundation (NSF) CAREER Award.
In 2000 he received the USC Zumberge Research
Award, and in 2002 he received the best paper award from the IEEE/IFIP
International Conference on Management of Multimedia Networks and Ser-
vices (MMNS). In 2000, he founded and is currently directing the
wireless networking laboratory at USC. His current research interests lie
in the areas of protocol design and analysis for mobile ad hoc and sensor
networks, mobility modeling, design and testing of multicast protocols, IP
micro-mobility, and network simulation.
E-mail: helmy@usc.edu
Saurabh Garg received his M.S. in computer science from the University of
Southern California in May 2003 and B.S. in engineering from University of
Delhi, India. He is currently working as a Programmer for National Center
for Ecological Analysis and Synthesis (NCEAS) affiliated with University
of California, Santa Barbara. He has worked with researchers in fields of
wireless networking, linguistics dialogue management and ecology. Before
joining NCEAS he was working for Institute of Creative Technologies at the
University of Southern California in the field of dialogue management. In
wireless networking, he has worked on network simulation, protocol design
and analysis.
E-mail: sgarg@usc.edu
Nitin Nahata received his M.S. in computer science from the University of
Southern California in December 2002. He is currently working as a software
engineer at Dynamix Technologies Inc.
E-mail: nnahata@usc.edu
Priyatham Pamu received his M.S. in computer science from the University
of Southern California in December 2002.
E-mail: pamu@usc.edu
 Mobile Networks and Applications 10, 115131, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Energy-Balanced Task Allocation for Collaborative Processing

in Wireless Sensor Networks
YANG YU and VIKTOR K. PRASANNA
Department of Electrical Engineering, University of Southern California, Los Angeles, CA 90089-2562, USA

Abstract. We propose an energy-balanced allocation of a real-time application onto a single-hop cluster of homogeneous sensor nodes
connected with multiple wireless channels. An epoch-based application consisting of a set of communicating tasks is considered. Each
sensor node is equipped with discrete dynamic voltage scaling (DVS). The time and energy costs of both computation and communication
activities are considered. We propose both an Integer Linear Programming (ILP) formulation and a polynomial time 3-phase heuristic. Our
simulation results show that for small scale problems (with
10 tasks), up to 5x lifetime improvement is achieved by the ILP-based approach,
compared with the baseline where no DVS is used. Also, the 3-phase heuristic achieves up to 63% of the system lifetime obtained by the
ILP-based approach. For large scale problems (with 60100 tasks), up to 3.5x lifetime improvement can be achieved by the 3-phase heuristic.
We also incorporate techniques for exploring the energy-latency tradeoffs of communication activities (such as modulation scaling), which
leads to 10x lifetime improvement in our simulations. Simulations were further conducted for two real world problems LU factorization
and Fast Fourier Transformation (FFT). Compared with the baseline where neither DVS nor modulation scaling is used, we observed up to
8x lifetime improvement for the LU factorization algorithm and up to 9x improvement for FFT.
Keywords: sensor networks, single-hop wireless networks, ILP, energy saving

1. Introduction
Wireless sensor networks (WSNs) are being developed for a
wide range of civil and military applications, such as target
tracking, infrastructure monitoring, habitat sensing, and bat-
tlefield surveillance [6,10]. WSNs usually contain a number
of networked sensor nodes with each sensor node consisting
of computation, communication, and sensing devices. These
sensor nodes collaborate with each other to realize certain ap-
plications.
For instance, in a target tracking application, up to thou-
sands of sensor nodes are dispersed over a specific area of
interest. The sensor nodes are usually organized into clus-
ters [13,31] with each cluster consisting of tens of sensor
nodes. Distributed signal detection and collaborative data
processing are performed within each cluster for detecting,
identifying, and tracking vehicles. Some of the operations
involved in such data processing include the LU factoriza-
tion [5] and the Fast Fourier Transformation (FFT) [7].
Energy efficiency is a key concern in WSNs. The large
number of sensor nodes involved in the system and the need
to operate over a long period of time require energy-aware
design and operation at all levels of abstraction, from the
physical layer to the application layer. However, while many
hardware techniques [1,14], network protocols [13,16], and
data processing algorithms [18,19] have been proposed for
energy-aware design, systematic mechanisms for designing
energy-aware collaborative processing between sensor nodes
still need to be addressed.
This work is supported by the DARPA Power Aware Computing and Com-
munication Program under contract no. F33615-C-00-1633. A preliminary
version of this paper appears in ACM LCTES 2003.
The state of the art in WSN design is largely ad-hoc sys-
tem planning and resource management are done without a
systematic methodology. This can lead to inefficient utiliza-
tion of the system. The main motivation of our efforts is to
develop techniques for systematic and rapid design and de-
ployment of WSN applications [3,25,31].
We focus on the development of energy-efficient collabo-
rative algorithms for WSNs based on high-level computation
models of WSNs. Such high-level models allow designers to
make informed decisions regarding energy and time tradeoffs
at the node and network level creating a modular, layered
paradigm for application development. Toward such a goal,
we study the following problem in this paper.
Energy-balanced task allocation problem. We consider a
single-hop cluster of homogeneous sensor nodes connected
through multiple wireless channels. Each sensor node is
equipped with dynamic voltage scaling (DVS) [30]. The tar-
get application consists of a set of communicating tasks.
Throughout the paper, the term activity refers to either a com-
putation task or a communication request. We consider an
epoch-based scenario [18], where an instance of the appli-
cation is executed during the beginning of each epoch and
must be completed before the end of the epoch. Such a re-
quirement is usually called the latency constraint. We use
the term period to indicate the length of each epoch. Also,
we assume that time-synchronization schemes (e.g., [9]) are
available within the cluster.
We consider the exclusive access constraint. Specifically, a
non-preemptive scheduling policy is employed by each sensor
node and each wireless channel. Also, at any time, a sensor
node can receive or send data by using at most one channel.
116
The underlying network protocol is assumed to be capable of
scheduling a communication activity over a specified chan-
nel according to the start and finish time of the activity. Such
a scheduling policy requires coarse-level bandwidth reserva-
tion mechanisms, which can be provided by, for example, a
time-division multiple-access (TDMA) protocol. Moreover,
we consider the task placement constraint, which is typically
required when certain tasks for sensing the raw data must be
allocated onto different sensor nodes.
A task allocation is defined as (1) the assignment of tasks
onto sensor nodes, (2) the voltage settings of tasks, (3) the
assignment of communication activities onto channels, and
(4) the scheduling of computation and communication activ-
ities. Our general goal is to find an allocation in order to
maximize the lifetime of the cluster. Toward such a goal, we
propose an energy-balanced task allocation such that the max-
imal energy dissipation among all sensor nodes during each
period is minimized, subject to the latency, exclusive access,
and task placement constraints.
Our contributions. The idea of energy-balanced task alloca-
tion to a single-hop cluster in WSNs is proposed. As we shall
see in section 2, most of the previous efforts in energy-aware
task allocation or resource management try to minimize the
overall energy dissipation of the system. This strategy may
not be suitable in the context of WSNs, since each sensor node
is equipped with its own energy source. Moreover, for event-
driven systems, applications often need to be executed after
the system has been working for sometime. In such a case, an
energy-balanced task allocation should also consider the fact
that the remaining energy can vary among sensor nodes.
To the best of the authors knowledge, this is the first work
for task allocation in WSNs that considers the time and en-
ergy costs of both the computation and communication ac-
tivities. We first present an integer linear programming (ILP)
formulation of our problem. The optimal solution of the prob-
lem can be obtained by using a commercial software pack-
age such as [26], though the running time of such a software
can be large. Next, we propose a polynomial time 3-phase
heuristic. Finally, we incorporate techniques that explore the
latency-energy tradeoffs of communication activities, such as
modulation scaling [23].
Our simulation results show that for small scale problems,
up to 5x lifetime improvement is achieved by the ILP-based
approach, compared with the case where no DVS is used.
Also, the 3-phase heuristic achieves up to 63% of the sys-
tem lifetime obtained by the ILP-based approach. For large
scale problems, the 3-phase heuristic achieves up to 3.5x life-
time improvement when only DVS is used. By incorporat-
ing modulation scaling, up to 10x lifetime improvement was
observed. Simulations were also conducted for application
graphs from two real world problems LU factorization and
FFT. We observed a lifetime improvement of up to 8x for the
LU factorization algorithm and up to 9x for FFT.
Paper organization. We discuss the related work in sec-
tion 2. The energy-balanced task allocation problem is de-
fined in section 3. The ILP formulation of the problem is
YU AND PRASANNA
given in section 4. The 3-phase heuristic is described in sec-
tion 5. Techniques, such as modulation scaling, are incor-
porated into our approaches in section 6. Simulation results
are demonstrated in section 7. Finally, we give concluding
remarks in section 8.
2. Related work
Extensive research efforts have studied the problem of energy-
efficient task allocation and scheduling with DVS in uni-
processor real-time systems, including [2,15,24,30]. Re-
cently, research interests have been shifted to multi-processor
systems. A list-scheduling based heuristic is proposed in [12],
to dynamically recalculate the priority of communicating
tasks. In [17], static and dynamic variable voltage schedul-
ing heuristics for real-time heterogeneous embedded systems
are proposed. An approach based on critical-path is used for
selecting the voltage settings of tasks. However, both [12]
and [17] assume that the task assignment is given. A sim-
ilar problem to the one studied in this paper is investigated
in [33]. A two-phase framework is presented to first deter-
mine the allocation of tasks onto processors and then the volt-
age settings of tasks using convex programming. In [34], a
dynamic processor voltage adjustment mechanism for a ho-
mogeneous multi-processor environment is discussed. How-
ever, the time and energy costs for communication activities
are not addressed in any of [12,33,34].
The goal of all the above works is to minimize the overall
energy dissipation of the system. While such a goal is reason-
able for tightly coupled systems, it does not capture the nature
of WSNs. The reason is that to minimize the overall energy
dissipation can lead to heavy use of energy-effective sensor
nodes, regardless of their remaining energy. The consequent
short lifetime of such sensor nodes will very likely hinder the
system from delivering required performance. This weakness
is a major motivation of the proposed energy-balanced task
allocation.
Our work considers the energy and time costs of both com-
putation and communication activities. As indicated by sev-
eral researches, wireless communication is a major source of
energy dissipation in WSNs. By incorporating techniques
such as modulation scaling, we can greatly improve the
energy-efficiency of the system.
Energy-balanced task allocation bears some resemblance
to load-balance in distributed computing. However, the com-
munication activities over the same wireless channel need to
be serialized such that run-time contentions can be avoided.
The serialization imposes new challenges that distinguish our
problem from most of the existing works for load-balance or
real-time scheduling in distributed systems.
3. Problem definition
3.1. System model
We consider a set of m homogeneous sensor nodes, PE =
{PEi : i = 1, . . . , m}, connected by a single-hop wireless
ENERGY-BALANCED TASK ALLOCATION
network with K communication channels. The homogene-
ity refers to the same processor and radio capabilities. Each
sensor node is equipped with D discrete voltage levels, listed
as V = {Vi : i = 1, . . . , D} in decreasing order. Each voltage
level in V corresponds to a specific computation speed (given
in cycles per second) of the processor. Let SPj denote the
speed of Vj . Let Ri denote the remaining energy of PEi . For
ease of analysis, we assume that the processors consume zero
power during idle state.
Regarding the exclusive access constraint, we assume that
a non-preemptive scheduling policy is employed by each sen-
sor node and each wireless channel. In other words, the
time duration scheduled for different computation (commu-
nication) activities over the same sensor node (wireless chan-
nel) cannot overlap with each other. Moreover, the under-
lying communication protocols are assumed to be capable
of scheduling communication activities according to the start
time of each activity in order to avoid run-time contentions.
We assume all channels have the same bandwidth. Let de-
note the time for transmitting one data unit between two sen-
sor nodes over any channel. For ease of analysis, we assume
that such a transmission costs the same amount of energy at
both the sender and the receiver, denoted by . Let s and s
denote the startup time and energy costs for communication.
The data transmission between two tasks on the same sensor
node is performed through the local memory with zero time
and energy costs.
For ease of analysis, we assume that the radios are com-
pletely shutdown in idle state. The energy cost for shutting
down and restarting the radio is assumed to be included in s .
Low power paging or signaling channel mechanisms can be
used for synchronization between sensor nodes when the ra-
dios are shutdown. However, the modeling of the power con-
sumption for such mechanisms is beyond the scope of this
paper. We also assume that computation and communication
activities can be parallelly executed on any sensor node.
3.2. Application model
An epoch-based application [18] consisting of a set of com-
municating tasks is considered. Let P denote the period of the
application, which is the length of each epoch. An instance of
the application is activated at time kP , and must be completed
by the relative deadline, (k + 1)P , where k = 0, 1, 2, . . . .
The structure of the application is represented by a directed
acyclic graph (DAG), G = (T , E), where node set T denotes
the set of n tasks, {Ti : i = 1, . . . , n}, and edge set E de-
notes the set of e directed communication activities between
tasks, {Ei : i = 1, . . . , e}. Every edge in E pointing from
node Ti to Tj , denoted as (i, j ), means that the output of
task Ti needs to be transmitted to Tj before Tj can start com-
putation. There is a precedence constraint on two tasks Ti
and Tj , if there is a path of alternate nodes and edges from Ti
to Tj in the DAG. Similarly, there is a precedence constraint
on two communication activities, (i, j ) and (i  , j  ), if there is
a path from Tj to Ti  . A task with no incoming edges is called
117
a source task. A task with no outgoing edges is called a sink
task.
For most applications in WSNs, the source tasks are used
for sensing or gathering raw data. For ease of analysis, the
task placement constraint is defined as that no two source
tasks can be assigned to the same sensor node. Nevertheless,
our models and approachs can be extended to handle the gen-
eral case that any pair of tasks must be or must not be assigned
to the same sensor node.
For any task Ti T , let Ci denote its workload in terms of
the worst-case number of required computation cycles. The
execution time of Ti on any voltage level Vj V , tij , can
be calculated as tij = Ci /(SPj ). The voltage level of a sen-
sor node is assumed to be dynamically switched, if necessary,
upon the arrival of a task instance. Because at most one switch
is needed for executing a task instance, the associated time
overhead is assumed to be included in the workload of the
task. From [4], the power consumption for executing a task
follows a monotonically increasing and strictly convex func-
tion of the computation speed, gi (), which can be represented
as a polynomial function of at least second degree. Hence, the
energy dissipation for executing Ti on Vj , eij , can be calcu-
lated as eij = gi (SPj )tij . The exact forms of gi () can vary
for different tasks based on their instruction components.
The communication load of any edge Ei E is rep-
resented by its weight wi , as the number of data units to
be transmitted. We assume that all the data of an edge is
transmitted in one data packet with variable size. For an
edge Ei = (j, k), let ti and ei denote the time and energy
costs of the corresponding communication activity, if tasks
Tj and Tk are not assigned to the same sensor node. We have
ti = s + wi and ei = s + wi .
3.3. Task allocation
Based on the above system and application models, a task al-
location is defined as (1) the assignment of tasks onto sensor
nodes, (2) the voltage settings of tasks, (3) the assignment of
communication activities onto channels, and (4) the schedul-
ing of computation and communication activities. Each task
can be assigned to exactly one sensor node with a fixed volt-
age setting. Also, each communication activity can be as-
signed to exactly one channel. An allocation is feasible if
it satisfies the latency, exclusive access, and task placement
constraints.
The system lifetime is defined as the time duration from the
time when the application starts execution to the time when
any sensor node in the cluster fails due to depleted energy.
A general solution to maximize the system lifetime is to allow
variable task allocations in different periods. Consequently,
the energy cost for each sensor node may vary in different pe-
riods. However, due to the high complexity raised by such a
solution, we assume that the task allocation remains the same
for all application periods. That is, the behavior of the system
repeats for each period and every sensor node spends the same
energy duration each period. Let Ei denote the energy dissi-
pation of PEi PE during each application period. Given an
118
allocation, the system lifetime (in number of periods) can be
calculated as mini {Ri /Ei }. A feasible allocation is optimal
if the corresponding system lifetime is maximized among all
the feasible allocations.
Note that a more complex definition of the system lifetime
would be the time period from the beginning of the applica-
tion execution to the time when not enough sensor nodes are
alive to deliver required performance. However, such a defin-
ition is quite application-specific. Thus, a simple but general
definition of the system lifetime is adopted in this paper. Now,
our task allocation problem can be informally stated as:
Find an allocation of a set of communicating tasks onto a
single-hop cluster that minimizes the maximal energy dis-
sipation among all sensor nodes during each application
period, normalized by their remaining energy.
4. Integer linear programming formulation
In this section, we present an ILP formulation of our task allo-
cation problem that captures the behavior of the system during
one application period. We first list the notations used in the
formulation as follows:
P: period of the application
tij , eij : time and energy costs of executing task Ti using
voltage level Vj
ti , ei : time and energy costs of edge Ei = (j, k), if Tj
and Tk are not assigned to the same sensor node
a||b: no precedence constraint exists for computation
(or communication) activities a and b
{xij }: a set of 01 variables such that xij equals one
iff Ti is assigned to PEj
{yij }: a set of 01 variables such that yij equals one
iff the voltage level of Ti is set to Vj
{zij }: a set of 01 variables such that zij equals one
iff Ei is assigned to the j th channel
{rij }: a set of 01 variables such that rij equals one
iff Ti and Tj are assigned to the same sensor node
{sij }: a set of 01 variables such that sij equals one
iff Ei and Ej are assigned to the same channel
{(i)}: a set of real variables indicating the time when
Ti starts execution
{(i)}: a set of real variables indicating the time when
Ti completes execution
{ (i)}: a set of real variables indicating the time when
Ei starts transmission
{(i)}: a set of real variables indicating the time when
Ei completes transmission
{pij }: a set of 01 variables such that zij equals one
iff the execution of Ti finishes before Tj starts
{qij }: a set of 01 variables such that qij equals one
iff the transmission of Ei finishes before Ej starts.
To capture the relative order imposed by the precedence
constraints among activities, we define the constraint set 1
shown in figure 1. It is easy to verify that the exclusive access
constraint for activities with precedence constraints is also
YU AND PRASANNA
enforced by constraint set 1. However, for activities that do
not have precedence constraints between them, an extra set of
constraints are needed (constraint set 2 in figure 2) to enforce
the exclusive access constraint. In addition, the task place-
ment constraint is captured by the constraint set 3 in figure 2.
The complete ILP formulation is given in figure 3, where
E is an auxiliary variable. In the figure, the factor |xik xj k |
means that the energy cost for (i, j ) is counted if exactly one
of Ti or Tj is assigned to PEk , but not both. Clearly, the pre-
sented formulation is nonlinear. It can be transformed into
an ILP formulation by standard linearization techniques [29].
Due to the space limitation, we omit the details of lineariza-
tion in this paper.
5. Heuristic approach
In this section, we describe an efficient 3-phase heuristic for
solving the task allocation problem. Initially, we assume
that the voltage levels for all tasks are set to the highest op-
tion (V1 ). In the first phase, the tasks are grouped into clusters
with the goal to minimize the overall execution time of the
application. In the second phase, task clusters are assigned to
sensor nodes such that the highest energy dissipation among
all sensor nodes, normalized by their remaining energy, is
minimized. In the last phase, the system lifetime is maxi-
mized by lowering the voltage levels of tasks. The details of
the heuristic are as follows.
Phase 1. A task cluster is defined as a set of tasks assigned to
the same sensor node with a specific execution order. Com-
munication between tasks within a cluster costs zero time and
energy. In this phase, we assume an unlimited number of sen-
sor nodes, implying that the number of clusters is also unlim-
ited. The main purpose of this phase is to eliminate communi-
cation activities in order to reduce the overall execution time
of the application.
The idea of phase 1 is similar to the algorithm pro-
posed in [22, pp. 123131]. However, traditional approaches
for task clustering usually assume a full connection among
processors such that all communication can be parallelized,
whereas in our problem, communication activities over the
same channel must be serialized. Thus, a new challenge is
to select a policy for the serialization that facilitates the re-
duction of the execution time of the application. We use a
simple first-come-first-serve policy to order the communica-
tion activities ready at different times. Activities ready at the
same time (such as those initiated by the same task) are exe-
cuted in a nondecreasing order of their communication loads.
Nevertheless, more sophisticated policies are also applicable.
The pseudo-code for phase 1 is shown in figure 4. In the
code, L denotes the overall execution time of the application
and C(i) denotes the cluster that contains task Ti . Initially,
every task is assumed to constitute a cluster by itself. We
then examine all the edges in a non-increasing order of their
weights. For each edge (i, j ) if the execution time of the ap-
ENERGY-BALANCED TASK ALLOCATION 119

Constraint set 1:
T
i T
x = 1, // every task can be assigned to exactly one sensor node

j ij
j yij = 1, // every task can be executed using exactly one voltage level
(i)  maxEl =(j,i)E

{(l)}, // Ti starts execution after receiving all input data
(i) = (i) + j (yij tij ), // execution time of Ti depends on its voltage level
Ti , Tj T
rij = 1 iff k = 1, . . . , m, xik = xj k // rij equals one if Ti and Tj are assigned to the same sensor node
E
i = (a, b) E
j zij = 1, // Ei can be assigned to exactly one channel
(i)  (a), // Ei starts transmission after Ta completes execution
(*) (i) = (i) + ti (1 rab ), // the transmission time of Ei depends on the locations of Ta and Tb
for any source tasks Ti
(i)  0, // all source tasks can start execution at time 0
for any sink task Ti
(i)
P . // all sink tasks must complete before the relative deadline

Figure 1. Constraint sets 1 for the ILP formulation.

Constraint set 2:
Ti , Tj T , such that i = j and Ti ||Tj
pij = 1 pj i ,
(j )  pij rij (i),
(i)  pj i rij (j ),
Ei , Ej E, such that Ei = (a, b),
Ej = (a, c), b = c
qij = 1 qj i ,
(j )  qij (1 rab )(1 rcd )(i),
(i)  qj i (1 rab )(1 rcd )(j ),
Ei , Ej E, such that Ei = (a, b),
Ej = (c, b), a = c
qij = 1 qj i ,
(j )  qij (1 rab )(1 rcd )(i),
(i)  qj i (1 rab )(1 rcd )(j ),
Ei , Ej E, such that Ei = (a, b),
Ej = (c, d), a = c, b = d, and Ei ||Ej
qij = 1 qj i ,
sij = 1 iff k = 1, . . . , K, zik = zj k ,
(j )  qij (1 rab )(1 rcd )sij (i),
(i)  qj i (1 rab )(1 rcd )sij (j ).
Constraint set 3:
Ti , Tj T , such that Ti and Tj are source tasks and i = j
rij = 0
// pij is the inverse of pj i
// if Ti and Tj are assigned to the same sensor node, Ti
// completes before Tj starts execution iff pij = 1
// if Ti and Tj are assigned to the same sensor node, Tj
// completes before Ti starts execution iff pj i = 1
// Communication activities from the same sensor node
// qij is the inverse of qj i
// Ei completes before Ej starts transmission iff qij = 1
// Ej completes before Ei starts transmission iff qj i = 1
// Communication activities to the same sensor node
// qij is the inverse of qj i
// Ei completes before Ej starts transmission iff qij = 1
// Ej completes before Ei starts transmission iff qj i = 1
// qij is the inverse of qj i
//sij equals one if Ei and Ej are assigned to the same channel
// if Ei and Ej are assigned to the same channel, Ei completes
// before Ej starts transmission iff qij = 1
// if Ei and Ej are assigned to the same channel, Ej completes
// before Ei starts transmission iff qj i = 1
// any two source tasks cannot be assigned to the same sensor node

Figure 2. Constraint sets 2 and 3 for the ILP formulation.

plication can be reduced by merging C(i) with C(j ) with-
out violating the task placement constraint, we perform the
merge. Otherwise, Ti and Tj remain in two different clusters.
In lines 3 and 6, the function Traverse() is called to traverse
the DAG in order to determine the schedule of the tasks and
hence L.
The pseudo code for Traverse() is shown in figure 5. In
the code, we maintain a queue of activities, Qact , that stores
all the ready computation or communication activities in their
expected execution order. We also maintain a timestamp for
each task cluster that indicates the finish time for all scheduled
tasks within the cluster. Similarly, we maintain a timestamp
120 YU AND PRASANNA

Minimize E 1. Initialize Qact

Subject to 2. Set the timestamps for all task clusters and channels to zero
PEk



 3. Append all source tasks to Qact with ready time set to zero
Ti T {xik j (yij eij )} + Ei =(a,b)E {ei |xak xbk |}

E 4. While Qact is not empty Do
Rk
and constraint sets 1, 2 and 3 5. Remove the first activity from Qact
6. If the removed activity is a computation activity, denoted
Figure 3. ILP formulation for the energy-balanced task allocation problem.
as Ti
7. Set (i) max{ready time of Ti , timestamp of C(i)}
1. Each task is assumed to constitute a cluster by itself
8. Set (i) to the expected completion time of Ti , i.e.,
2. Set E as the list of edges in a nondecreasing order (i) (i) + ti1
of the edge weights
9. Set the timestamp of C(i) to (i)
3. L Travese()
10. Insert all communication activities initiated by Ti into Qact
4. While E is not empty Do with ready time set to (i) in a nondecreasing order of
5. Remove the first edge from E, denoted as (i, j ) their communication loads

6. L Traverse() as if C(i) and C(j ) are merged 11. Else

7. If L < L and to merge C(i) and C(j ) does not 12. Let Ei = (a, b) denote the removed communication
activity
violate the task placement constraint
13. Find the channel with the smallest timestamp, say the j th
8. Merge C(i) and C(j )
channel
9. L L
14. Set (i) max{ready time of Ei , timestamp of the j th
10. If L > P , Return failure channel}
15. Set (i) to the expected completion time of Ei , i.e.,
Figure 4. Pseudo-code for phase 1.
(i) (i) + ti
for each channel that indicates its nearest available time. The 16. Set the timestamp of the j th channel to (i)
timestamps are used to schedule the computation and com-
17. Set the ready time of any unscheduled communication
munication activities in lines 7, 13 and 14. In lines 9 and 14, activities from Ta to (i)
the timestamps are updated based on the execution time of
the scheduled activities. The actions in lines 17 and 18 are 18. Set the ready time of any unscheduled communication
important to ensure that the radio can be tuned to at most one activities to Tb to (i)
channel at any time. 19. If all the communication activities to Tb have been
scheduled
Phase 2. In this phase, we assign the task clusters from 20. Insert Tb into Qact with ready time set to (i)
phase 1 onto the actual sensor nodes in PE. Note that multiple
clusters can be assigned to the same sensor node. Based on 21. Return the largest timestamp among all clusters
the contained tasks and the corresponding communication ac-
tivities, we first calculate the energy dissipation of each clus- Figure 5. Pseudo-code for function Traverse().
ter. Let  = [1 , 2 , . . . , c ] denote the list of all tasks clus-
ters and i denote the energy dissipation of i . The normal- plication based on the resulting assignment. Compared with
ized energy dissipation (norm-energy for short) of a sensor Traverse(), the modification in TraverseAssigned() is that in
node is given as the sum of the energy dissipation of the clus- line 7 of figure 5, each computation activity is scheduled on
ters assigned to the sensor node, normalized by the remaining the sensor node that it is assigned to. Thus, timestamps are
energy of the sensor node. maintained for all sensor nodes, instead of clusters.

The pseudo-code of phase 2 is shown in figure 6. Ini-
tially,  is sorted into a non-increasing order of energy dis-
sipation of clusters. Then, for each cluster in , we calcu-
lated the norm-energy of every sensor node as if the cluster
is assigned to the sensor node (called expected norm-energy).
We then assign the cluster to the sensor node that gives the
minimal expected norm-energy. In the code, function Tra-
verseAssigned() is used to find the execution time of the ap-
Phase 3. The voltage levels of tasks are adjusted in this phase
with the goal to maximize the system lifetime. An iterative
greedy heuristic is used (shown in figure 7). Let E denote
the maximum of the norm-energy among all sensor nodes.
The sensor node that determines E is called the critical node.
In each iteration, we find the task such that by lowering its
current voltage level to the next level, E can be decreased the
most. The increased latency caused by lowering the voltage
ENERGY-BALANCED TASK ALLOCATION 121

1. Sort  in a non-increasing order of the energy

dissipation of clusters
2. While  is not empty Do
3. Select the first element in 
4. Calculate the expected norm-energy for each
sensor node (set to infinity if two source tasks
are assigned to the same sensor node)
5. Assign to the sensor node that gives the minimal
expected norm-energy
6. Update the norm-energy of the sensor node
7. Remove from 
8. L TraverseAssigned()
9. If L > P , Return failure (a)

Figure 6. Pseudo-code for phase 2. Task time cost energy cost

Vh Vl Vh Vl
1. For each PEi , sort EDi in a non-increasing order
T1 10 33 20 6
2. Do T2 60 199 120 36
T3 10 33 20 6
3. i1
T4 10 33 20 6
4. Let PEr denote the critical sensor node and T5 20 66 40 12
E denote the norm-energy of PEr T6 10 33 20 6
T7 10 33 20 6
5. While i
|EDr | Do
6. Select the ith item in EDr ; (b)
let Tj denote the corresponding task
Figure 8. An application example: (a) application graph; (b) time and energy
7. If L + td j
P costs for executing tasks at voltage levels Vh and Vl .

8. L L + tdj
9. Lower the voltage of Tj to the next level
10. Update edj in EDr ; resort EDr if necessary
11. Find the new critical sensor node, PEr  ; update E
12. If r = r 
13. r r ; i 1
14. Else i i + 1
15. L TraverseAssigned()
16. Until E can not be reduced any more
Figure 7. Pseudo-code for phase 3.
is added to L. Since the schedule of activities can be changed
by the latency increment, L is re-computed by traversing the
DAG every time it reaches P (in line 15).
In figure 7, edj denotes the energy gain by lowering the
current voltage of Tj to the next level, while tdj denotes the
incurred increment in latency. The array composed by edj s
for all tasks assigned to PEi is denoted as EDi .
Time complexity analysis. In phase 1 (figure 4), the While
iteration is executed e times. Function Traverse() in line 6
takes O(n + e) time. Thus, phase 1 needs O(e(n + e)) time.
In phase 2 (figure 6), the ordering in line 1 takes O(c log c)
time. The outer iteration is executed c times. The results
of m possible assignments are compared in line 5. The tra-
verse in line 8 takes O(n + e) time. Hence, phase 2 takes
O(c log c + mc + n + e) time. In phase 3 (figure 7), the sorting
in line 1 takes O(n log n) time. The number of voltage switch-
ing in line 9 is bounded by dn. To update EDr in line 10
needs O(log n) time. Let p denote the number of times for
calling TraverseAssigned() in line 12. The time complexity
of phase 3 is O(dn log n + p(n + e)). Although p equals dn
in the worst case, it was observed in our simulations that p
usually equals 1 or 2. Thus, the overall time complexity of
the heuristic is O((e + p)(n + e) + mc + dn log n + c log c),
which is O(dn(n + e + log n) + e2 + mn) in the worst
case.
An illustrative example. We illustrate the execution of the
above heuristic through a simple example. We assume a clus-
ter of 3 sensor nodes connected by 2 channels. Each sensor
node have two voltage levels, Vh and Vl , with SPh = 1 and
SPl = 0.3. We assume that it costs one time and energy unit
for transmitting one data unit over any channel. The applica-
tion graph is shown in figure 8(a), with each circle represent-
ing a task. The number close to each circle is the required
122
workload, while the number on each edge is the weight of the
edge. The time and energy costs for executing tasks at the
two voltage levels are given in figure 8(b). We assume that
P = 250 time units.
The clustering steps in phase 1 is shown in figure 9. In this
phase, the voltage levels of all tasks are set to Vh . The sorted
edge list with respect to edge weights is {(T4 , T6 ), (T1 , T2 ),
(T3 , T6 ), (T6 , T7 ), (T2 , T7 ), (T5 , T6 ), (T1 , T3 )}. The table in
figure 9 traces the execution of the algorithm, where Li is the
execution time of the application at the completion of step i.
The subfigures (a) through (e) correspond to the application
graph at the completion of steps 0, 1, 2, 3 and 5, respectively.
The clusters are marked with polygons in dash line. Note that
in steps 6 and 7, the clustering is not performed due to the
task placement constraint.
During phase 2, we first calculate the energy dissipa-
tion for each cluster 190 energy units for cluster 1 =
{T1 , T2 , T7 }, 100 for the cluster 2 = {T3 , T4 , T6 }, and 50 for
cluster 3 = {T5 }. Since the remaining energy for the three
sensor nodes are the same, we simply assign 1 to PE1 , 2 to
PE2 , and 3 to PE3 .
Finally, we adjust the voltage levels of tasks. Since PE1 is
the critical node, we first set the voltage level of T2 to Vl ,
which reduces E1 to 106 and increases L from 80 to 219.
Next, we set the voltage level of T1 to Vl , which further de-
creases E1 to 92 and increases L to 242. After this step, the
critical node becomes PE2 with E2 = 0.1. Since the latency
constraint is 250, our heuristic terminates.
In the above example, we decreases the norm-energy of
the critical sensor node from 0.19 to 0.1, implying a system
lifetime improvement by a factor around 2.
6. Incorporating energy-latency tradeoffs for
communication activities
While DVS has been widely applied into various applications
for energy saving in computation activities, techniques for ex-
ploring the energy-latency tradeoffs of communication activ-
ities are gaining interest. An important observation [11] is
that in many channel coding schemes, the transmission en-
ergy can be significantly reduced by lowering the transmis-
sion power and increasing the duration of the transmission.
Techniques such as modulation scaling [23] have been pro-
posed for implementing such tradeoffs. Recently, algorithms
for applying such techniques in the context of packet trans-
missions or data gathering in wireless networks have been
studied in [11,23,32].
Our approaches can be extended to incorporate the above
tradeoffs. In the following, we discuss through the example
of modulation scaling that explores the tradeoffs by adapting
the modulation level to match the traffic load.
For ease of analysis, we focus on the Quadrature Am-
plitude Modulation (QAM) scheme [28]. The techniques
presented in this paper are extendible to other modulation
schemes as well. Given a communication activity Ei with
YU AND PRASANNA
a packet of si bits, assuming a fixed symbol rate Ri , the trans-
mission time can be calculated as [23]:
si
i = , (1)
bi Ri
where bi is the modulation level in terms of the constellation
size (number of bits per symbol). The corresponding energy
dissipation can be modeled as a function of i , denoted as
fi (i ). We have [23]
   
fi (i ) = Ci 2si /(i Ri ) 1 + Di i Ri , (2)
where Ci is determined by the quality of transmission (in
terms of bit error rate) and the noise power, and Di is a
device-dependent parameter that determines the power con-
sumption of the electronic circuitry of the sensor nodes. The
energy-latency tradeoffs for transmitting 1 bit is plotted in fig-
ure 10. The settings for Ci , Di and Ri are extracted from [23].
Also, we may estimate the energy dissipation for receiving the
packet as
fi (i ) = Di Ri i . (3)
In practice, the value of bi is typically set to positive even
integers, resulting in discrete values of i . For any communi-
cation activity Ei E, let tij denote the time cost with bi set
to the j th modulation level. Also, let eijs and eijr denote the
corresponding sending and receiving energy costs. We can
calculate the values of tij s, eijs s, and eijr s based on equa-
tions (1)(3).
To modify our ILP formulation, a set of 01 variables {uij }
are needed to indicate the modulation level of the communica-
tion activities. Specifically, uij equals one iff the modulation
level of Ei is set to the j th level. Moreover, we replace the
constraint set marked with * in figure 1 with the following
one, which states that the transmission time of Ei = (a, b)
depends on the modulation level for Ei and the locations of
Ta and Tb :
 
Ei = (a, b) E, (i) = (i) + uij tij (1 rab ).
j
Moreover, we change the constraint on the auxiliary vari-
able E in figure 3 as follows:
PEk



Ti T {xik j (yij eij )}
+
Rk




Ei =(a,b)E {xak (1 xbk ) j (uij eij ) + (1 xak )xbk
s r
j (uij eij )}
Rk

E.
For the 3-phase heuristic, we assume that both voltage and
modulation levels of the system are set to the highest options
in phases 1 and 2. We then slightly modify phase 3, such
that the energy savings achieved by lowering the modulation
levels of communication activities are also examined. The
modified pseudo code is shown in figure 11. One concern
ENERGY-BALANCED TASK ALLOCATION 123

(a) (b)

(c) (d)

Step i edge examined L if clustering clustering? Li

0 145
1 (T4 , T6 ) 135 yes 135
2 (T1 , T2 ) 120 yes 120
3 (T3 , T6 ) 100 yes 100
4 (T6 , T7 ) 100 no 100
5 (T2 , T7 ) 80 yes 80
6 (T5 , T6 ) no 80
7 (T1 , T3 ) no 80

(f)

(e)

Figure 9. Clustering steps for the application in figure 8.

is that to decrease the transmission energy at the sender, we energy does not cause the value of E to increase. By doing so,
actually increase the receiving energy at the receiver. Thus, our heuristic can handle the situation in highly dense WSNs,
in lines 13 and 14 of figure 11, we ensure that the modulation where the receiving energy is comparable with the sending
scaling is performed only when the increase in the reception energy.
124 YU AND PRASANNA

7. Simulation results

A simulator based on the system and application models pre-

sented in section 3 was developed to evaluate the performance
of our approach using application graphs from both a syn-
thetic approach and real world problems. The goals of our
simulations are (1) to measure and compare the performance
of the 3-phase heuristic against the ILP-based approach; and
(2) to evaluate the impact of the variations in several key sys-
tem parameters on the performance of the heuristic, includ-
ing the tightness of the latency constraint, the relative time
and energy costs of communication activities compared with
computation activities, and the number of voltage levels.
The evaluation metrics are based on the system lifetime
obtained by different approaches. Let LT ILP and LT heu de-
note the system lifetime obtained by the ILP-based approach
Figure 10. Energy-latency tradeoffs for transmitting one bit of data. and the 3-phase heuristic, respectively. In addition, let LT raw
denote the system lifetime obtained by assuming that no volt-
age or modulation scaling is available (i.e., every sensor node
runs and transmits data at the highest speed). Since we do not
1. For each PEi , sort EDi in a non-increasing order
have a stand alone approach to obtain LT raw , LT raw was cal-
2. Do culated based on the value of E obtained after phase 2 of the
3-phase heuristic.
3. i1
Unless otherwise stated, all the data presented in this
4. Let PEr denote the critical sensor node section is averaged over more than 100 instances so that a
5. While i
|EDr | Do 95% confidence interval with a 10% (or better) precision is
achieved.
6. Select the ith component in EDr ; let a denote the
corresponding activity
7.1. Synthetic application graphs
7. If L + tda > P , i = i + 1
Simulation setup. The structure of the application graph was
8. Else
generated using a method similar to the one described in [8].
9. L L + tda The only difference is that we enforce multiple source tasks
10. If a is a computatin activity in the generation of the DAG.
According to Rockwells WINS node [27], the power
11. Lower the voltage level of a to the next available consumption of an Intel StrongARM 1100 processor with
option 150 MIPS is around 200 mW. This implies that the time and
12. Else energy costs per instruction are around 5 nsec and 1 nJ. Also,
the power of the radio module used in WINS is 100 mW at
13. If to lower the modulation level of a to the next
available option does not increase E
100 Kbps, implying that the time and energy costs for trans-
mitting a bit are around 10 sec and 1 J. In the following,
14. Lower the modulation level of a to the next we set the parameters for our simulator such that the time and
available option energy costs for computation and communication activities
15. Else i i + 1 roughly follow the above data.
We set the maximum computation speed of each sensor
16. If any voltage or modulation scaling is performed
node to 102 Mcps (million cycles per second) and the mini-
17. Update ed a and tda ; resort EDr if necessary mum speed to 0.3 102 Mcps. It is assumed that other levels
18. Find the new critical sensor node, PEr  ; update E of computation speed are uniformly distributed between the
maximum and minimum speeds. The computation require-
19. If r = r  ments of the tasks followed a gamma distribution with a mean
20. r r ; i 1 value of 2 105 and a standard deviation of 105. The power
function of task Ti , gi (SP), was of the form ai (SP/108 )bi ,
21. L TraverseAssigned()
where ai and bi were random variables with uniform distrib-
22. Until E can not be reduced any more ution between 2 and 10, and 2 and 3 [20], respectively. For
example, suppose ai = bi = 2. Then, to execute a task of
Figure 11. Pseudo-code for the modified phase 3 that incorporates modulca- 2 105 instructions costs 2 msec and 4 mJ in the highest
tion scaling. speed, and 6.7 msec and 1 mJ in the lowest speed.
ENERGY-BALANCED TASK ALLOCATION
The time and energy costs of communication activities are
determined by the number of data units to transmit and the
values of and . Based on the data for WINS, we set
= 10 sec and = 1 J. To focus on the main issues,
we set the startup energy dissipation of the radio to be zero.
To study the effect of different communication load with re-
spect to the computation load, the number of bits per com-
munication activity follows a uniform distribution between
200 CCR (1 0.2), where CCR (communication to compu-
tation ratio) is a parameter indicating the ratio of the average
execution time of the communication activities to that of the
computation activities. Intuitively, a larger value of CCR im-
plies a relatively heavier communication loads compared with
the computation loads. Note that by varying CCR, we abstract
not only the variations in the amount of transmitted data, but
also the variations in the relative speed of computation and
communication devices. In our simulations, CCR was varied
within [0, 20].
The period of the application, P , was generated in the fol-
lowing way. We first define the distance of a node in the appli-
cation DAG as the number of edges in the longest path from
the source to the node. Nodes are then divided into layers,
with nodes in each layer having the same value of distance.
Since the average time to execute a task in the highest speed
is 2 msec, the computation time required for a layer is esti-
mated as 2 p/m
msec, where p is the number of tasks in tic achieved up to 63% of the solution obtained by the ILP-
the layer. By doing so, we implicitly assume full parallelism
in executing the tasks at each layer. In addition, the expected
number of communication activities initiated by a task is es-
timated as its out-degree subtracted by 1. Assuming there are
in total q communication activities requested by all the tasks
in a specific layer, the corresponding time cost is estimated
as 2 CCR q/K
msec. P is then set to the sum of the com-
putation and communication time cost of all layers over u,
where u [0, 1] is a parameter that approximates the overall
utilization of the system. The setting of u is important as it
determines the latency laxity for trading against energy. Intu-
(a)
Figure 12. Lifetime improvement of our approaches for samll scale problems (3 sensor nodes, 3 voltage levels, 2 channels, CCR = 1): (a) lifetime improve-
ment achieved by the ILP-based approach; (b) performance comparison of the ILP-based approach and the 3-phase heuristic.
125
itively, a larger value of u implies a tighter latency constraint
and hence less latency laxity.
The remaining energy of sensor nodes follows a uniform
distribution between Emean (1 0.3), where Emean is a fairly
large number.
Small scale problems. We first conducted simulations for
small scale problems, with 3 sensor nodes, 3 voltage levels,
2 channel, and 710 tasks. The number of source tasks in the
application graph is set to 2, while the maximal in-degree and
out-degree for each node are set to 3. A commercial software
package, LINDO [26], was used to solve the ILP problems.
Due to the large running time for solving some problem in-
stances, LINDO was interrupted after two hours of execution
if the optimal solution was not yet found. Then, the best so-
lution obtained so far was returned. We observed that in most
cases, LINDO was able to find the optimal solution within
two hours.
The data shown in figure 12 is averaged over more than
70 instances so that each data point has a 95% confidence in-
terval with a 10% precision. In figure 12(a), we illustrate the
lifetime improvement achieved by the ILP-based approach,
which is calculated as LT ILP /LT raw 1. We can see an im-
provement around 3x5x. Figure 12(b) shows the perfor-
mance ratio of the 3-phase heuristic over the ILP-based ap-
proach, i.e., LT heu /LT ILP . We can see that the 3-phase heuris-
based approach for the conducted simulations.
While the running time of the heuristic is around zero, the
average running time of the ILP-based approach ranges from
550 sec (n = 7, u = 0.5) to 5900 sec (n = 10, u = 0.8)
on a Sun Blade1000 machine with a UltraSparc III 750 Mhz
CPU.
Large scale problems. A set of simulations were conducted
for evaluating the performance of the 3-phase heuristic for
problems with 10 sensor nodes, 8 voltage levels, 4 channels,
60100 tasks, CCR [0, 20], and u [0, 1]. The number of
(b)
126 YU AND PRASANNA

(a) (b)
Figure 13. Lifetime improvement of the 3-phase heuristic for large scale problems (10 sensor nodes, 8 voltage levels, 4 channels, 60100 tasks): (a) lifetime
improvement vs. system utilization (u) and communication to computation ratio (CCR); (b) lifetime improvement vs. number of tasks (CCR = 4).

source tasks in the application graph is set to 6. The maximal

in-degree and out-degree for each node are set to 5. Due to the
large size of the problems, it is impractical to obtain the opti-
mal solutions by using the ILP-based approach. Thus, we use
the lifetime improvement achieved by the 3-phase heuristic as
the evaluation metric, which is calculated as LT heu /LT raw 1.
The simulation results are shown in figure 13.
An improvement up to 3.5x in the system lifetime can be
observed from figure 13(a). We can see that the improvement
increases when u decreases, as the latency laxity increases
accordingly. The lifetime improvement saturates when u ap-
proaches 0, i.e., the latency constraint approaches . The
curve with u = 0.0 gives the upper bound of the improve-
ment that can be achieved by our heuristic with respect to
variations in CCR. Figure 14. Miss rate of the 3-phase heuristic (10 sensor nodes, 8 voltage
levels, 4 channels, 60 tasks, CCR = 0).
The effect of CCR is more complicated. For example,
when u = 0.5, the lifetime improvement increases when
CCR
6 and decreases when CCR is beyond 6. This the total number of instances) of a heuristic is another key
is because when CCR is small, the computation activities issue. Note that in our simulations, not all instances are guar-
dominate the overall energy costs of the application. By anteed to have feasible solutions. We observed that the miss
increasing CCR, we actually increase the latency constraint rate of the 3-phase heuristic is significant only when CCR is
without increasing the computation load, which in turn can close to zero. Thus, we show the miss rate with CCR = 0
be traded for lifetime improvement. However, when CCR in figure 14. Also, the running time of the heuristic is around
reaches some threshold value, the communication energy cost 0.5 msec on a Sun Blade1000 machine with a UltraSparc III
becomes more significant than that of the computation activ- 750 Mhz CPU.
ities. Thus, the lifetime improvement achieved by reducing
computation energy becomes limited. We shall see later that Impact of the number of voltage levels. We also studied
this shortcoming can be overcome by incorporating modula- the impact of the variations in the number of voltage levels.
tion scaling into our heuristic. Simulations were conducted with 10 sensor nodes, 60 tasks,
Figure 13(b) shows the lifetime improvement with number 4 channels, CCR = 2, u {0.2, 0.5, 0.8, 1.0} and 1 to 10
of tasks, n varying from 60 to 100. We can see that the per- voltage levels. The results are demonstrated in figure 15.
formance of our approach is quite stable with respect to the The plots show that when u > 0.2, the performance of
variation in n. the heuristic can be significantly improved by increasing the
The miss rate (defined as the ratio of the number of in- number of voltage levels from 1 to 4. Further increase in the
stances that an approach fails to find a feasible solution to number of voltage levels does not improve the performance
ENERGY-BALANCED TASK ALLOCATION
much. This is understandable since the energy behaves as a
monotonically increasing and strictly convex function of the
computation speed. The first derivative of the energy function
tends to when the speed tends to . Thus, the most portion
of energy saving is obtained by changing the speed from the
highest option to some lower options, which can be efficiently
achieved with 4 voltage levels per sensor node.
When u = 0.2, the latency laxity is so large that the volt-
age level of most tasks can be set to the lowest option. Thus,
there is almost no improvement by increasing the number of
voltage levels beyond 2.
Incorporating modulation scaling. We used modulation
scaling to illustrate the energy-latency tradeoffs for commu-
nication activities. We assume that all sensor nodes have the
identical settings for parameters Ci , Di and Ri . From [23],
we set Di = 107 [23]. To investigate the impact of different
energy/time ratio for data transmission, we set Ci to 107 and
106 for different instances. The modulation level, bi , was set
Figure 15. Impact of variation in number of voltage levels (10 sensor nodes,
4 channels, 60 tasks, CCR = 2).
(a)
Figure 16. Lifetime improvement of the 3-phase heuristic incorporated with modulation scaling (10 sensor nodes, 8 voltage levels, 4 channels, 3 modulation
levels, 60 tasks): (a) small energy/time ratio for communication activities (Ci = 107 ); (b) large energy/time ratio for communication activities (Ci = 106 ).
127
to even numbers between 2 and 6. We set Ri = 1.7 104 so
that when bi = 6, it roughly takes 10 sec and 1 J to trans-
mit a bit (as shown in figure 10).
The simulations were conducted with 10 sensor nodes,
8 voltage levels, 3 modulation levels ({2, 4, 6}), 60 tasks,
u {0.0, 0.2, 0.5, 0.8, 1.0}, and CCR [0, 20]. Com-
pared with figure 13, we can observe a significant amount of
performance improvement in figure 16. For example, when
u = 0.5, the highest lifetime improvement increases from
3x in figure 13(a) to 6x in figure 16(a) and even 10x in fig-
ure 16(b). The difference in performance improvement of fig-
ures 16(a) and (b) is because that a larger Ci leads to larger
energy/time ratio of communication activities, which in turn
gives more advantage in reducing the communication energy
by utilizing modulation scaling.
Similar to figure 13, larger improvement is observed when
u becomes smaller. In addition, the miss rate of the heuristic
exhibits a similar trend as the cases with DVS only.
7.2. Application graphs from real world problems
In addition to synthetic application graphs, we also consid-
ered application graphs of two real world problems: LU fac-
torization algorithm [5] and Fast Fourier Transformation [7].
These two algorithms are widely used as kernel operations for
various signal processing, such as beamforming [21].
LU factorization. Figure 17(a) gives the sequential program
for the LU factorization without pivoting, where s denotes
the dimension of the matrix. The application graph of the
algorithm for the special case of s = 5 is given in figure 17(b).
Each Tk,k represents a pivot column operation and each Tk,j
represents an update operation. The total number of tasks in
the application graph equals (s 2 + s 2)/2. Also, we assume
the input matrix is available at the sensor node where task T1,1
is assigned.
We performed simulations with 10 sensor nodes, 8 voltage
levels, 4 channels, 3 modulation levels, and the matrix di-
(b)
128 YU AND PRASANNA

MatrixFactorization(a) mension, s, varying from 5 to 20. Regarding the energy/time

1. For k = 1 to s 1 Do ratio for data transmission, we set Di = 106 . It is easy to
2. For i = k + 1 to s Do // Tk,k verify that the computation requirement of any task Tk,j is
3. aik = aik /akk s k ALU operations. Further, for any task, Tk,j , the size
4. For j = k + 1 to s Do of data transmitted by any communication activity to the task
5. For i = k + 1 to s Do // Tk,j is s k units in the matrix. We examined two cases with u
6. aij = aij aik /akj set to 0.5 and 0.8. In both cases, CCR was selected from
{1.0, 3.0, 5.0, 8.0, 10.0}.
(a) The lifetime improvement achieved by our 3-phase heuris-
tic for the LU factorization algorithm is shown in figure 18.
It can be observed that the performance of the heuristic im-
proves when CCR increases or u decreases. The lifetime im-
provement approaches 8x when CCR = 10.0. Also, very
few improvement was observed during our simulations by
setting CCR beyond 10.0. The least amount of lifetime im-
provement is around 15% when u = 0.8, CCR = 1.0, and
s = 20.

Fast Fourier Transformation (FFT). The recursive, one-

(b)
Figure 17. Matrix factorization algorithm: (a) sequential algorithm; (b) ex-
ample application graph with a 4 4 matrix.
dimensional FFT algorithm is given in figure 19(a). In the
figure, A is an array of length l which holds the coefficients
of the polynomial and array Y is the output of the algorithm.
The algorithm consists of two parts: recursive calls (lines 3, 4)
and the butterfly operation (lines 6, 7). For an input vector of
size l, there are 2 l 1 recursive call tasks and l log l but-
terfly operation tasks (we shall be assuming l = 2k for some
integer k). For example, the application graph with four data
points is given in figure 19(b). The 7 tasks above the dashed
line are the recursive call tasks, while the 8 tasks below the
line are butterfly operation tasks.
We performed simulations used 10 sensor nodes, 8 volt-
age levels, 4 channels, 3 modulation levels. Regarding the
energy/time ratio for data transmission, we set Di = 106 .
The vector size was varied from 4 to 64 incrementing by the
power of 2. We also examined two cases with u set to 0.5 and
0.8. In both cases, CCR was selected from {1.0, 3.0, 5.0, 8.0}.

(a) (b)
Figure 18. Lifetime improvement for the matrix factorization algorithm (10 sensor nodes, 8 voltage levels, 4 channels, 3 modulation levels): (a) u = 0.5;
(b) u = 0.8.
ENERGY-BALANCED TASK ALLOCATION 129

FFT(A, w) The lifetime improvement achieved by our 3-phase heuris-

1. Set l = length(A) tic for the FFT algorithm is shown in figure 20. Again, the
2. If l = 1, return A performance of the heuristic improves when CCR increases or
3. Y (0) = FFT((A[0], A[0], . . . , A[l 2]), w2 ) u decreases. The lifetime improvement is close to 10x when
4. Y (1) = FFT((A[1], A[3], . . . , A[l 1]), w2 ) CCR = 8.0 and l = 64. The least amount of lifetime im-
5. For i = 0 to l/2 Do provement is around 75% when u = 0.8, CCR = 1.0, and
6. Y [i] = Y (0)[i] + wi Y (1)[i] l = 4.
Note that the above two example applications have exactly
7. Y [i + l/2] = Y (0) [i] wi Y (1) [i]
one source task that initially holds the entire data set, im-
8. Return Y
plying that data dissemination within the cluster is required.
However, our technique is also applicable to applications
(a) where data are locally sensed or gathered at each individual
sensor node. For example, in figure 19(b), input data can be
generated by tasks T4T7 through local sensing. Thus, the
recursive calls above the dashed line to disseminate the data
become unnecessary.

8. Concluding remarks

In this paper, we have investigated the problem of allocating

(b)
Figure 19. Fast Fourier Transformation (FFT) algorithm: (a) sequential algo-
rithm; (b) example application graph with 4 points.
an epoch-based real-time application to a single-hop cluster of
homogeneous sensor nodes with multiple wireless channels.
A new performance metric has been proposed to balance the
energy dissipation among all the sensor nodes. We have pre-
sented both an ILP formulation and a polynomial time heuris-
tic. Also, we have incorporated techniques that explore the
energy-latency tradeoffs of communication activities.
We have demonstrated through simulations that for small
scale problems, a lifetime improvement up to 5x is achieved
by the ILP-based approach, compared with the case where no
DVS is used. Also, the performance of the 3-phase heuris-
tic achieves up to 63% of the system lifetime obtained by the
ILP-based approach. For large scale problems, a lifetime im-
provements up to 10x was observed when both voltage and
modulation scaling were used. Simulations were also con-
ducted for application graphs from LU factorization and FFT.
The 3-phase heuristic achieves a lifetime improvement of up

(a) (b)
Figure 20. Lifetime improvement for the FFT algorithm (10 sensor nodes, 8 voltage levels, 4 channels, 3 modulation levels): (a) u = 0.5; (b) u = 0.8.
130
to 8x for the LU factorization algorithm and an improvement
of up to 9x for the FFT algorithm.
In the future, we would like to validate our approaches us-
ing real systems. We are particularly interested in advanced
applications for WSNs, where systematic methodologies for
task allocation are mostly required for rapid and automated
system design.
References
[1] G. Asada, M. Dong, T.S. Lin, F. Newberg, G. Pottie and W.J. Kaiser,
Wireless integrated network sensor: Low power systems on a chip, in:
ESSCIRC 98 (1998).
[2] H. Aydin, R. Melhem, D. Moss and P.M. Alvarez, Determining opti-
mal processor speeds for periodic real-time tasks with different power
characteristics, in: 13th Euromicro Conf. on Real-Time Systems (June
2001).
[3] A. Bakshi, J. Ou and V.K. Prasanna, Towards automatic synthesis of
a class of application-specific sensor networks, in: Internat. Conf. on
Compilers, Architecture, and Synthesis for Embedded Systems (CASES)
(October 2002).
[4] T.D. Burd, T.A. Pering, A.J. Stratakos and R.W. Brodersen, A dynamic
voltage scaled microprocessor system, IEEE Journal of Solid-State Cir-
cuits 35(11) (2000) 15711580.
[5] M. Conard, M. Marrakchi, Y. Robert and D. Trystram, Parallel
Gaussian elimination on an MIMD computer, Parallel Computing 6
(1988) 275295.
[6] S. Conner, L. Krishnamurthy and R. Want, Making everyday life easier
using dense sensor networks, in: ACM UBICOMP (2001).
[7] T.H. Cormen, C.E. Leiserson and R.L. Rivest, Introduction to Algo-
rithms (MIT Press, Cambridge, MA, 1990).
[8] R.P. Dick, D.L. Rhodes and W. Wolf, TGFF: Task graphs for free,
in: Internat. Workshop on Hardware/Software Codesign (March 1998)
pp. 97101.
[9] J. Elson, L. Girod and D. Estrin, Fine-grained network time synchro-
nization using reference broadcasts, in: Symposium on Operating Sys-
tems Design and Implementation (OSDI) (December 2002).
[10] D. Estrin, L. Girod, G. Pottie and M.B. Srivastava, Instrumenting the
world with wireless sensor networks, in: Internat. Conf. on Acoustics,
Speech and Signal Processing (ICASSP) (May 2001).
[11] A.E. Gamal, C. Nair, B. Prabhakar, E. Uysal-Biyikoglu and S. Zahedi,
Energy-efficient scheduling of packet transmissions over wireless net-
works, in: IEEE INFOCOM (2002).
[12] F. Gruian and K. Kuchcinski, LEneS: Task scheduling for low-energy
systems using variable supply voltage processors, in: Design Automa-
tion Conf. (DAC) (2001) pp. 449455.
[13] W. Heinzelman, A.P. Chandrakasan and H. Balakrishnan, An applica-
tion specific protocol architecture for wireless microsensor networks,
IEEE Transactions on Wireless Networking (2002) 660670.
[14] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler and K. Pister, System
architecture directions for networked sensors, in: 9th Internat. Conf.
on Architectural Support for Programming Languages and Operating
Systems (2000).
[15] I. Hong, G. Qu, M. Potkonjak and M.B. Srivastava, Synthesis tech-
niques for low-power hard real-time systems on variable voltage
processors, in: IEEE Real-Time Systems Symposium (RTSS) (Decem-
ber 1998).
[16] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion:
A scalable and robust communication paradigm for sensor networks,
in: ACM/IEEE Internat. Conf. on Mobile Computing and Networking
(MOBICOM) (2000).
[17] J. Luo and N.K. Jha, Static and dynamic variable voltage scheduling
algorithms for real-time heterogeneous distributed embedded systems,
in: VLSI Design (January 2002).
YU AND PRASANNA
[18] S.R. Madden, M.J. Franklin, J.M. Hellerstein and W. Hong, TAG: a
Tiny AGgregation service for ad-hoc sensor networks, in: Symposium
on Operating Systems Design and Implementation (OSDI) (December
2002).
[19] C. Meesookho, S. Narayanan and C.S. Raghavendra, Collaborative
classification applications in sensor networks, in: 2nd IEEE Sensor Ar-
ray and Multichannel Signal Processing Workshop (August 2002).
[20] P. Meja-Alvarez, E. Levner and D. Moss, An integrated heuristic ap-
proach to power-aware real-time scheduling, in: Workshop on Power-
Aware Computer Systems (February 2002).
[21] R.A. Mucci, A comparison of efficient beamforming algorithms, IEEE
Transactions on Acoustic, Speech, Signal Processing 22 (1984) 548
558.
[22] V. Sarkar, Partitioning and Scheduling Programs for Execution on Mul-
tiprocessors (MIT Press, Cambridge, MA, 1989).
[23] C. Schurgers, O. Aberhorne and M.B. Srivastava, Modulation scaling
for energy-aware communication systems, in: ISLPED (2001) pp. 96
99.
[24] Y. Shin, K. Choi and T. Sakurai, Power optimization of real-time em-
bedded systems on variable speed processors, in: IEEE/ACM Internat.
Conf. on Computer-Aided Design (2000) pp. 365368.
[25] M. Singh and V.K. Prasanna, A hierarchical model for distributed col-
laborative computation in wirelss sensor networks, in: 5th Workshop
on Advances in Parallel and Distributed Computational Models (April
2003).
[26] The LINDO System Inc., http://www.lindo.com.
[27] The WINS Project, Rockwell Science Center, http://wins.rsc.
rockwell.com.
[28] T. Ue, S. Sampei, N. Morinaga and K. Hamaguchi, Symbol rate and
modulation level-controlled adaptive modulation/TDMA/TDD system
for high-bit rate wireless data transmission, IEEE Transactions on Ve-
hicular Technology 47(4) (1998) 11341147.
[29] H.P. Williams, Model Building in Mathematical Programming (Wiley,
New York, 1999).
[30] F. Yao, A. Demers and S. Shenker, A scheduling model for reduced
CPU energy, IEEE Annual Foundations of Computer Science (1995)
374382.
[31] Y. Yu, B. Krishnamachari and V.K. Prasanna, Issues in designing mid-
dleware for wireless sensor networks, IEEE Network Magazine (Spe-
cial Issue on Middleware Technologies for Future Communication Net-
works) 18(1) (2004) 1521.
[32] Y. Yu, B. Krishnamachari and V.K. Prasanna, Energy-latency tradeoffs
for data gathering in wireless sensor networks, in: Proceedings of IEEE
INFOCOM (March 2004).
[33] Y. Zhang, X. Hu and D.Z. Chen, Task scheduling and voltage selection
for energy minimization, in: Design Automation Conf. (DAC) (2002).
[34] D. Zhu, R. Melhem and B. Childers, Scheduling with dynamic volt-
age/speed adjustment using slack reclamation in multi-processor real-
time systems, in: IEEE Real-Time Systems Symposium (RTSS) (Decem-
ber 2001).
Yang Yu is a Ph.D. degree candidate in the Depart-
ment of Electrical Engineering at the University of
Southern California (USC). He received both BS and
MS degrees in computer science from Shanghai Jiao-
Tong University in China. His research interests in-
clude energy-aware resource management for wire-
less sensor networks, especially in algorithmic so-
lutions for energy-aware communication scheduling
and task allocation. He is a student member of the
IEEE.
E-mail: yangyu@usc.edu
ENERGY-BALANCED TASK ALLOCATION
Viktor K. Prasanna is a Professor of Electrical En-
gineering and Computer Science at the University
of Southern California (USC). He is also a mem-
ber of the NSF supported Integrated Media Sys-
tems Center (IMSC) and an associate member of the
Center for Applied Mathematical Sciences (CAMS)
at USC. His research interests include high perfor-
mance computing, parallel and distributed systems,
network computing, and embedded systems. He re-
ceived his BS in electronics engineering from Ban-
galore University, MS from the School of Automation, Indian Institute of
Science, and Ph.D. in computer science from the Pennsylvania State Univer-
131
sity. He has published extensively and consulted for industries in the above
areas. He is the steering committee co-chair of the International Parallel
and Distributed Processing Symposium (IPDPS) (merged IEEE International
Parallel Processing Symposium (IPPS) and Symposium on Parallel and Dis-
tributed Processing (SPDP)). He is the steering committee chair of the In-
ternational Conference on High Performance Computing (HiPC). He serves
on the editorial boards of the Journal of Parallel and Distributed Computing.
He is the editor-in-chief of the IEEE Transactions on Computers. He was
the founding chair of the IEEE Computer Societys Technical Committee on
Parallel Processing. He is a fellow of the IEEE.
E-mail: prasanna@usc.edu
 Mobile Networks and Applications 10, 133149, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Efficient and Robust Protocols for Local Detection and

Propagation in Smart Dust Networks ,
IOANNIS CHATZIGIANNAKIS, SOTIRIS NIKOLETSEAS and PAUL SPIRAKIS
Computer Technology Institute (CTI) and Patras University, P.O. Box 1122, 261 10 Patras, Greece

Abstract. Smart Dust is a set of a vast number of ultra-small fully autonomous computing and communication devices, with very restricted
energy and computing capabilities, that co-operate to quickly and efficiently accomplish a large sensing task. Smart Dust can be very useful
in practice, i.e., in the local detection of a remote crucial event and the propagation of data reporting its realization. In this work we make
an effort towards the research on smart dust from an algorithmic point of view. We first provide a simple but realistic model for smart dust
and present an interesting problem, which is how to propagate efficiently information on an event detected locally. Then we present various
smart dust protocols for local detection and propagation that are simple enough to be implemented on real smart dust systems, and perform,
under some simplifying assumptions, a rigorous average case analysis of their efficiency and energy consumption (and their interplay). This
analysis leads to concrete results showing that our protocols are very efficient and robust. We also validate the analytical results by extensive
experiments.
Keywords: wireless sensor networks, algorithms, data propagation, stochastic processes, simulation

1. Introduction
Networked sensors (or Smart Dust) are very large systems,
comprised of a vast number of homogenous ultra-small fully
autonomous computing and communication devices that co-
operate to achieve a large sensing task. Each device has one
or more sensors, embedded processors and low-power radios,
and is normally battery operated. Examining each such sin-
gle device individually, might appear to have small utility.
The realization of Smart Dust, however, lies in using and co-
coordinating a vast number of such devices.
Smart Dust is a useful case of dynamic environments of
networked sensors that are spread over a global system and
try to communicate and compute efficiently and quickly, hav-
ing only partial knowledge of the global conditions and hav-
ing poor energy and computing resources. Typically, these
networked sensors coordinate to perform a common task. De-
signing protocols to coordinate such systems (i.e., create a dy-
namic and efficient network of these sensors) and monitoring
their behavior as they operate in complex and dynamic global
environments is of great importance for information gathering
and processing in many practical situations.
As an example, [11] points that integrated low-power sens-
ing devices will permit remote object monitoring and track-
ing in inhospitable physical environments such as remote
geographic regions or toxic urban locations. They will also
enable low maintenance sensing in the field (vehicles, equip-
ment, personnel), the office buildings (projectors, furniture,
Preliminary versions of this work have appeared in the 2nd ACM Work-
shop on Principles of Mobile Computing (POMC, 2002 [9]) and the 3rd
Workshop on Mobile and Ad-hoc Networks (WMAN, 2003 [6]).
This work has been partially supported by the IST Program of the Eu-
ropean Union under contract numbers IST-1999-14186 (ALCOM-FT) and
IST-2001-33116 (FLAGS).
books, people), the factory floor (motors, small robotic de-
vices).
There are many possible models for such networked sen-
sors. In this work, we consider networked sensors where
(a) all nodes in the network are homogenous and constrained
by low availability of resources (energy, communication) and
(b) the data being sensed by the nodes must be transmitted
to a fixed control center located far away from the sensors.
Thus direct communication between the sensor nodes and the
control center is impossible and/or expensive, since there are
no high-energy nodes through which communication can
proceed. This is the general framework for MITs AMPS
project [20], which focuses on innovative energy-optimized
solutions at all levels of the system hierarchy, from the phys-
ical layer and communication protocols up to the application
layer.
To motivate the challenges in designing such sensor net-
works, we can consider the following scenario where local
detection and fast propagation to the authorities of the real-
ization of a crucial event can be achieved using smart dust.
Think about thousand of disposable sensors scattered (e.g.,
thrown from an aircraft) over a forest. Each of these sensors
can monitor the temperature at a single, very small geograph-
ical area. The sensors coordinate to establish an efficient, dy-
namic and short-lived communication network, dividing the
task of monitoring the terrain and offering continuous moni-
toring of the environment in order to alert the authorities as
soon as possible after a forest fire is detected by some sensor.
Several aspects of such systems of autonomous networked
entities emerge, which are quite different from those posed
by standard computer networks. Such aspects include the
very poor and highly restricted resources (e.g., very low bat-
tery power, low computing capabilities, total absence of syn-
chrony and anonymity). Network protocols must be designed
134
to achieve fault tolerance in the presence of individual node
failure while minimizing energy consumption. Another im-
portant aspect is the scalability to the change in network size,
node density and topology. The network topology changes
over time as some nodes may die, or possibly because new
nodes join later.
This work, continuing our line of research on communica-
tion in ad-hoc mobile networks [7,8], is an attempt towards
capturing the underlying foundational and algorithmic issues
in the design of such systems, abstracting accurately enough
the real technological specifications involved and providing
some first concrete results for the efficiency of a variety of
smart dust protocols using an average case analysis. We focus
in this paper on the efficient use of smart dust in local detec-
tion and propagation protocols. We first provide an abstract
model for smart dust systems which is simple enough to allow
an analysis to develop, being however at the same time quite
realistic, in terms of the technological specifications of real
smart dust systems it captures. Then we define the problem
of local detection and propagation using smart dust and also
propose some concrete performance and robustness measures
for the average case analysis of protocols for this problem.
Our results
For the local detection and propagation problem using smart
dust, we provide three protocols. All protocols are simple
enough to be implemented in real smart dust systems despite
the severe energy and computing power limitations of such
systems. Furthermore, we give a rigorous average case analy-
sis for the efficiency of these protocols. We consider a variety
of performance and robustness criteria, such as propagation
time, number of particle to particle transmissions (which also
characterizes energy consumption and time efficiency, assum-
ing an efficient MAC protocol) and fault-tolerance:
1. Our first protocol, which we call the local target protocol
(LTP), uses a fast and cheap search phase which is assumed
to always return a nearby particle towards the authorities,
uniformly in some range. We show that LTP is efficient,
in the sense that it achieves a propagation time and an en-
ergy consumption whose expected ratio over the optimal
solutions is at most /2 1.57.
2. Our second protocol, the min-two uniform targets proto-
col (m2TP), applies the simple idea of getting at least two
particles towards the authorities and selecting the best in
terms of propagation progress. It is, in fact, an optimized
and more efficient version of the local target protocol, and
has an expected time and energy ratio over the optimal so-
lutions which is at most 2 /8 1.24.
3. Next we provide tight upper bounds to the distribution of
the number of particle to particle data transmissions (and
thus the efficiency) of a generalized target protocol.
4. We propose a new protocol which we call the Sleep
Awake protocol (SWP), that explicitly uses the energy
saving characteristics, such as the alteration of sleep and
awake time periods, of smart dust particles. By using both
CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
analytic and extensive experimental means, we investigate
the relation between (a) the success probability and (b) the
time efficiency of the protocol, to the maximum sleeping
time period, for various values of other parameters, such
as particle density, particle distribution and angle . We
interestingly note that the new protocol is efficient, despite
the fact that the particles are allowed to enter a sleeping
mode in order to save energy.
All protocols mentioned above are shown to be robust in
the following sense: (a) the protocols use the search and the
backtrack phases to explore the active (non-faulty) next par-
ticles. Thus, the fact (demonstrated both by analysis and sim-
ulation) that the protocols succeed with high probability, ex-
hibits also fault-tolerance properties of the protocols; (b) our
findings showing that the protocols succeed even in the case
of low densities also implies robustness.
Discussion of selected related work
In the last few years, Sensor Networks have attracted a lot of
attention from researchers at all levels of the system hierarchy,
from the physical layer and communication protocols up to
the application layer.
At the MAC level, many researchers have done research
work in an effort to minimize the power consumption. [27]
presents a contention-based protocol that tries to minimize en-
ergy consumption due to node idle listening, by avoiding the
overhearing among neighboring nodes. A recent work [30]
exploits a similar method for energy savings, and further re-
duce idle listening by avoiding any use of out-of-channel sig-
naling. Additionally, their protocol trades off per-node fair-
ness for further energy savings.
For establishing communication and routing information
to the control center, mobile ad-hoc routing protocols [24]
may be used in sensor networks. However, although proto-
cols for mobile ad-hoc networks take into consideration en-
ergy conservation issues, most of them are not really suitable
for sensor networks. [19] presents a routing protocol suitable
for sensor networks that makes greedy forwarding decisions
using only information about a nodes immediate neighbors
in the network topology. This approach achieves high scala-
bility as the density of the network increases. [14] presents a
clustering-based protocol that utilizes randomized rotation of
local cluster heads to evenly distribute the energy load among
the sensors in the network. [21] introduces a new energy ef-
ficient routing protocol that does not provide periodic data
monitoring (as in [14]), but instead nodes transmit data only
when sudden and drastic changes are sensed by the nodes. As
such, this protocol is well suited for time critical applications
and compared to [14] achieves less energy consumption and
response time.
A family of negotiation-based information dissemination
protocols suitable for wireless sensor networks is presented
in [15]. Sensor Protocols for Information via Negotiation
(SPIN) focus on the efficient dissemination of individual sen-
sor observations to all the sensors in a network. However,
in contrast to classic flooding, in SPIN sensors negotiate with
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST
each other about the data they possess using meta-data names.
These negotiations ensure that nodes only transmit data when
necessary, reducing the energy consumption for useless trans-
missions.
A data dissemination paradigm called directed diffusion
for sensor networks is presented in [18], where data-generated
by sensor nodes is named by attributevalue pairs. An ob-
server requests data by sending interests for named data; data
matching the interest is then drawn down towards that node
by selecting a single path or through multiple paths by us-
ing a low-latency tree. [17] presents an alternative approach
that constructs a greedy incremental tree that is more energy-
efficient and improves path sharing.
We note that, as opposed to the work presented in this
paper, the above research focuses on energy consumption
without examining the time efficiency of their protocols. Fur-
thermore, these works contain basically protocol design and
technical specifications, while quantitative aspects are only
experimentally evaluated and no theoretical analysis is given.
Note also that our protocols are quite general in the sense
that (a) do not assume global network topology informa-
tion, (b) do not assume geolocation information (such as GPS
information) and (c) use very limited control message ex-
changes, thus having low communication overhead.
Finally, our third protocol is using a similar approach to
the recent work of [26], where a new technique called Sparse
Topology and Energy Management (STEM) is proposed that
aggressively puts nodes to sleep. Interestingly, the analysis
and experiments of STEM show improvements of nearly two
orders of magnitude compared to sensor networks without
topology management.
Some recent work
In [4] the authors present a new protocol for data propaga-
tion that avoids flooding by probabilistically favoring certain
(close to optimal) data transmissions. As shown by a geom-
etry analysis, the protocol is correct, since it always prop-
agates data to the sink, under ideal network conditions (no
failures). Using stochastic processes, they show that the pro-
tocol is very energy efficient. Also, when part of the network
is inoperative, the protocol manages to propagate data very
close to the sink, thus in this sense it is robust. They finally
present and discuss large-scale experimental findings validat-
ing the analytical results.
In [5], the authors have implemented and experimentally
evaluated two variations of LTP, under new, more general and
realistic modelling assumptions. They comparatively study
LTP to PFR, by using extensive experiments, highlighting
their relative advantages and disadvantages. All protocols are
very successful. In the setting considered there, PFR seems
to be faster while the LTP based protocols are more energy
efficient.
In [12], Euthimiou et al. study the problem of energy-
balanced data propagation in wireless sensor networks. The
energy balance property guarantees that the average per sen-
sor energy dissipation is the same for all sensors in the net-
135
work, during the entire execution of the data propagation pro-
tocol. This property is important since it prolongs the net-
works lifetime by avoiding early energy depletion of sen-
sors. They propose a new algorithm that in each step decides
whether to propagate data one-hop towards the final destina-
tion (the sink), or to send data directly to the sink. This ran-
domized choice balances the (cheap) one-hop transimssions
with the direct transimissions to the sink, which are more ex-
pensive but bypass the sensors lying close to the sink. Note
that, in most protocols, these close to the sink sensors tend to
be overused and die out early.
In [1], the authors propose a new energy efficient and fault
tolerant protocol for data propagation in smart dust networks,
the Variable Transmission Range Protocol (VTRP). The basic
idea of data propagation in VTRP is the varying range of data
transmissions, i.e., they allow the transmission range to in-
crease in various ways. Thus data propagation in the protocol
exhibits high fault-tolerance (by bypassing obstacles or faulty
sensors) and increases network lifetime (since critical sensors,
i.e., close to the control center are not overused). They im-
plement the protocol and perform an extensive experimental
evaluation and comparison to a representative protocol (LTP)
of several important performance measures with a focus on
energy consumption. The findings indeed demonstrate that
the protocol achieves significant improvements in energy ef-
ficiency and network lifetime.
In [23], Nikoletseas et al. (a) propose extended versions
of two data propagation protocols: the SleepAwake Prob-
abilistic Forwarding Protocol (SW-PFR) and the Hierarchi-
cal Threshold sensitive Energy Efficient Network protocol
(H-TEEN). These non-trivial extensions aim at improving the
performance of the original protocols, by introducing sleep
awake periods in the PFR protocol to save energy, and in-
troducing a hierarchy of clustering in the TEEN protocol to
better cope with large networks areas; (b) they have imple-
mented the two protocols and performed an extensive exper-
imental comparison (using simulation) of various important
measures of their performance with a focus on energy con-
sumption; (c) they investigate in detail the relative advantages
and disadvantages of each protocol and discuss and explain
their behavior; (d) in the light above they propose and discuss
a possible hybrid combination of the two protocols towards
optimizing certain goals. Efficient collision avoidance proto-
cols, particularly useful for multipath data propagation, have
been proposed in [10].
A brief description of the technical specifications of state-
of-the-art sensor devices, a discussion of possible models
used to abstract such networks and a presentation of some
characteristic protocols for data propagation in sensor net-
works, along with an evaluation of their performance analy-
sis, can be found in the recent book chapter of Boukerche and
Nikoletseas [3].
2. The model
Smart dust is comprised of a vast number of ultra-small ho-
mogenous sensors, which we call grain particles. Each
136
Figure 1. A smart dust cloud.
smart-dust grain particle is a fully-autonomous computing
and communication device, characterized mainly by its avail-
able power supply (battery) and the energy cost of computa-
tion and transmission of data. Such particles cannot move.
Each particle is equipped with a set of monitors (sensors)
for light, pressure, humidity, temperature, etc. Each particle
has two communication modes: a broadcast (digital radio)
beacon mode (for low energy short signals) and a directed
to a point actual data transmission mode (usually via a laser
beam). Also, in a variation of our model capturing energy sav-
ing specifications, each particle may alternate (independently
of other particles) between a sleeping and an awake mode.
During sleeping periods grain particles cease any communi-
cation with the environment, thus they are unable to listen,
receive and propagate data transmitted by other particles. De-
pending on the specific application the sensing part may cease
or not during the sleeping mode. In the case where sensing
is not ceased during sleeping mode, detection of the crucial
event wakes the particle up.
We adopt here (as a starting point) a two-dimensional
(plane) framework: a smart dust cloud (a set of particles)
is spread in an area (for a graphical presentation, see fig-
ure 1). Note that a two-dimensional setting is also used in
[14,15,17,18,21].
Definition 2.1. Let d (usually measured in numbers of par-
ticles/m2 ) be the density of particles in the area. Let R be
the maximum (beacon/laser) transmission range of each grain
particle. A receiving wall W is defined to be an infinite line in
the smart-dust plane. Any particle transmission within range
R from the wall W is received by W.
We assume that W has very strong computing power, able
to collect and analyze received data and has a constant power
supply and so has no energy constraints. The wall represents,
in fact, the authorities (the fixed control center) who the real-
ization of a crucial event should be reported to. Note that a
wall of appropriately big (finite) length suffices. We plan to
conduct an analysis of the (expected and/or with high proba-
bility) deviation of the transmitted data from the vertical to the
CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
wall position in order to provide upper bounds on the walls
length needed.
Furthermore, we assume that there is a set-up phase of the
smart dust network, during which the smart cloud is dropped
in the terrain of interest, when using special control messages
(which are very short, cheap and transmitted only once) each
smart dust particle is provided with the direction of W. By as-
suming that each smart-dust particle has individually a sense
of direction (e.g., through its magnetometer sensor), and using
these control messages, each particle is aware of the general
location of W.
We feel that our model, although simple, depicts accu-
rately enough the technological specifications of real smart
dust systems. Similar models are being used by other re-
searchers in order to study sensor networks (see [14,21]). In
contrast to [18,19], our model is weaker in the sense that no
geolocation abilities are assumed (e.g., a GPS device) for the
smart dust particles leading to more generic and thus stronger
results. In [16] a thorough comparative study and description
of smart dust systems is given, from the technological point
of view. In the following section we report some basic techni-
cal characteristics which we took into account when defining
the model of smart dust we use here.
3. Technological specifications of smart dust devices
New technology is changing the nature of sensors and the way
they interface with data acquisition and control systems. Re-
searchers have developed an open-source hardware and soft-
ware platform that combines sensing, communications, and
computing into a complete architecture. The first commer-
cial generation of this platform was dubbed the Rene Mote,
and several thousand of these sensors have been deployed at
commercial and research institutions worldwide to promote
the development and application of wireless sensor networks.
The platforms development community is based on the
open-source model, which has become well known with the
increasingly popular Linux operating system. Most develop-
ment work is done in the public domain, and it includes the
hardware design and software source code. Users of the tech-
nology contribute their developments back to the community
so that the base of code and hardware design grows rapidly.
It is worth noting that currently, a number of research insti-
tutions in the U.S. are working on centimeter-scale (and even
smaller) distributed sensor networks [2,29].
3.1. Hardware design of wireless sensors
The basic MICA hardware uses a fraction of a Watt of power
and consists of commercial components a square inch in size.
The hardware design consists of a small, low-power radio and
processor board (known as a mote processor/radio, or MPR,
board) and one or more sensor boards (known as a mote sen-
sor, or MTS, board). The combination of the two types of
boards form a networkable wireless sensor.
The MPR board includes a processor, radio, A/D converter,
and battery. The processor is an ATMEL ATMEGA, but there
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST
CPU speed 4 MHz
Memory ROM: 128 Kb FLASH
SDRAM: 4 Kb
EEPROM: 4 Kb
Power supply 2 AA batteries
Power consumption 0.75 mW
Processor current draw 5.5 mA (active current)
< 20 A (sleep mode)
Radio current draw 12 mA (transmit current)
1.8 mA (receive current)
< 1 A (sleep mode)
Output device 3 LEDs
I/O port Expansion connected (51 pin)
Serial port (proprietary 16-pin)
Network Wireless 4 Kbits/s at 916 MHz (ISM band)
Radio range depends on antennae configuration
Figure 2. MPR300CB specifications.
are other processors that would meet the power and cost tar-
gets. The processor runs at 4 MHz, has 128 Kb of flash mem-
ory and 4 Kb of SDRAM. In a given network, thousands of
sensors could be continuously reporting data, creating heavy
data flow. Thus, the overall system is memory constrained,
but this characteristic is a common design challenge in any
wireless sensor network.
The MPR modules contain various sensor interfaces,
which are available through a small 51-pin connector that
links the MPR and MTS modules. The interface includes: an
8-channel, 10-bit A/D converter; a serial UART port; and an
I2C serial port. This allows the MPR module to connect to a
variety of MTS sensor modules, including MTS modules that
use analog sensors as well as digital smart sensors. The MPR
module has a guaranteed unique, hard-coded 64-bit address.
The processors, radio, and a typical sensor load consumes
about 100 mW in active mode. This figure should be com-
pared with the 30 A draw when all components are in sleep
mode. Figure 2 shows a synopsis of the MPR specs.
The MTS sensor boards currently include light/tempera-
ture, two-axis acceleration, and magnetic sensors and 420 mA
transmitters. The wireless transmission is at 4 Kbps rate and
the transmission range may vary. Researchers are also devel-
oping a GPS board and a multisensor board that incorporates
a small speaker and light, temperature, magnetic, accelera-
tion, and acoustic (microphone) sensing devices. The MICA
developers community welcomes additional sensor board de-
signs.
3.2. Software and the TinyOS
A considerable portion of the challenge faced by the devel-
opers of MICA devices is in the software embedded in the
sensors. The software runs the hardware and networkmaking
sensor measurements, routing measurement data, and control-
ling power dissipation. In effect, it is the key ingredient that
makes the wireless sensor network produce useful informa-
tion.
To this end, a lot of effort has gone into the design of a
software environment that supports wireless sensors. The re-
sult is a very small operating system named TinyOS, or Tiny
137
Software footprint 3.4 Kb
Transmission cost 1 J/bit
Inactive state < 25 A
Peak load 20 mA
Typical CPU usage < 50%
Events propagate thru stack < 40 s
Figure 3. TinyOS key facts.
Microthreading Operating System, which allows the network-
ing, power management, and sensor measurement details to
be abstracted from the core application development. The
operating system also creates a standard method of develop-
ing applications and extending the hardware. Although tiny,
this operating system is quite efficient, as shown by the small
stack handling time. Figure 3 lists the key points of TinyOS.
4. The problem
An adversary A selects a single particle, p, in the plane-cloud
and allows it to monitor a local crucial event E. The general
propagation problem P is the following:
How can particle p, via cooperation with the rest of the
cloud, propagate information about event E to the receiv-
ing wall W?
Definition 4.1. Let hopt (p, W) be the (optimal) number of
hops (direct, vertical to W transmissions) needed to reach
the wall, in the ideal case in which particles always exist in
pair-wise distances R in the vertical line from p to W. Let
 be a smart-dust propagation protocol, using a transmis-
sion path of length L(, p, W) to send info about event E to
wall W. Let h(, p, W) be the number of hops (transmis-
sions) taken to reach W. The hops efficiency of protocol 
is the ratio
h(, p, W)
Ch = .
hopt (p, W)
Clearly, the number of hops (transmissions) needed char-
acterizes the energy consumption and the time needed to
propagate the information E to the wall. Remark that hopt =
d(p, W)/R, where d(p, W) is the (vertical) distance of p
from the wall W.
In the case where protocol  is randomized, or in the case
where the distribution of the particles in the cloud is a random
distribution, the number of hops h and the efficiency ratio Ch
are random variables and we study here their expected values.
The reason behind these definitions is that when p (or any
intermediate particle in the information propagation to W)
looks around for a particle as near to W as possible to pass
its information about E, it may not get any particle in the per-
fect direction of the line vertical to W passing from p. This
difficulty comes mainly from three causes: (a) due to the ini-
tial spreading of particles of the cloud in the area and because
particles do not move, there might not be any particle in that
direction; (b) particles of sufficient remaining battery power
138
may not be available in the right direction; (c) particles may
temporarily sleep (i.e., not listen to transmissions) in order
to save battery power.
Remark. Note that any given distribution of particles in the
smart dust cloud may not allow the ideal optimal number of
hops to be achieved at all. In fact, the least possible number
of hops depends on the input (the positions of the grain par-
ticles). We have chosen, however, to compare the efficiency
of our protocols to the ideal case. A comparison with the best
achievable number of hops in each input case will of course
give better efficiency ratios for our protocols.
5. The local target protocol (LTP)
Let d(pi , pj ) the distance (along the corresponding vertical
lines towards W) of particles pi , pj and d(pi , W) the (verti-
cal) distance of pi from W. Let info(E) the information about
the realization of the crucial event E to be propagated. In this
protocol, each particle p that has received info(E) from p
(via, possibly, other particles) does the following:
Search phase. It uses a periodic low energy broadcast of
a beacon in order to discover a particle nearer to W than
itself (i.e., a particle p where d(p , W) < d(p , W)).
Direct transmission phase. Then, p sends info(E) to p
via a direct line (laser) time consuming transmission.
Backtrack phase. If consecutive repetitions of the search
phase fail to discover a particle nearer to W, then p sends
info(E) to p (i.e., to the particle that it originally received
the information).
Note that one can estimate an a-priori upper bound on the
number of repetition of the search phase needed, by using the
probability of success of each search phase. This bound can
be used to decide when to backtrack.
Also note that the maximum distance d(p , p ) is R, i.e.,
the beacon transmission range (for a graphical representation
see figures 4, 5).
To enable a first step towards a rigorous analysis of smart
dust protocols, we make the following simplifying assump-
tion. The search phase takes zero time and always finds a p
Figure 4. Example of the search phase.
CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
(of sufficiently high battery) in the semicircle of center p, in
the direction towards W. Note that this assumption on al-
ways finding a particle can be relaxed in the following ways:
(a) by repetitions of the search phase until a particle is found.
This makes sense if at least one particle exists but was sleep-
ing during the failed searches; (b) we may consider, instead
of just the semicircle, a cyclic sector defined by circles of ra-
diuses R R, R and also take into account the density
of the smart cloud; (c) if the protocol during a search phase
ultimately fails to find a particle towards the wall, it may back-
track.
In this analysis we do not consider the energy spent in the
search phase. Note, however, that even the case where this
is comparable to the energy spent in actual data transmission,
the number of hops accounts for both (total energy spent is
upper bounded by a multiple of actual data transmission en-
ergy).
We also assume that the position of p is uniform in the
arc of angle 2a around the direct line from p vertical to W.
Each data transmission (one hop) takes constant time t (so the
hops and time efficiency of our protocols coincide in this
case). We also assume that each target selection is random
independent of the others, in the sense that it is always drawn
uniformly in the arc (, ).
We are aware of the fact that the above assumptions may
not be very realistic in practice, however, they allows us to
perform a first effort towards providing some concrete analyt-
ical results.
Lemma 5.1. The expected hops efficiency of the local tar-
get protocol in the -uniform case is E(Ch )  / sin , for
large hopt . Also 1
E(Ch )
/2 1.57, for 0

/2.
Proof. Due to the protocol, a sequence of points is gener-
ated, p0 = p, p1 , p2 , . . . , ph1 , ph where ph1 is a particle
within Ws range and ph is part of the wall. Let i be the
(positive or negative) angle of pi with respect to pi1 s verti-
Figure 5. Example of a transmission.
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 139

cal line to W. It is: Thus, the distribution function of i , for any 0

, is

h1

h
( )2 2 2
d(pi1 , pi )
d(p, W)
d(pi1 , pi ). Fi () = P{i
} = 1 =
2 2
i=1 i=1
and the probability density function is, for any 0

:
Since the (vertical) progress towards W is then i =  
d(pi1 , pi ) = R cos i , we get: fi () =
d
P{i
} =
2
1

.
d


h1

h
cos i
hopt
cos i . The expected local progress is:

i=1 i=1
2(1 cos )
From Walds equation for the expectation of a sum of a ran- E(cos i ) = cos fi () d = . (2)
0 2
dom number of independent random variables (see [25]), then

E(h 1)E(cos i )
E(hopt ) = hopt
E(h)E(cos i ).
 We remark that
Now, i, E(cos i ) = cos x(1/2) dx = sin /. Thus
2
E(h) 1 lim E(Ch ) = lim =1

= E(Ch )
+ . 0 0 2 sin a
sin hopt sin hopt and
Assuming large values for hopt (i.e., events happening (/2)2 2
far away from the wall, which is the most interesting case lim E(Ch ) = = 1.24.
/2 2(1 0) 8
in practice since the detection and propagation difficulty in-
creases with distance) we have (since for 0

/2 it is Lemma 6.2. The expected hops efficiency of the min-two
1
/ sin
/2) we get the result.  uniform targets protocol is 1
E(Ch )
2 /8 1.24 for
large h and for 0

/2.
6. Local optimization the min two uniform targets
protocol (m2TP) We remark that, w.r.t. the expected hops efficiency of the
local target protocol, the min-two uniform targets protocol
Note that the same basic framework holds for any situation in achieves, because of the one additional search, a relative gain
which the local (vertical) progress in the direction towards W which is (/2 2 /8)/(/2) 21.5%. We experimentally
(i.e., i ) is of the same, independent, distribution. I.e., it investigate the further gain of additional (i.e., m > 2) searches
always holds (via the Walds equation) that in section 10.
R E(h) R 1


+
E(i ) hopt E(i ) hopt 7. Tight upper bounds to the hops distribution of the
R 1 general target protocol
E(Ch ) = (1)
E(i ) E(cos i )
for large h. To understand the power of this, let us assume Consider the particle p (which senses the crucial event) at dis-
that the search phase always returns two points p , p each tance x from the wall. Let us assume that when p searches in
uniform in (, ) and that the protocol selects the best of the sector S defined by angles (, ) and radius R, another
the two points, with respect to the local (vertical) progress. particle p is returned in the sector with some probability den-



sity f ( p ) dA, where p = (xp , yp ) is the position of p
Lemma 6.1. The expected hops efficiency of the min two in S and dA is an infinitesimal area around p .
uniform targets protocol in the -uniform case is
Definition 7.1 (Horizontal progress). Let x be the projec-
2 tion of the line segment (p, p ) on the line from p vertical
E(Ch ) ,
2(1 cos ) to W.
for 0

/2 and for large h.
We assume that each search phase returns such a particle,
Proof. Let i1 , i2 the angles of the particles found and let with independent and identical distribution f ().
i = min{|i1 |, |i2 |}. Then, for any 0

, it is:
  Definition 7.2 (Probability of significant progress). Let
P{i > } = P |i1 | > |i2 | >
 2 m > 0 be the least integer such that P{x > R/m}  p,
2 2 where 0 < p < 1 is a given constant.
=
2
 
2 Lemma 7.1. For each continuous density f () on the sector S
= .
and for any constant p, there is always an m > 0 as above.
140 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS

Proof. Remark that f () defines a density function f() on reported to the wall. Since for all f () it is h  x/R = hopt
( ) its distribution func-
(0, R] which is also continuous. Let F we get that

tion. Then we want 1 F (R/m)  p, i.e., to find the first m EP (h) mx/Rq/p (m + 1)q
such that 1 p  F (R/m). Such an m always exists since

.
 hopt x/R p
F is continuous in [0, 1]. 
Theorem 7.6. The above upper bound process P estimates
Definition 7.3. Consider the (discrete) stochastic process P the expected number of hops to the wall with a guaranteed
in which with probability p the horizontal progress is R/m efficiency ratio (m + 1)/(1 p)p at most.
and with probability q it is zero, where q = 1 p.
Let Q the actual stochastic process of the horizontal
Example. When for p = 0.5 we have m = 2 and the effi-
progress implied by f ().
ciency ratio is 3, i.e., the overestimate is 3 times the optimal
number of hops.
Lemma 7.2. PP {h
h0 }
PQ {h
h0 }.

Proof. The actual process Q makes always more progress 8. The sleepawake protocol (SWP)
than P . 
We now present a new protocol for smart dust networks which
Now let t = x/(R/m) = mx/R. Consider the integer we call the sleepawake protocol. In contrast to the previ-
random variable H such that P{H = i} = q i (1 q) for any ous protocols, we now assume that we can explicitly use the
i  0. Then H is geometrically distributed. Let H1 , . . . , Ht periods that a particle is in awake mode or in sleeping mode.
be t random variables, independent and identically distributed During sleeping periods grain particles cease any communi-
according to H . Clearly then cation with the environment, thus they are unable to listen,
receive and propagate data transmitted by other particles.
The procedures of search transmission and backtrack are
Lemma 7.3. PP {number of hops is h} = P{H1 + +
the same as in the LTP.
Ht = h}.
In the above procedure, propagation of info(E) is done in
two steps; (i) particle p locates the next particle (p ) and
The probability generating function of H is transmits the information and (ii) particle p waits until the
next particle (p ) succeeds in propagating the message further
H (s) = P{H = 0} + P{H = 1}s + + P{H = i}s i + ,
towards W. In both steps particle p will remain awake. This
i.e., is done to speed up the backtrack phase in case p does not
succeed in discovering a particle nearer to W. Note, however,

p
H (s) = p 1 + qs + q 2 s 2 + + q i s i + = . that as soon as p succeeds to propagate data, p resumes its
1 qs sleepawake mode.
 Propagation protocols for such energy-restricted systems
But the probability generating function of t = H1 + +
should at least guarantee that the wall eventually receives the
Ht is then just (p/(1 qs))t by the convolution theorem of
messages that report a crucial event. The success of such
generating functions. This is just the generating function of
protocols depends on the density d of grain particles/m2 and
the t-fold convolution of geometric random variables, and it is
their distribution, the distribution of sleeping and awake time
exactly the distribution of the negative binomial distribution
periods and, of course, on the angle of the search beacon.
(see [13], vol. 1, p. 253). Thus,
We below provide some first results on the interplay be-
tween these parameters. In particular, we focus on the rela-
Theorem 7.4. tion between the maximum sleeping time period and the other
  parameters, thus allowing to program the smart-cloud energy
t t
PP {the number of hops is h} = p (q)h saving specifications accordingly.
h
  To simplify the analysis we assume that the grain particles
t +h1 t h
= pq . are uniformly distributed on the smart-dust plane. Thus, in
h the area inspected during a search phase of beacon angle
between R and R + R, the number of grain particles is
Corollary 7.5. For the process P , the mean and variance of  
the number of hops are:
N =d R (R R)
2 2

tq tq

E(h) = , Var(h) = .  2RR R2 d  2dRR. (3)

p p2
Now, we assume that the sleeping/awake time durations al-
Note that the method sketched above, finds a distribution ternate independently in each particle and have lengths s, w,
that upper bounds the number of hops till the crucial event is respectively (this can be easily achieved if during the start-up
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 141

phase, the first awake period w is set using a random bit gen- msgHandler rcvInfo(msg) {
erator, or is hardcoded into the particle by the manufacturer). Timer.stop( );
Thus, the probability that at least one of the N particles in the HOLDER = true;
PREVIOUS = sender(msg);
senders beacon search area is awake is: }
P1 = P{at least one particle is awake} msgHandler rcvReqBeacon(msg) {
 N
s initiator = sender(msg);
=1 send(initiator)[BeaconMsg];
s+w }
 2d RR
s
=1 . msgHandler rcvBeacon(msg) {
s+w remember(sender(msg), power(msg));
Thus the probability that the event report eventually }
reaches the wall is: msgHandler rcvSuccess(msg) {
Timer.start(PERIODs);

PowerDisable( );
P{success} = P1h P{h = h0 }, }
h0 =1
msgHandler rcvFail(msg) {
where P{h = h0 } is the probability density function of the Timer.stop( );
random variable h. HOLDER = true;
Let now = s/w, i.e., represents the energy saving OUTf = OUTf {sender(msg)};
specifications of the smart dust particles (a typical value for
may be 100). Then, Figure 6. The Message Handler procedures.

Definition 8.1. The energy saving specification is: eventHandler SensorEvent {

PowerEnable( );
s 1
en = =1 . Timer.stop( );
s+w 1+ HOLDER = true;
}
By taking d such that 2Rd  n(1 + ) we get P1 = eventHandler Timer.fired( ) {
1 en . Then, by the Bernoulli inequality, we have P1
E(h)
 if (power == Enable) {
1 E(h)en . This probability is non-zero when Timer.start(PERIODs);
PowerDisable( );
n > ln E(h). } else {
Timer.start(PERIODw);
This final condition allows to set the technical specifica- PowerEnable( );
tions and the propagation time accordingly in order to guar- }
antee that the crucial event is eventually reported to the wall.
Figure 7. The Event Handler procedures.

9. Implementation aspects and details
We now proceed by providing a more detailed description of
the protocols implementation in our simulation environment.
We also discuss implementation aspects of our protocols in
current technology wireless sensor networks.
We assume that the particles are equipped with TinyOS,
an event driven operating system suitable for smart dust [28].
The pseudo-code presented in figures 68 demonstrates how
to implement the SWP protocol. Note that the implementa-
tion of LTP is very similar.
At every particle, we use a Boolean variable HOLDER to
denote the status of the particle. It is set to true only if the
site holds info(E) (the information about the realization of the
crucial event E to be propagated). A variable PREVIOUS
records the particle from which info(E) was received. A set
OUTf is used to store any particle that failed to propagate a
message towards the wall (this set is used for backtracking
purposes).
In addition, for SWP we use a decreasing clock timer
CLOCK that can be explicitly activated, deactivated and set to
a given value, and two constant variables PERIODw, PERI-
ODs provided by the implementer that indicate the length of
the awake and sleeping periods of the particle. For example,
in TinyOS, to create a timer that expires every PERIODw ms
we use the statement Timer.start(TIMER_REPEAT,
PERIODw); and Timer.stop( ); terminates the timer.
Each time the timer expires, a Timer.fired( ) event
is triggered that invokes a function implemented by the
user.
Initially, the Boolean variables HOLDER and EXECUTING
are set to false, the variable PREVIOUS is set to itself and
the set OUTf is empty. For SWP, each particle is at awake
mode with its CLOCK set to a period chosen randomly in
the range (0, PERIODw + PERIODs]. In TinyOS this can
be implemented using the method Random.rand( ) of the
built-in 16-bit Linear Feedback Shift Register pseudo-random
number generator.
142 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS

The protocols use five types of messages: info(E), fail,
success, requestBeacon, beacon. The first message type is
used to propagate the actual information on the crucial event,
while the next two (fail and success) are special control mes-
sages used to signify a failure or a success in the attempt
to propagate info(E) towards the receiving wall W. The
requestBeacon and beacon messages are used by the search
phase. Interestingly, in TinyOS, radio communication follows
the Active Message (AM) model, in which each packet on the
network specifies a handler ID that will be invoked on recip-
ient nodes. When a message is received, the receive event
associated with that handler ID is signaled. Thus we only
need to define one message handler per message type. Fig-
ure 6 depicts the five message handlers implemented by the
protocols.
Remark that the beacon message handler assumes that the
communication module is capable of measuring the signal
strength of the message received by executing the function
power(msg). Similarly, the function sender(msg) is
used to extract the originator of a message, assuming that
the message structure maintains such kind of information.
Also, the function remember(. . .) adds the information
to a temporary buffer. Using these primitives by sending a
requestBeacon message the particle initiates the search phase
and then, after waiting for a sufficient period of time (so that
all neighbors respond to the request by sending a beacon),
the DetectNeighbors( ) procedure processes the tem-
porary buffer and returns a set containing those particles that
responded to the broadcast of the search beacon, ordered by
the distance of the particles (i.e., d(p , p )).
Apart from the message event handlers, the protocols use
two additional types of events: (i) SensorEvent created
by the sensors of the particle when a crucial event is real-
ized (i.e., when the particle is selected by the adversary A)
and (ii) Timer.fired( ) created by the Timer when the
counting has finished. Figure 7 depicts the two generic event
handlers implemented by our protocols. Remark that the
PowerDisable( ) and PowerEnable( ) will force the
particle to enter a snooze mode where only the Timer is ac-
tive.
Based on the above event driven functionality, particle p
executes continuously the Main task, shown in figure 8.
TinyOS provides a two-level scheduling hierarchy consisting
of tasks and hardware event handlers. Tasks are used to per-
form longer processing operations, such as background data
processing, and can be preempted by hardware event handler.
Remark that the post operation places the task on an internal
task queue which is processed in FIFO order.

task main {
if (HOLDER == true) { 10. Experimental evaluation
next = SenseNeighbours( );
if (next == nil) In this section we report on four sets of experiments that aim
BackTrack( ); to validate the theoretical analysis of the previous sections.
else { We have implemented the three protocols using C++ and the
send(next)[info(E)]; data types for two-dimensional geometry of LEDA [22]. Each
send(PREVIOUS)[success];
HOLDER = false;
class is installed in an environment that generates sensor fields
} given some parameters (such as the area of the field, the dis-
} tribution function used to drop the particles), and performs a
network simulation for a given number of repetitions, a fixed
post main( ); number of particles and certain protocol parameters. After
}
the execution of the simulation, the environment stores the
procedure SenseNeighbors { results on files so that the measurements can be represented
send( )[reqBeacon]; in a graphical way. Each experiment was conducted for more
than 10,000 times in order to achieve good average results.
Set tempSet = DetectNeighbors( );
if (tempSet == empty) In the first set of experiments, we investigate (a) the impact
return nil; of the angle and (b) the number of targets found during the
search phase, on the hops efficiency of the Local target pro-
Set out = tempSet OUTf ; tocol when considering the ideal case where the search phase
if (out == empty)
always finds a particle (of sufficiently high battery) in (, )
return nil;
(we call the measured efficiency, the ideal hops efficiency). In
return out.first( ); figure 9 we observe that for both protocols, as 0, the
} ideal hops efficiency Ch 1. Actually, the ideal Ch initially
decreases very fast with increasing , while having a limiting
procedure Backtrack { behavior of no further significant improvement when
40.
send(PREVIOUS)[fail];
HOLDER = false;
Figure 10 shows the effect of finding more than one target dur-
Timer.start(PERIODs); ing the search phase; as the number of targets increases, the
PowerDisable( ); ideal hops efficiency Ch 1. We note a similar threshold
} behavior, for a total number of 4 targets.
In the second set of experiments we study the performance
of the LTP and m2TP protocols in more realistic cases by gen-
Figure 8. The Main task and the SenseNeighbors, Backtrack procedures. erating a variety of sensor fields in a 100 m 100 m square.
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 143

In these fields, we drop n [100, 5000] particles uniformly efficiency remains unaffected. This is a result of our choice
distributed on the smart-dust plane, i.e., 0.01
d
0.5. not to include the failed searches in our measurements, that
Each smart dust particle has a radio range of R = 5 m. For is, the measurements include only the search phases that re-
carrying out identical repetitions on our experiments we ex- sulted in finding a particle p closer to W. To get a more com-
plicitly place a particle at position (x, y) = (0, 50) and we plete view on the effect of density, figure 13 shows the failure
assume that this particle detects the event. The wall is located rate (i.e., the number of times that the protocols backtracked)
at x = 100. In this set of experiments, the particle p dis- for different values of d. We observe that for low density
covered in the search phase can be located anywhere within (i.e., d
0.1) both protocols almost always use the backtrack
the cyclic sector defined by circles of radiuses 0, R and an-
gles (, ). Note that this experimental setup is based on
that used in [14,17,21]. Also, remark that the efficiency is
measured over the successful tries, i.e., we do not take into
account those runs that backtracked, however we keep track
of the total number of times that the protocol was required to
backtrack.
In figure 11 we observe that opposed to the ideal case (i.e.,
when the search phase always returns a particle on the semi-
circle), we do not get significant improvement in the hops ef-
ficiency as the angle is reduced. This is basically because
the discovered particle p might be close to p and thus the
local improvement made is of limited significance. Note that
the min-two uniform targets protocol (m2TP) achieves better
efficiency compared to the local target protocol (LTP).
Figure 12 depicts the effect of density d on the hops ef-
ficiency of the two protocols. Interestingly, we observe that Figure 11. Hops efficiency for angles [5, 90] for d = 0.3.
even for quite low density of particles (i.e., d
0.2) the hops

Figure 12. Hops efficiency for density d [0.01, 0.5] and = 90.

Figure 9. Ideal hops efficiency for angles [5, 90].

Figure 10. Ideal hops efficiency for different number of targets. Figure 13. Failure rate for density d [0.01, 0.5] and = 90.
144 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS

Figure 14. Probability of success (P{success}) over particle density d = [0.01, 0.3] for various angles = {45, 60, 90}, and random distribution.

Figure 15. Average hops efficiency (Ch ) over particle density d = [0.01, 0.3] for various angles = {45, 60, 90}, and random distribution.

mechanism, while when d  0.2 the failure rate drops very
fast to zero. This can be justified by taking into account the
average degree of each particle for various density d.
In the third set of experiments, we evaluate the perfor-
mance of the SWP protocol in the case when all the particles
remain awake (i.e., en = 0). We consider this a first step to in-
vestigate (a) the impact of the angle and (b) the effect of the
particles density d on the probability of success (P{success}),
hops efficiency (Ch ) and average number of backtracks. The
particle density was 0.01
d
0.3 and used three different
angles, = {45, 60, 90} (in degrees). The reported experi-
ments for the three different performance measures we con-
sidered are illustrated in figures 1416.
Examining figure 14, that shows the probability of suc-
cess (P{success}), we first observe that for particle density
d < 0.05 (i.e., throwing a small number of particles) the
protocol fails to propagate the critical event (i.e., the suc-
cess probability is zero). However, the probability of suc-
cess increases very fast exhibiting a threshold-like behaviour
as the particle density increases, and the protocol almost al-
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 145

Figure 16. Average number of backtracks over particle density d = [0.01, 0.3] for various angles = {45, 60, 90}, and random distribution.

Figure 17. Probability of success (P{success}) over en (where en = s/(s + w)) for various particle densities d = {0.15, 0.2, 0.25}, fixed angle = 90 and
random distribution.

ways succeeds to propagate the critical event when d > 0.15.
As expected (due to equation (3)), setting a smaller angle
reduces the probability of success in each density case
since the number of particles that respond to the search phase
gets smaller. So for = 60 the P{success} gets close to 1
when d > 0.2 while for = 45, P{success} 1 when
d > 0.25.
Regarding the average hops efficiency (Ch ) we interest-
ingly observe in figure 15 that even for a small particle den-
sity, the hops efficiency is close to the optimal. Actually, as
the particle density crosses d = 0.05 (i.e., when P{success} >
0) the hops efficiency gets close to 2.6. In fact, when d = 0.1,
Ch = 1.74 while no further gain is achieved if we throw more
particles (i.e., increase d). This is because of a sufficiently
large density leads to many particles found in the search, of
which particles already some are close to the vertical line.
Similar results hold for = 60 and = 45, although at high
particle densities, the hops efficiency is slightly better.
Finally, in figure 16 we can see the average number of
backtracks performed by the protocol in the attempt to prop-
146 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS

Figure 18. Average hops efficiency (Ch ) over en (where en = s/(s + w)) for various particle densities d = {0.15, 0.2, 0.25}, fixed angle = 90 and random
distribution.

Figure 19. Average number of backtracks over en (where en = s/(s + w)) for various particle densities d = {0.15, 0.2, 0.25}, fixed angle = 90 and random
distribution.

agate the critical event to the wall. We observe that for low
density (i.e., d
0.1) the protocols almost always uses the
backtrack mechanism, while when d  1.5 the number of
backtracks performed drops very fast to zero. Furthermore,
we observe that the number of backtracks is initially high
but decreases with a fast rate as the particle density increases.
This can be justified by taking into account the average degree
of each particle for various density d. More specifically, when
= 90 and d = 0.15 the protocol almost always succeeds in
propagating the crucial event without the need to backtrack.
The last set of experiments aims to evaluate the impact
of the energy saving specification en on the performance of
the protocol. Again, we measure the probability of success
(P{success}), hops efficiency (Ch ) and average number of
backtracks over energy saving specification (en), for three dif-
ferent particle densities (d = 0.15, 0.2, 0.25) and three differ-
ent angles = {45, 60, 90} (in degrees). We have set the
awake period w = 2 and the sleeping period s [0, 15]
thus making en [0, 0.88] (recall that en = s/(s + w)).
Figures 1719 show the measured performance for the dif-
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 147

Figure 20. Probability of success (P{success}) over en (where en = s/(s + w)) for various angles = {45, 60, 90}, fixed particle density d = 0.2 and random
distribution.

Figure 21. Average hops efficiency (Ch ) over en (where en = s/(s + w)) for various angles = {45, 60, 90}, fixed particle density d = 0.2 and random
distribution.

ferent particle densities d and figures 2022 for the different
angles .
In figures 17 and 20, that show the probability of suc-
cess for different particle densities d and angles , we ob-
serve that the protocol experiences a threshold behavior when
en = 0.75: when en
0.75 the probability of success
is close to 1 while for en > 0.75, P{success} drops very
fast to zero. In other words, even if we set the particles to
be awake only the 25% of each sleepawake cycle, it does
not affect the success of the protocol to propagate the in-
formation to W. However, in figures 18 and 20 we ob-
serve a similar threshold behavior for the average hops ef-
ficiency when en = 0.5: the hops efficiency remains unaf-
fected when en
0.5 while it decreases very fast (i.e., Ch
increases) when en > 0.5. Interestingly, figures 19 and 21
show that for the same threshold value (en = 0.5) the pro-
tocol almost always succeeds without the need to backtrack,
while for en > 0.5 the number of backtracks increases very
fast with en. Thus, although for en
0.8 the probability
of success is close to 1, setting the energy saving specifica-
tion to en = 0.5 seems to be more reasonable. This leads to
the conclusion that by setting the particles to be active only
the 50% of the overall period for which the protocol is exe-
cuted, we manage to decrease the energy requirements while
148
Figure 22. Average number of backtracks over en (where en = s/(s + w)) for various angles = {45, 60, 90}, fixed particle density d = 0.2 and random
distribution.
keeping the hops efficiency (and thus time efficiency) unaf-
fected.
11. Conclusions and future work
We presented here a model for Smart Dust and three basic
protocols (and their average case performance) for local de-
tection and propagation. We plan to investigate protocols that
trade-off hops efficiency and time, as well as study the fault-
tolerance of protocols as a function of smart dust parameters
(such as density of the cloud, the energy saving character-
istics, etc.). We also intend to investigate alternative back-
track mechanisms and study their effect on the efficiency and
fault-tolerance of the protocol. Also, we are currently work-
ing towards the design of local protocols than can monitor
the spreading of a time-sequence of events (i.e., tracking pro-
tocols). We plan to provide performance comparisons with
other protocols mentioned in the related work section. Fi-
nally, we plan to also explicitly introduce sensor faults and
study the performance (efficiency, fault-tolerance) of our pro-
tocols in this case.
References
[1] T. Antoniou, A. Boukerche, I. Chatzigiannakis, G. Mylonas and
S. Nikoletseas, A new energy efficient and fault-tolerant protocol for
data propagation in smart dust networks using varying transmission
range, in: Proc. 37th ACM/IEEE Annual Simulation Symposium ANSS,
Arlington, VA, USA, April 1822 (2004) pp. 167177.
[2] Berkeley Wireless Research Center (2001), http://bwrc.eecs.
berkeley.edu
[3] A. Boukerche and S. Nikoletseas, Protocols for data propagation in
wireless sensor networks: a survey, in: Wireless Communications Sys-
tems and Networks, ed. M. Guizani (Kluwer Academic, 2004) chap-
ter 2, pp. 2351.
CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
[4] I. Chatzigiannakis, T. Dimitriou, S. Nikoletseas and P. Spirakis, A prob-
abilistic algorithm for efficient and robust data propagation in smart
dust networks, in: Proc. 5th European Wireless Conference (EW 2004),
Barcelona, Spain, February 2427 (2004) pp. 344350.
[5] I. Chatzigiannakis, T. Dimitriou, M. Mavronicolas, S. Nikoletseas and
P. Spirakis, A comparative study of protocols for efficient data propa-
gation in smart dust networks, in: Proc. International Conference on
Parallel and Distributed Computing, EUPOPAR 2003 (2003). Also in
Parallel Processing Letters 13(4) (2003) 615627.
[6] I. Chatzigiannakis and S. Nikoletseas, A sleepawake protocol for in-
formation propagation in smart dust networks, in: Proc. 3nd Interna-
tional Workshop on Mobile and Ad-Hoc Networks, 17th Annual In-
ternational Parallel and Distributed Processing Symposium (IPDPS
2003), Nice, France, April 2125 (2003) p. 225.
[7] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, An efficient com-
munication strategy for ad-hoc mobile networks, in: Proc. 15th Inter-
national Symposium on Distributed Computing (DISC 2001), pp. 285
299. See also brief announcement in: Proc. 20th Annual Symposium on
Principles of Distributed Computing (PODC 2001), pp. 320322.
[8] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, On the average and
worst-case efficiency of some new distributed communication and con-
trol algorithms for ad-hoc mobile networks, in: Proc. 1st ACM Work-
shop on Principles of Mobile Computing (POMC 2001), pp. 119, in-
vited paper.
[9] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, Smart dust protocols
for local detection and propagation, in: Proc. 2nd ACM Workshop on
Principles of Mobile Computing (POMC 2002), pp. 916.
[10] I. Chatzigiannakis, A. Kinalis and S. Nikoletseas, Wireless sensor net-
works protocols for efficient collision avoidance in multi-path data
propagation, in: Proc. of the ACM Workshop on Performance Evalua-
tion of Wireless Ad Hoc, Sensor, and Ubiquitous Networks, PE-WASUN
(2004).
[11] D. Estrin, R. Govindan, J. Heidemann and S. Kumar, Next century
challenges: scalable coordination in sensor networks, in: Proc. 5th
ACM/IEEE International Conference on Mobile Computing (MOBI-
COM 1999) (1999) pp. 263270.
[12] H. Euthimiou, S. Nikoletseas and J. Rolim, Energy balanced data prop-
agation in wireless sensor networks, in: Proc. 4th International Work-
shop on Algorithms for Wireless, Mobile, Ad-Hoc and Sensor Networks
(IPDPS/WMAN04) (Computer Society Press, 2004) p. 225.
[13] W. Feller, An Introduction to Probability Theory and Its Applications
(Wiley, New York, 1957).
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST
[14] W.R. Heinzelman, A. Chandrakasan and H. Balakrishnan, Energy-
efficient communication protocol for wireless microsensor networks,
in: Proc. 33rd Hawaii International Conference on System Sciences
(HICSS 2000) (2000) p. 8020.
[15] W.R. Heinzelman, J. Kulik and H. Balakrishnan, Adaptive protocols
for information dissemination in wireless sensor networks, in: Proc.
5th ACM/IEEE International Conference on Mobile Computing (MO-
BICOM 1999) (1999) pp. 174185.
[16] S.E.A. Hollar, COTS dust, M.Sc. Thesis in engineeringmechanical en-
gineering, University of California, Berkeley, USA (2000).
[17] C. Intanagonwiwat, D. Estrin, R. Govindan and J. Heidemann, Impact
of network density on data aggregation in wireless sensor networks,
Technical Report 01-750, University of Southern California Computer
Science Department (November, 2001).
[18] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion:
A scalable and robust communication paradigm for sensor networks,
in: Proc. 6th ACM/IEEE International Conference on Mobile Comput-
ing (MOBICOM 2000) (2000) pp. 5667.
[19] B. Karp, Geographic routing for wireless networks, Ph.D. Dissertation,
Harvard University, Cambridge, MA (2000).
[20] -adaptive multi-domain power aware sensors (April 2001),
http://www-mtl.mit.edu/research/icsystems/uamps
[21] A. Manjeshwar and D.P. Agrawal, TEEN: A routing protocol for en-
hanced efficiency in wireless sensor networks, in: Proc. 2nd Interna-
tional Workshop on Parallel and Distributed Computing Issues in Wire-
less Networks and Mobile Computing (PDCIWNMC02), 16th Annual
International Parallel and Distributed Processing Symposium (IPDPS
2002), Fort Lauderdale, FL, USA, April 1519, 2002 (2004) p. 1956.
[22] K. Mehlhorn and S. Nher, LEDA: A Platform for Combinatorial and
Geometric Computing (Cambridge University Press, 1999).
[23] S. Nikoletseas, I. Chatzigiannakis, H. Euthimiou, A. Kinalis, A. Anto-
niou and G. Mylonas, Energy efficient protocols for sensing multiple
events in smart dust networks, in: Proc. 37th ACM/IEEE Annual Simu-
lation Symposium (ANSS) (2004).
[24] C.E. Perkins, Ad Hoc Networking (Addison-Wesley, Boston, MA,
2001). (Reference to book spans across different chapters since the text
is discussing the case of porting existing mobile ad-hoc routing proto-
cols to smart dust networks.)
[25] S.M. Ross, Stochastic Processes, 2nd ed. (Wiley, 1995).
[26] C. Schurgers, V. Tsiatsis and M.B. Srivastava, STEM: Topology
management for energy efficient sensor networks, in: Proc. IEEE
Aerospace Conference 2002, Vol. 3 (2002) pp. 10991108.
[27] S. Singh and C.S. Raghavendra, PAMAS: Power aware multi-access
protocol with signalling for ad hoc networks, ACM Computer Com-
munication Review 28(3) (1998) 525.
[28] TinyOS, A component-based os for the network sensor regime (Octo-
ber, 2002), http://webs.cs.berkeley.edu/tos/
[29] Wireless Integrated Sensor Networks (April 2001), http://www.
janet.ucla.edu/WINS/
[30] W. Ye, J. Heidemann and D. Estrin, An energy-efficient MAC protocol
for wireless sensor networks, in: Proc. 12th IEEE International Con-
ference on Computer Networks (INFOCOM 2002) (2002) pp. 947957.
Ioannis Chatzigiannakis is a Researcher of Re-
search Unit 1 (Foundations of Computer Science,
Relevant Technologies and Applications) at the
Computer Technology Institute (CTI), Greece. He
has received his B.Eng. degree from the University
of Kent, UK in 1997 and his Ph.D. degree from the
Computer Engineering and Informatics Department
of Patras University, Greece in 2003, under the su-
pervision of Prof. Paul Spirakis. His research in-
terests include distributed computing, mobile com-
puting and algorithmic engineering. He has served as an external reviewer
149
in major international conferences. He has participated in several European
Union funded R&D projects, and worked in the private sector.
E-mail: ichatz@cti.gr
Sotiris E. Nikoletseas is currently a Lecturer Pro-
fessor at the Computer Engineering and Informatics
Department of Patras University, Greece and also a
Senior Researcher and Director of Research Unit 1
(Foundations of Computer Science, Relevant Tech-
nologies and Applications) at the Computer Tech-
nology Institute (CTI), Greece. His research in-
terests include probabilistic techniques and random
graphs, average case analysis of graph algorithms
and randomized algorithms, algorithmic applications
of probabilistic techniques in distributed computing (focus on ad-hoc mobile
networks and wireless sensor networks), algorithmic applications of combi-
natorial and probabilistic techniques in fundamental aspects of modern net-
works (focus on network reliability and stability), approximation algorithms
for computationally hard problems. He has published over 80 scientific ar-
ticles in major international conferences and journals and has co-authored
a book on probabilistic techniques, a chapter in the Handbook of Random-
ized Computing (Kluwer Academic) and several chapters in books of inter-
national circulation in topics related to distributed computing. He has been
invited speaker in international scientific events and Universities and he has
delivered several tutorials and keynote talks. He has been a reviewer for
important computer science journals and has served in the Program and Or-
ganizing Committees of International Conferences and Workshops. He has
participated in many European Union funded R&D projects.
E-mail: nikole@cti.gr
Paul G. Spirakis born in 1955, got his Ph.D. from
Harvard University in 1982. He became an Assistant
Professor at NYU (the Courant Institute) the same
year. He was then elected as a tenured Associate
Professor at Patras University, Greece, in the De-
partment of Computer Engineering and Informatics.
He became a Full Professor in the same Department
in 1990. He served as the Chairman of the Depart-
ment of Computer Engineering and Informatics for
six years. Since 1996 he is the Director of the Re-
search and Academic Computer Technology Institute (RACTI) of Greece.
His research interests include algorithms, probabilistic techniques, distrib-
uted and parallel computing and average case analysis. Recently P. Spirakis
is active in the area of algorithmic aspects of game theory and also in algo-
rithmic aspects of ad-hoc and sensor networks. Paul Spirakis has won the top
prize of the Greek Mathematical Society in 1973 as a student. Since then he
won several awards and many competitive grants. He was appointed a dis-
tinguished Visiting Researcher of Max Planck Informatik. He served as the
National Representative of Research in Informatics in the EU for four years
and is now a Member of the ISTAG group, a high level EU group respon-
sible for Research Planning in Informatics. He is currently one of the two
Vice Chairs of the European Association for Theoretical Computer Science
(EATCS). He is a high level consultant of the Greek State and Industry in
Informatics. Paul Spirakis has published extensively in many journals and
conferences of his field (more than 150 publications currently). The journals
and conferences where his work appears are among the most competitive
worldwide. He coauthored two books with Cambridge University Press and
seven books in Greek. Paul Spirakis serves as an Editor in many scientific
journals of computer science and usually in the scientific commitees of some
of the most prestigeous computer science conferences. He also serves peri-
odically as a high level research evaluator for the EU and Greece.
E-mail: spirakis@cti.gr
 Mobile Networks and Applications 10, 151168, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Training a Wireless Sensor Network

A. WADAA, S. OLARIU and L. WILSON
Department of Computer Science, Old Dominion University, Norfolk, VA 23529-0162, USA

M. ELTOWEISSY
Department of Computer Science, Virginia Tech, Falls Church, VA 22043, USA

K. JONES
NASA Langley Research Center, Hampton, VA 23681, USA

Abstract. The networks considered in this paper consist of tiny energy-constrained commodity sensors massively deployed, along with one
or more sink nodes providing interface to the outside world. Our contribution is to propose a scalable energy-efficient training protocol for
nodes that are initially anonymous, asynchronous and unaware of their location. Our training protocol imposes a flexible and intuitive coor-
dinate system onto the deployment area and partitions the anonymous nodes into clusters where data can be gathered from the environment
and synthesized under local control. An important by-product of the training protocol is a simple and natural data fusion protocol as well
as an energy-efficient protocol for routing data from clusters to the sink node. Being energy-efficient, our training protocol can be run on
either a scheduled or ad-hoc basis to provide robustness and dynamic reconfiguration. We also outline a way of making the training protocol
secure by using a parameterized variant of frequency hopping.
Keywords: wireless sensor networks, self-organization, dynamic coordinate system, training, clustering, security, energy-efficient protocols

1. Introduction mation from raw local data obtained by individual sensor

Recent advances in nano-technology have made it possible
to develop a large variety of Micro Electro-Mechanical Sys-
tems (MEMS) miniaturized low-power devices that inte-
grate sensing, special-purpose computing and wireless com-
munications capabilities [1820,44]. It is expected that these
small devices, referred to as sensor nodes, will be mass-
produced and deployed, making their production cost negli-
gible. Individual sensor nodes have a small, non-renewable
power supply and, once deployed, must work unattended.
For most applications, we envision a massive deployment of
sensor nodes, perhaps in the thousands or even tens of thou-
sands [23,40].
Aggregating sensor nodes into sophisticated computa-
tional and communication infrastructures, called wireless sen-
sor networks, will have a significant impact on a wide array
of applications ranging from military, to scientific, to indus-
trial, to health-care, to domestic, establishing ubiquitous wire-
less sensor networks that will pervade society redefining the
way in which we live and work [28,32]. The novelty of
wireless sensor networks and the tremendous potential for
a multitude of application domains has triggered a flurry of
activity in both academia and industry. We refer the reader
to [1,2,21,30,35,38] for a summary of recent applications of
wireless sensor networks.
The fundamental goal of a sensor network is to produce,
over an extended period of time, globally meaningful infor-
This work was supported, in part, by a grant from the Commonwealth of
Virginia Technology Research Fund (SE 2001-01) through the Common-
wealth Information Security Center.
nodes. Importantly, this goal must be achieved in the context
of prolonging as much as possible the useful lifetime of the
network and ensuring that the network remains highly avail-
able and continues to provide accurate information in the face
of security attacks and hardware failure. The sheer number of
sensor nodes in a sensor network combined with the unique
characteristics of their operating environment (anonymity of
individual sensors, limited power budget and a possibly hos-
tile environment), pose unique challenges to the designers of
protocols. For one thing, the limited power budget at the
individual sensor node level mandates the design of ultra-
lightweight data gathering, fusion, and communication pro-
tocols. An important guideline in this direction is to perform
as much local data processing at the sensor level as possible,
avoiding the transmission of raw data through the sensor net-
work. Recent advances in hardware technology are making
it plain that the biggest challenge facing the sensor network
community is the development of ultra-lightweight commu-
nication protocols ranging from training, to self-organization,
to network maintenance, to security, to data collection and
fusion, to routing, among many others [28,31,37].
There are several possible techniques that can be used to
interface sensor networks to the outside world and, in par-
ticular, to harvest the information they produce. Perhaps the
simplest involves using one or several special sink nodes de-
ployed alongside with the sensor nodes. In this scenario, the
raw data collected by individual sensor nodes is fused, in
stages, and forwarded to the sink nodes that provide the in-
terface to the outside world. However, in some applications,
it is impossible or impractical to deploy sink nodes within the
152
sensor network. In such cases the task of harvesting the infor-
mation produced by the sensor network and that of providing
an interface to the outside world may be performed by aircraft
and/or helicopters over-flying the sensor network, or by laser
transmission to a satellite constellation. In this latter case,
the bulk of the inter-sensor communications is by radio, since
such communications are point to multi-point, while special-
ized sensors acting as local sinks communicate with the satel-
lite constellation using laser beams, for example.
1.1. Securing wireless sensor networks
It is anticipated that in most application domains, sensor net-
works will constitute an information source that is a mission-
critical system component and will, thus, require commensu-
rate security protection. If an adversary can thwart the work
of the network by perturbing the information produced, stop-
ping production, or pilfering information, then the usefulness
of sensor networks will be drastically curtailed. Thus, secu-
rity is a major issue that must be resolved in order for the
potential of wireless sensor networks to be fully exploited.
The task of securing wireless sensor networks is compli-
cated by the fact that the sensors are mass-produced anony-
mous devices with a severely limited energy budget, and,
initially no knowledge of their location in the deployment
environment. Security must be provided even though sensor
nodes are unattended and vulnerable to a vast array of attacks
[3,4,10,22,30,42].
Wireless sensor networks are sufficiently different from
ad-hoc networks that security solutions designed specifically
for the former do not apply to the latter [10,32,42]. Indeed,
in was recently noted that the ultra-lightweight protocols im-
posed by the stringent energy limitations may leave not much
room for advanced encryption schemes. Consequently, pro-
tection against overhearing in military applications and pri-
vacy protection in personal systems needs to be inherently
built into the concepts underlying sensor network models and
protocols from the beginning. Reliability is expected to be a
result of the large number of sensors deployed for a specific
task. However, this can only be obtained if defective sensors
can be excluded from the communication, and the sensors are
calibrated either individually or collectively, either before
deployment or continuously in their environment.
1.2. Our contributions
We view our main contribution at several levels:
First we propose a virtual infrastructure a dynamic co-
ordinate system for a massively deployed collection of
anonymous sensor nodes. This coordinate system yields,
at no extra cost, a clustering scheme: two nodes are in the
same cluster only if they have the same coordinates.
We then go on to show that training the sensor nodes the
process through which nodes learn their coordinates can
be performed by a protocol that is at the same time light-
weight and secure. Indeed, we outline a way of making the
WADAA ET AL.
training protocol secure by using a parameterized variant
of frequency hopping.
Next, we show that in a trained wireless sensor network
routing and data fusion can be performed by very simple
and energy-efficient protocols.
Finally, we show how to design the coordinate system such
as to minimize the power expended in collecting and rout-
ing data.
The remainder of this paper is organized as follows. Sec-
tion 2 discusses the sensor node model used throughout the
work. In particular, it discusses genetic material with which
sensor nodes are endowed prior to deployment and which will
be key in securing sensor networks. Section 3 discusses wire-
less sensor networks, as a conglomerate of individual sensor
nodes that have to self-organize and self-govern. In particu-
lar, therein we discuss interfacing sensor networks with the
outside world, a work model for sensor networks, as well as a
brief preview of the training process. Next, section 4 proposes
routing and data fusion algorithms in a trained sensor net-
work. Section 5 is the backbone of the entire paper, presenting
the theoretical underpinnings of the training process. We note
that within this section we discuss the details of a lightweight
synchronization protocol for sensor networks. Section 6 dis-
cusses the longevity of sensor networks in terms of a number
of system parameters. Section 7 takes a look at the problem
of evaluating the energy expenditure per sensor node. Sec-
tion 8 shows how to choose the coronas in such a way that the
energy expended for conveying the results to the sink node is
minimized. Finally, section 9 offers concluding remarks and
maps out areas for future investigations.
2. The sensor node model
We assume a sensor node to be a device that possesses three
basic capabilities: sensory, computation, and wireless com-
munication as illustrated in figure 1. The sensory capability is
necessary to acquire data from the environment; the computa-
tional capability is necessary for aggregating data, processing
control information, and managing both sensory and commu-
nication activity. Finally, the wireless communication capa-
bility is necessary for sending (receiving) aggregated data and
control information to (from) other sensors or the sink.
We assume that individual sensor nodes operate subject to
following fundamental constraints.
(a) Sensor nodes are anonymous they do not have fabrica-
tion-time identities.
(b) Sensor nodes are tiny, commodity devices that are mass-
produced in an environment where testing is a luxury.
(c) Each sensor has a non-renewable energy budget; when
the on-board power supply is exhausted, the sensor node
is expired.
(d) In order to save energy, each sensor node is in sleep mode
most of the time, waking up at random points in time for
short intervals under the control of an internal timer.
TRAINING A WIRELESS SENSOR NETWORK
Figure 1. The anatomy of a sensor node.
(e) Each sensor has a modest transmission range, perhaps a
few meters. This implies that out-bound messages sent
by a sensor can reach only the sensors in its proximity,
typically a small fraction of the sensors deployed.
(f) Individual sensor nodes must work unattended once de-
ployed it is either infeasible or impractical to devote at-
tention to individual sensor nodes.
At any point in time, a sensor, will be engaged in perform-
ing one of a finite set of possible operations, or will be asleep.
Three basic operations are sensing (to collect raw measure-
ments), data fusion and/or aggregation (to derive target data
from raw measurements), routing (to communicate raw mea-
surements, target data, and control data). We assume each
operation performed by a sensor consumes a known fixed
amount of energy and that a sleeping sensor performs no op-
eration and consumes, essentially, no energy.
It is worth mentioning that while the energy budget can
supply short-term applications, sensors dedicated to work
over years may need to scavenge energy from the specific en-
vironment they are placed into, employing light, temperature,
vibration, kinetics, magnetic fields, etc.
2.1. Genetic material
The nodes genetic material plays a key role in driving the
functionality of different node protocols. To illustrate this, we
consider protocols implementing our proposed security solu-
tion for the sensor network.
We assume that at pre-deployment time the sensor nodes
are injected, in a secure environment, with the following ge-
netic material:
a standard pseudo-random number generator (one of
public-domain algorithms available);
a set of secret seeds to be used as parameters for the ran-
dom number generator;
an initial time (at this point all the sensor nodes are syn-
chronous to the sink node).
It is important to note that immediately after deployment
all the clocks are synchronous. In time, however, clocks will
drift and periodic re-synchronization becomes necessary. For
reasons of simplicity, we assume that synchronization is al-
ways done to the master clock running at the sink. As we will
show in detail later in this section, one of our main contribu-
tions is a light-weight re-synchronization protocol.
153
Classical frequency hopping mechanisms have been used
as a means of combating jamming both hostile and non-
hostile and of implementing frequency diversity and interfer-
ence averaging in a non-hostile context [13,45]. Typically
these mechanisms offered little cryptographic value. Cryp-
tographic techniques such as encryption, on the other hand,
are customarily used to address security problems in all but
the physical layer in the network. The key idea behind our
proposed security solution is that by extending classical fre-
quency hopping techniques using symmetric key cryptogra-
phy, security problems in the physical layer, as well as in
other layers in the network can be uniformly addressed in a
unified framework; we call this framework randomized fre-
quency hopping.
We are now in a position to show how the genetic material
is used in support of secure communications in a sensor net-
work. For this purpose, it is useful to imagine three sequences
of random numbers as follows:
an infinite sequence of t1 , t2 , . . . , ti , . . . of time epoch
lengths;
an infinite sequence n1 , n2 , . . . , ni , . . . of frequency sets
drawn from a large universe, e.g., the ISM band;
for every ni (i
1), an infinite permutation f1i , f2i , . . . of
frequencies from ni .
Importantly, these sequences can be generated locally by each
sensor node using the injected genetic material and, therefore,
do not need to be communicated after deployment.
We assume that time is ruled into epochs. During the ith
time epoch, of length ti , a frequency set ni will be used sub-
ject to a hopping pattern described by the hopping sequence
f1i , f2i , . . . . Thus, as long as a sensor node is synchronous
to the sink, it knows the current time epoch, the offset into
the epoch, the set of frequencies in use during the epoch, as
well as the hopping pattern in force during the epoch. To
an outside observer, however, successive epoch lengths, hop-
ping sets, and hopping patterns appear as the product of an
unknown random process. Given that techniques are known
to discover a hopping sequence by monitoring transmissions,
security can only be provided if the design modifies the hop-
ping sequence in less time than is required to discover the
sequence. The choice of frequency hopping parameters de-
termines the time required to discover the sequence (the mag-
nitude of the challenge to an adversary).
154
2.2. Making sensor nodes tamper-resistant
The most obvious tamper resistance strategies are hardware-
based and involve special hardware circuits within the sen-
sor node to protect sensitive data, special coatings or tamper
seals. However, hardware solutions to the tampering prob-
lem require extra circuitry that increases the cost and hard-
ware complexity of sensor nodes. Worse yet, the additional
hardware is very likely to consume valuable energy, already
in short supply. Also special coatings and seals may offer
protection against some but, certainly, not all tampering at-
tempts. Indeed, it is assumed that a sufficiently capable adver-
sary can extract confidential information, thus compromising
the sensor node. Thus, not surprisingly, tamper resistance or
tamper protection is not found in present-day sensor nodes
[3,4,10]. Since wireless sensor networks must function unat-
tended, the potential for physical tampering attacks is signif-
icant. It is worth noting that while pre-deployment tamper
detection may be worthwhile, post-deployment tamper de-
tection is of little use in wireless sensor networks since, in
the vast majority of applications, inspecting individual sensor
nodes is not an option. Also, physical tampering may com-
promise only the node attacked (ideal), the immediate neigh-
borhood of the node attacked, or the entire network. To cope
with these conditions, our solution subscribes to the notion of
self-guarding in that each sensor node should be able to de-
tect, independently, physical tampering and should react such
that the impact of the attack is minimal. Our solution to en-
dow individual sensor nodes with tamper resistance does not
require additional or more sophisticated hardware.
In order to set the stage for discussing our solution, we note
that the tampering threat model assumes that the adversary is
either trying to force open an individual sensor node in-
situ; or
is physically removing the sensor node from the deploy-
ment area.
We guard against the first threat by blanking out the mem-
ory, triggered by a simple switch. We guard against the sec-
ond threat by relying on local data that the sensor can collect,
thus establishing a unique signature of its neighborhood that
is difficult to forge. To be more specific, immediately after
deployment each sensor transmits, during its wake time, on
a specified sets of frequencies, using a frequency hopping
sequence established prior to deployment. This allows in-
dividual sensor nodes to collect an array of signal strengths
from the sensors in their locale. It is important to recall that
sensors do not have identities and that, consequently, the ar-
ray of signal strengths is the only data available to the sensor
node. This array, establishes, in the obvious way, a signature
of the neighborhood of the node. For this reason the array
will be referred to as the nodes neighborhood signature array
(NSA, for short). If the node is removed from the area of de-
ployment, it will notice changes in the signals received when
compared to its NSA and erase its own memory to prevent the
tampering agent from gaining access to information secret to
WADAA ET AL.
the sensor network. Note also that tampering attempts that in-
volve the removal of several sensor nodes simultaneously will
also be defeated since some node in the set of removed nodes
will notice changes in its NSA and can alert the others.
3. Structure and organization of a wireless sensor
network
We envision a massive deployment of sensor nodes, perhaps
in the thousands or even tens of thousands. The sensor nodes
are aggregated into sophisticated computational and com-
munication infrastructures, called wireless sensor networks,
whose goal is to produce globally meaningful information
from data collected by individual sensor nodes. However,
the massive deployment of sensors nodes in a sensor net-
work, combined with anonymity of individual sensors, lim-
ited power budget and in many applications a hostile envi-
ronment, pose daunting challenges to the design of protocols
for sensor networks. For one thing, the limited power bud-
get at the individual sensor node level mandates the design of
ultra-lightweight communication protocols. Likewise, issues
concerning how the data collected by individual sensor nodes
could be queried and accessed and how concurrent sensing
tasks could be executed internally are of particular signifi-
cance. An important guideline in this direction is to perform
as much local data processing at the sensor level as possible,
avoiding the transmission of raw data through the sensor net-
work. Indeed, it is known that it costs 3 J of energy to transmit
1 Kb of data a distance of 100 meters. Using the same amount
of energy, a general-purpose processor with the modest speci-
fication of 100 million instructions/Watt executes 300 million
instructions [31,37].
As a consequence, the sensor network must be multi-hop
and only a limited number of the sensor nodes count the sink
among their one-hop neighbors. For reasons of scalability,
it is assumed that no sensor node knows the topology of the
network.
Our work focuses on the design of ultra-lightweight orga-
nization and communication protocols for a class of wireless
sensor networks consisting of a single sink node and a large
number of sensors nodes randomly deployed in the transmis-
sion range of the sink.
A basic management problem in wireless sensor networks
is to balance the utility of the activity in the network against
the cost incurred by the network resources to perform this
activity. The scarce resource in the network that is of primary
concern is energy.
3.1. Interfacing sensor networks
We assume that the sensor network is connected to the outside
world (e.g., point of command and control, the Internet, etc.)
through a gateway node. The gateway node may or may not
be collocated with the sensor nodes in the deployment area.
Referring to figure 2, we note that the interface with the out-
side world may be achieved by a helicopter or aircraft over-
flying the sensor network, and collecting information from a
select group of reporting nodes. In such scenarios communi-
cation between individual sensor nodes is by radio, while the
TRAINING A WIRELESS SENSOR NETWORK
Figure 2. A sensor network with a mobile external gateway.
Figure 3. A sensor network with a central sink node.
reporting nodes are communicating with the external gateway
by laser.
One can easily have a mobile sink, or collection of mobile
sinks for fault tolerance, assume the role of the gateway in
the network. In case the sink is collocated with the sensor
network, it can also be in charge of performing any necessary
training and maintenance operations.
A somewhat complementary view, illustrated in figure 3 is
to have a sink node collocated with the sensor nodes play the
role of the gateway. In this case, the sink node has a full range
of computational capabilities, can send long-range directional
broadcasts to all sensors, can receive messages from nearby
sensors, and has a steady power supply. However, since the
sink is a single point of failure in this model, we envision
that in practice multiple (backup) sink nodes will exist in the
network.
3.2. A work model for wireless sensor networks
We take the view that the sensor network performs the tasks
mandated by an end-user (perhaps a point of command and
155
Figure 4. Illustrating the transaction-based network management.
control) that is remote from the network itself. Assuming
the sensor network model depicted in figure 3, the sink node
serves as the interface between the end user and the network.
We characterize the work activity in the network in terms of
an event model. Under the event model, the utility of the sen-
sor network is measured by the time period during which it
guarantees a specific Quality of Service (QoS) for detection
and notification of event types of interest to the application.
Based on this work model, we propose a hierarchical multi-
level network management approach, as illustrated in figure 4.
The hierarchy involves the following layers:
application layer: high-level consumers of information
produced by the sensor network;
event layer: provides the interface between the sensor net-
work and the application layer.
We now discuss each layer in detail. Referring to figure 4,
the application layer issues high-level requests, of a coarse
semantic granularity defined in terms of application-level ab-
stractions, referred to as Application events (A-events, for
short) to be performed by the sensor network. The A-event is
a task that takes the form of a tuple consisting of a high-level
action, along with a desired level of QoS. As an example, the
A-event (Fire, p) requires that the occurrence of fire be de-
tected in the area of interest with probability at least p. Here,
of course, p specifies the requested QoS.
The event layer provides the interface between the ap-
plication layer and the sensor network. This layer receives
A-events, i.e., high-level tasks and QoS requests from the ap-
plication layer, considers the current state of the sensor net-
work and its capabilities including the remaining energy bud-
get both globally and within the individual clusters, and then
negotiates a contract with the application layer before com-
mitting the network. Due to this negotiation, the network will
not squander resources needlessly by attempting to carry out
an A-event that it does not currently have the resources to pro-
vide. Also a set of A-events queueing for service in the event
layer will be prioritized in order to get the greatest benefits
from the sensor network. After a contract has been agreed
upon, the event layer translates the corresponding A-event
into individual tasks, termed primitive events (P-events, for
short), assigned to individual clusters. The clusters must then
perform these tasks at the QOS level required and send the
data back to the sink for further consolidation and analysis
156
in the event layer. The polished information from this ef-
fort is provided to the application layer for proper dissemi-
nation.
To continue our example, assume that the event layer de-
termines that the A-event (Fire, p) is feasible for the sensor
network. Assuming that the occurrence of fire is predicated
on high temperature, low humidity and the presence of smoke,
the event layer will then translate (Fire, p) into the following
(P-events):
(Temperature, t0 , q): detect with probability larger than
q whether the temperature reading is higher than thresh-
old t0 .
(Smoke, q  ): detect with probability larger than q  that
there is smoke.
(Humidity, h0 , q  ): detect with probability higher than q 
whether the humidity is lower than threshold h0 .
On the other hand, if the A-event (Fire, p) is infeasible for
the sensor network, the event layer will negotiate with the
application layer for a new task, for example, (Fire, p ) with
p < p.
3.3. Training a wireless sensor network
It was recognized that some applications require sensory data
with some location awareness, encouraging the development
of communication protocols that are location aware and per-
haps location dependent. The practical deployment of many
sensor networks will result in sensors initially unaware of
their location: they must be trained in this vital information.
Further, due to limitations in form factor, cost per unit and
energy budget, individual sensor nodes are not expected to be
GPS-enabled. Moreover, many probable application environ-
ments limit satellite access.
The localization problem is for individual sensor nodes to
determine, as closely, as possible their geographic coordinates
in the area of deployment. Prominent solutions to the localiza-
tion problem are based on multilateration [79,12,16,29,33].
Most of these solutions assume the existence of several an-
chor nodes that are aware of their location (perhaps by endow-
ing them with a GPS-like devices). Sensor nodes receiving
location messages from at least three sources can approxi-
mate their own locations. For a good survey of localization
protocols for wireless sensor networks we refer to [25].
In some other applications, exact geographic location is
not necessary: all that the individual sensor node need is
coarse-grain location awareness. There is an obvious trade-
off: coarse-grain location awareness is lightweight but the
resulting accuracy is only a rough approximation of the ex-
act geographic coordinates. Figure 5 illustrates a possible
way of inducing such a coarse-grain location awareness by
an overflying aircraft or helicopter. All that the individual
sensor nodes need is to determine their approximate distance
to three different positions of the training agent. We omit the
details.
Our approach is different: we obtain this coarse-grain
location awareness by the training protocol that imposes a
WADAA ET AL.
Figure 5. Acquiring coarse-grain location awareness.
coordinate system onto the sensor network. An interesting
by-product of our training protocol is that it provides a par-
titioning into clusters and a structured topology with natural
communication paths. The resulting topology will make it
simple to avoid collisions between transmissions of nodes in
different clusters, between different paths and also between
nodes on the same path. This is in contrast with the major-
ity of papers that assume routing along spanning trees with
frequent collisions.
Clustering was proposed in large-scale networks as a
means of achieving scalability through a hierarchical ap-
proach. For example, at the medium access layer, clustering
helps increase system capacity by promoting the spatial reuse
of the wireless channel; at the network layer, clustering helps
reducing the size of routing tables and striking a balance be-
tween reactive and proactive routing. It is intuitively clear that
wireless sensor networks benefit a great deal from clustering;
indeed, separating concerns about inter-cluster management
and the intra-cluster management can substantially decrease,
and load balance the management overhead. Given the im-
portance of clustering, a number of clustering protocols for
wireless sensor networks have been proposed in the recent
literature [5,11,15]. However, virtually all clustering proto-
cols for wireless sensor networks assume tacitly or explicitly
that individual sensor nodes have identities.
As it turns out, our clustering protocol has the following
desirable features:
lightweight as a by-product of training;
organizes anonymous asynchronous nodes;
a cluster is the locus of all nodes having the same coordi-
nates; and
individual nodes need not know the identity of other nodes
in their cluster.
TRAINING A WIRELESS SENSOR NETWORK
Figure 6. A trained sensor network.
In the remainder of this work we assume a wireless sensor
network that consists of a sink and a set of sensors randomly
deployed in its broadcast range as illustrated in figure 3. For
simplicity, we assume that the sink node is centrally placed,
although this is not really necessary. The task of training
refers to imposing a coordinate system onto the sensor net-
work in such a way that each sensor belongs to exactly one
sector.
The coordinate system divides the sensor network area into
equiangular wedges. In turn, these wedges are divided into
sectors by means of concentric circles or coronas centered
at the sink and whose radii are determined to optimize the
transmission efficiency of sensors-to-sink transmission as will
be discussed later. Sensors in a given sector map to a cluster,
the mapping between clusters and sectors is one-to-one.
Referring to figure 6, the task of training a sensor network
involves establishing:
Coronas: The deployment area is covered by k coronas de-
termined by k concentric circles of radii r1 < r2 < < rk
centered at the sink node.
Wedges: The deployment area is ruled into a number of an-
gular wedges centered at the sink node.
As illustrated in figure 6, at the end of the training period
each sensor node has acquired two coordinates: the identity of
the corona in which it lies, as well as the identity of the wedge
to which it belongs. Importantly, the locus of all the sensor
nodes that have the same coordinates determines a cluster.
4. Routing and data fusion in a trained sensor network
The main goal of this section is to show that once a wireless
sensor network has been trained, both routing and data fusion
become easy and straightforward.
4.1. Routing
The routing problem in sensor networks differs rather sub-
stantially from routing in other types of wireless networks.
157
Figure 7. Illustrating communication paths to the sink.
For one thing, individual sensor nodes do not have unique
identifiers; thus, standard addressing methods do not work di-
rectly. For another, the stringent energy limitations present in
sensor network render the vast majority of conventional rout-
ing protocols impractical.
Given the importance of routing, it is not surprising to see
that a number of routing protocols specifically designed for
wireless sensor networks were proposed in the literature. For
example, in [21] Intanagonwiwat et al. describe directed dif-
fusion and a companion routing protocol based on interest ta-
bles at the expense of maintaining a cache of information in-
dexed by interest area at each node. Shah and Rabaey [34]
responds to client requests by selecting paths that maximize
the longevity of the network rather than minimize total power
consumed by a path with path options established by local
flooding. The protocols of Kulik et al. [24] are based on
a push-pull system where the nodes send metadata first us-
ing routing that is optimal for point-to-point communication,
but does not benefit from established predefined paths. Other
routing protocols include rumor routing [6], and multi-path
routing [14], among others. As we are about to demonstrate,
our training protocol provides a novel solution to the routing
problem by yielding energy-efficient paths based routing.
Recall that sensor networks are multi-hop. Thus, in order
for the sensing information to be conveyed to the sink node,
routing is necessary. Our cluster structure allows a very sim-
ple routing process as described below. The idea is that the
information is routed within its own wedge along a virtual
path joining the outermost sector to the sink, as illustrated
in figure 7. The collection of all the virtual paths (one per
wedge) defines a tree. In this tree, each internal node, except
for the root, has exactly one child, largely eliminating MAC
level contention in sending sensor information to the sink.
Recently, a number of MAC layer protocols for wire-
less sensor networks have been proposed in the literature
[36,41,43]. Its worthwhile to note that in our routing scheme
by appropriately staggering transmissions in neighboring
wedges, collision and, therefore, the need for retransmissions
158
is completely eliminated. Thus, our training protocol implies
an efficient MAC protocol as well.
4.2. Data fusion
Once sensory data was collected by a multitude of sensor
nodes, the next important task is to consolidate the data in
order to minimize the amount of traffic to the sink node. We
place the presentation in the context of our work model. To be
more specific, we assume that the cluster identified by (i, j )
that is, the set of sensor nodes located in sector Ai,j , where
i is the corona identifier, and j is the wedge identifier, are
to perform a certain task T . A number of sensors in sec-
tors A1,j , A2,j , . . . , Ai1,j are selected to act as routers of
the data collected by the sensors in Ai,j to the sink. Collec-
tively, these sensors are the support sensors of task T .
It is, perhaps, of interest to describe the process by which
the sensors associated with T are selected. To begin, dur-
ing a time interval of length
the sink will issue a call for
work specifying the identity j of the wedge in which the task
is to be performed, as well as the identity i of the corona in
which data is to be collected. The sensor nodes in wedge j
that happen to wake up during the interval
and that have
an appropriate energy level stay awake and will participate in
the task either as either data collectors or as routers depending
on their respective position within the wedge. It is intuitively
clear that by knowing the number of sensors, the density of
deployment and the expected value of sleep periods, one can
fine-tune
in such a way that a suitable number of routers
will be awake in wedge j in support of T . Likewise, we can
select the set D of data collecting sensors in Ai,j . Let S de-
note the set of support sensors for T . It is appropriate to recall
that a by-product of the call for work is that all the sensors in
S are synchronized. In order to make the task secure the sen-
sors in S will share a secret key that allows them access to
a set of time epochs, a set of frequencies to be used in each
time epoch, and a hopping sequence to be used within each
epoch. For details we refer the reader to the description of the
randomized frequency hopping security framework proposed
in section 2.
Assume that the results of the data collection specific to
task T can be partitioned into 2m (m
0), disjoint groups.
Thus, each sensor performing data collection will encode its
data in a string of m bits.
Since, typically, D contains a large number of sensors, it
is important to fuse individual results into a final result that
will be sent to the sink node. We now outline two possible
solutions to the data fusion problem. Using the algorithm of
Nakano and Olariu [26] that does not require sensors to have
identities, the sensors in D acquire temporary identities rang-
ing from 1 to |D|. Using their newly acquired identities, indi-
Table 1
Illustrating temperature ranges and their encoding.
Temp 5160 6170 7180 8190 91100 101110 111115 116120
Code 0010 0011 0100 0101 0110 0111 1000
WADAA ET AL.
vidual data values are being transmitted to the sensor whose
identity is 1 who will perform data fusion and will send the
final result to the sink node as discussed in section 7. The
advantage of this data fusion scheme is that there is no data
loss and all the collected values will be correctly fused. There
are, however, many disadvantages. For one thing, the initial-
ization algorithm of [26] requires every sensor in D to expend
an amount of energy proportional with log |D|. For another,
the final result of the data collection is concentrated in a sin-
gle sensor (i.e., the sensor with temporary identity 1), who is
a single point of failure.
We now propose a much simpler data fusion scheme that
involves some data loss but that is fault tolerant and does not
require the sensors in D to have unique identities. The idea
is that the sensors in D transmit the data collected bit by bit
starting, say, left to right as follows: a value of 0 is not trans-
mitted, while a 1 will be transmitted. The sensors in Ai1,j
that have been elected as routers in support of transaction T
pick up the values transmitted. The following disambiguation
scheme is used:
No bit is received in this case a 0 is recorded;
A bit of 1 is received in this case a 1 is recorded;
A collision is recorded in this case a 1 is recorded.
It is clear that as a result of this disambiguation scheme,
every sensor in Ai1,j that is in support of T stores the logical
OR of the values stored by sensors in D. Note also that while
there was loss of information in the process of fusing data,
no further loss can occur in traversing the path from Ai1,j
to the sink: this is because all routers in Ai1,j transmit the
same bit string.
4.3. An example
For an example of data fusion consider a sensor network that
is tasked to monitor and report the temperature in cluster Ai,j .
Referring to table 1, for the application at hand temperatures
below 111 F are considered to be non-critical and if such a
temperature is reported no specific action is to be taken. By
contrast, temperatures above 111 F are considered to be criti-
cal and they trigger a further monitoring action. The encoding
featured in table 1 is specifically designed to reflects the rela-
tive importance of various temperature ranges. For example,
the temperature ranges in the non-critical zone are twice as
large as those in the critical zone. Also, notice that the left-
most bit differentiates critical from non-critical temperatures.
Thus, if the sink nodes receives a reported temperature whose
leftmost bit is a 1, then further action is initiated; if, on the
other hand, the leftmost bit is 0, then no special action is nec-
essary.
121125 126130 131135 136140 141145 146150
1001 1010 1011 1100 1101 1110 1111
TRAINING A WIRELESS SENSOR NETWORK
Figure 8. Illustrating data fusion: (a) trading loss in data aggregation/re-
porting for energy; (b) trading energy for lossless data aggregation/reporting.
Let us see how our data fusion works in this context. Re-
ferring to figure 8(a) assume that a group of three sensors d0 ,
d1 , and d2 in Ai,j have collected data and are about to transmit
it to the sensors s0 and s1 in Ai1,j . The values collected are
encoded, respectively, as 0110, 0101 and 0110. Thus, none of
the values indicates a critical situation. After transmission and
disambiguation, the sensors in Ai1,j will store 0111 which is
the logical OR of the values transmitted. Notice that although
the data fusion process involves loss of information, we do
not loose critical information. This is because the logical OR
of non-critical temperatures must remain non-critical. Con-
versely, if the logical OR indicates a critical temperature, one
of the fused temperatures must have been critical and thus ac-
tion must be initiated. It is also interesting to note that when
the sensors in Ai1,j transmit to those in Ai2,j no further
loss of information occurs.
There is an interesting interplay between the amount of
loss in data aggregation (fusion) and the amount of energy
expended to effect it. As we are about to show, if we are will-
ing to expend slightly more energy, lossless data aggregation
can be achieved.
The corresponding tradeoff is interesting in its own right
being characteristic of choices that present themselves in the
design of protocols for wireless sensor networks. For illus-
tration purposes, assume that it is necessary to determine the
maximum of the bit codes stored by the sensors in Ai,j .
To solve this problem, all the sensors in Ai,j that have col-
lected relevant information engage in the following protocol
that is guaranteed to aggregate the values into the maximum.
Assume that each sensor stores an n-bit code for the range.
Starting with the highest significant bit to the lowest:
1. Sensors in Ai,j that have a 0 in position p listen for two
time slots; if in any of these slots a 1 or a collision mes-
sage is received, they terminate their participation in the
protocol.
2. Sensors that have a 1 in position d transmit in the first time
slot and sleep in the second.
159
3. Sensors in Ai1,j do the following:
3.1. Any sensor that has received a 1 or a collision in the
first slot, echoes a 1 in the second.
3.2. Any sensor that has not received a transmission in the
first slot sleeps in the second slot.
Figure 8(b) illustrates how the maximum of the values col-
lected by sensors d0 , d1 , and d2 in Ai,j is correctly communi-
cated to the support sensors s0 , and s1 in Ai1,j . In this case,
we assume d0 , and d1 are not in direct communication range
of each other. Note that s0 receives a collision corresponding
to the third most significant bit; consequently it echoes a 1,
thereby enabling d1 to terminate the protocol. Similarly, s1
receives a collision, and echoes a 1 for the same bit position
(not shown in the figure). It is easy to confirm that by exploit-
ing the associatively of the maximum, the simple protocol that
we just outlined correctly forwards to the sink the maximum
of the values stored by sensors in Ai,j .
5. Our lightweight training protocol
Our proposed model for a sensor network assumes that after
deployment the sensor nodes must be trained before they can
be operational in the network. Recall that sensor nodes do
not have identities and are initially unaware of their location.
It follows that untrained nodes are not addressable and can-
not be targeted to do work in the network. The main goal of
this section is to present, in full detail, our lightweight highly
scalable training protocol for wireless sensor networks. The
key advantage of this protocol is that each node participating
in the training incurs an energy cost that is logarithmic in the
number of clusters and wedges defined by the protocol. Being
energy efficient, this training can be repeated on a scheduled
or ad-hoc basis providing robustness and dynamic reorgani-
zation.
After deployment nodes sleep until wakened by their in-
dividual timers. Thus, each node sleeps for a random period
of time, wakes up briefly and if it hears no messages of in-
terest, selects a random number x and returns to sleep x time
units. Clocks are not synchronized but over any time inter-
val [t, t +
t] a percentage directly proportional to
t of the
nodes are expected to wake up briefly. During this time inter-
val the sink continuously repeats a call to training specifying
the current time and a rendezvous time. Thus, in a probabilis-
tic sense a certain percentage of nodes will be selected for
training. The time interval
t can be adjusted to control the
percentage of nodes that are selected. Using the synchroniza-
tion protocol we describe in section 5.1 the selected sensors
nodes reset their clocks and set their timer appropriately be-
fore returning to sleep.
5.1. The synchronization protocol
It is natural to assume that, just prior to deployment, the sen-
sor nodes are synchronized. However, due to natural clock
160 WADAA ET AL.

drift, re-synchronization is necessary. Re-synchronization is

done with respect to the master clock running at the sink.
Suppose that the sink dwells micro-seconds on each fre-
quency in the hopping sequence. For the purpose of show-
ing how synchronization is effected, assume that time is ruled
into epochs as discussed before. For every i (i
1), we let li
stand for ti / ; thus, epoch ti involves a hopping sequence
of length li . We can think of the epoch ti as being partitioned
into li slots, each slots using its own frequency selected by
virtue of the hopping sequence out of the set ni of frequen-
cies associated with epoch ti . It is clear that determining the
epoch and the position of the sink in the hopping sequence
corresponding to the epoch is sufficient for synchronization.
Our synchronization protocol is predicated on the assump-
Figure 9. Illustrating corona training.
tion that clock drift is bounded. Specifically, assume that
whenever a sensor node wakes up during its local time epoch time ruled into slots s1 , s2 , . . . , sk1 and that the sensors can
ti the master clock is in one of the time epochs ti1 , ti , or ti+1 . synchronize2 to the master clock running at the sink node.
Using its genetic information, the sensor node knows the In time slot s1 all the sensors are awake and the sink trans-
last frequencies i1 , i and i+1 on which the sink will mits with a power level corresponding to rk/2 . In other words,
dwell in the time epochs ti1 , ti , and ti+1 , respectively. Its in the first slot the sensors in the first k/2 coronas will re-
strategy, therefore, is to tune in, cyclically, to these frequen- ceive the message above a certain threshold, while the others
cies, spending /3 time units on each of them. It is clear that, will not. Accordingly, the sensors that receive the signal set
eventually, the sensor node meets the sink node on one of the- b1 = 0, the others set b1 = 1.
ses frequencies. Assume, without loss of generality, that the Consider a k-leaf binary tree T and refer to figure 9. In the
node meets the sink on frequency in some (unknown) slot s figure the leaves are numbered left to right from 1 to k. The
of one of the epochs ti1 , ti , or ti+1 . To verify the synchro-
edges of T are labeled by 0s and 1s in such a way that an
nization, the node will attempt to meet the sink in slots s + 1,
edge leading to a left subtree is labeled by a 0 and an edge
s + 2 and s + 3 at the start of the next epoch. If a match is
leading to a right subtree is labeled by a 1. Let l (1  l  k),
found, the node declares itself synchronized. Otherwise, the
be an arbitrary leaf and let b1 , b2 , . . . , blog k be the edge labels
node will repeat the above process.
of the unique path leading from the root to l. It is both well
We note that even if the sensor node declares itself syn-
known and easy to prove by a standard inductive argument
chronized with the sink, there is a slight chance that, it is not.
that
The fact that the node has not synchronized will be discov-
ered quickly and it will again attempt to synchronize. There

log k

are ways in which we can make the synchronization proto- l =1+ bj 2log kj (1)
col deterministic. For example, the hopping sequence can be j =1
designed in such a way that the last frequency in each epoch (for example, applying equation (1) to leaf 7 we have: 7 =
is unique and it is not used elsewhere in the epoch. How- 1 + 0 23 + 1 22 + 1 21 + 0 20 ).
ever, this entails less flexibility in the design of the hopping Referring again to figure 9, let the interior nodes of the
sequence and constitutes, in fact, an instance of a differential tree be numbered in preorder from 1 to k 1 and let T  be the
security service where the level of security is tailored to suit tree consisting of the interior nodes only. Let u be an arbitrary
the application or the power budget available. node in T  , and let b1 , b2 , . . . , bi1 be the edge labels on the
unique path from the root to u. We take note of the following
5.2. The corona training protocol technical result.
The main goal of this subsection is to present the details of
the corona training protocol. The wedge training protocol is Lemma 5.1. Let p(u) be the preorder number of u in T  .
similar and will not be discussed further. Then, we have
Let k be an integer1 known to the sensor nodes and let the

i1
k coronas be determined by concentric circles of radii r1 < p(u) = 1 + cj ,
r2 < < rk centered at the sink node. j =1
The idea of the corona training protocol is for each individ-
where
ual sensor node to learn the identity of the corona to which it
belongs. For this purpose, each individual sensor node learns 1 if bj = 0,
a string of log k bits from which the corona number can be de- cj = k
if bj = 1.
termined easily. To see how this is done, it is useful to assume 2j
1 For simplicity we shall assume that k is a power of two. 2 See section 5.1.
TRAINING A WIRELESS SENSOR NETWORK
Proof. The proof is by induction on the depth i of node u in
T  . To settle the basis, note that for i = 1, u must be the root
and p(u) = 1, as expected.
For the inductive step, assume the statement true for all
nodes in T  of depth less that u. Indeed, let v be the parent of
u and consider the unique path of length i 1 joining the root
to u. Clearly, nodes u and v share b1 , b2 , . . . , bi2 and, thus,
c1 , c2 , . . . , ci2 . By the inductive hypothesis,


i2
p(v) = 1 + cj . (2)
j =1
On the other hand, since v is the parent of u, we can write
1 if u is the left child of v,
p(u) = p(v) + k (3)
otherwise.
2i1
Notice that if u is the left child of v we have bi1 = 0 and
ci1 = 1; otherwise bi1 = 1 and ci1 = k/2i1 . This
observation, along with (2) and (3) combined, allows us to
write


i2

i1
p(u) = 1 + cj + ci1 = 1 + cj
j =1 j =1
completing the proof of the lemma.  ers will sleep minimizing the power expenditure. Yet another
For further reference we also need the following technical
result.
Lemma 5.2. Let u be an arbitrary node of the tree T  and let
n(u) denote its inorder number in T  . Let m be the left-to-
right rank among the leaves of T of the rightmost leaf of the
left subtree of T rooted at u. Then, n(u) = m.
Proof. We proceed by induction on the inorder number of
a node in T  . Indeed, if n(u) = 1, then u must be the left-
most leaf in T  and, thus, its left subtree in T consists of the
leftmost leaf of T t, settling the base case.
Assume that the statement true for all nodes of T  with
inorder number smaller than that of u. we shall distinguish
between the following two cases.
Case 1. v is an ancestor of u in T  . Let T  (v) be the subtree
of T  rooted at v. In this case u must be the leftmost leaf in the
right subtree of T  (v). Let q be the left-to-right rank among
the leaves of T of the rightmost leaf of the left subtree of
T  (v). By the inductive hypothesis n(v) = q. Since u is a
leaf in T  it has exactly two children in T , namely the leaves
of ranks q +1 and q +2. Thus, in this case, n(u) = n(v)+1 =
q + 1, as claimed.
Case 2. u is an ancestor of v in T  . Let T  (u) be the subtree
of T  rooted at u. In this case v must be the rightmost leaf in
the left subtree of T  (u). Assume that n(v) = r. Observe that
v has exactly two leaf children T . By the induction hypothesis
these children have ranks r and r + 1. Thus, in this case,
n(u) = n(v) + 1 = r + 1, as claimed.
This completes the proof of lemma 5.2.
161
With these technicalities out of the way, we now return to
the corona training protocol. In our setting, the preorder and
inorder numbers of internal nodes in T correspond, respec-
tively, to time slots in the training protocol and to the trans-
mission ranges used by the sink. More precisely, consider an
arbitrary integer i (2  i  log k 1), and assume that at
the end of time slot s a sensor node has learned the leftmost
i 1 bits b1 , b2 , . . . , bi1 . The following important result is
implied by lemma 5.1 and lemma 5.2.
Corollary 5.3. Having learned bits b1 , b2 , . . . , bi1 a sensor
node must wake up in time slot z = 1 + i1 j =1 cj to learn bit
bi . Moreover in time slot z the sink node uses a transmission
range of rinorder(z) .
(To illustrate corollary 5.3, refer again to figure 9 where the
internal nodes are labeled by their preorder numbers. Con-
sider the node labeled 2. It is easy to verify that its inorder
number is 4. Thus, all the nodes in the subtree rooted at 2 will
be awake in slot 2 and the sink node will transmit with a range
of r4 . Consequently, the sensor nodes at a distance from the
sink not exceeding r4 will receive the signal, while the others
will not.)
It is also worth noting that only the sensor nodes that need
to be awake in a given time slot will stay awake, the oth-
interesting feature of the training protocol we just described
is that individual sensor nodes sleep for as many contiguous
slots as possible before waking up, thus avoiding repeated
wakesleep transitions that are known to waste energy.
Securing the training protocol is especially important since
training is a prerequisite for subsequent network operations.
As argued in [22] our parameterized frequency hopping
scheme guarantees that the physical layer of wireless com-
munications is secure.
At the same time, in case the corona training process has
to be aborted before it is complete, corollary 5.3 guarantees
that if the training process re-starts at some later point, every
sensor node knows the exact time slots when it has to wake
up in order to learn its missing bits.
6. Reasoning about the longevity of the sensor network
The main goal of this section is to explore the energy require-
ments of the sensor network in terms of a model of work. In-
deed, we adopt a transaction-based model whereby the sensor
network is subjected to a set T of transactions. Each transac-
tion involves the nodes in a sector (i.e., a cluster) and involves
performing local sensing by the sensors, data fusion and send-
ing the resulting information to the sink. Recall that, as dis-
cussed in section 4.1, one of the key benefits of our training
is that transmitting the result of the transaction from a sector
to the sink node amounts to routing the information along a
path lying within the same wedge (see also figure 7). Thus,
we associate each transaction with such a path. We will now
analyze the energy expended by sensor nodes to fulfill their
 path-related duties.
162 WADAA ET AL.

the condition (ri + ri2 )/2
ri1 is equivalent to ri

ri1
ri1 ri2 , confirming our intuitive idea that
corona widths are non-decreasing.

Similarly, the condition ri1
ri ri2 implies that
rk rk1 ri ri1 r2
     
rk1 rk2 ri1 ri2 r1
which can be interpreted as limiting the growth of consec-
utive coronas.
Let n denote the total number of sensor nodes deployed in
wedge W . We assume a uniform deployment with density .
In particular, with A standing for the area of wedge W , we
can write
2
n = A = r . (5)
2 k
Let n1 , n2 , n3 , . . . , nk stand for the number of nodes de-
ployed in the sectors A1 , A2 , A3 , . . . , Ak , respectively. Since
the deployment is uniform, it is easy to confirm that for
every i (1  i  k),
 2 
Figure 10. Illustrating a wedge W and the associated sectors. ni = Ai = ri ri1
2
. (6)
2
Let N denote the number of sector-to-sink paths (hence-
Table 2 forth, simply denoted by paths) that the wedge W sees during
Summary of system parameters.
the lifetime of the sensor network. By our previous discus-
Parameter Description sion there is a one-to-one map between paths and transac-
rk Radius of circle of deployment
tions. Thus, N equals the total number T of transactions that
Deployment density the wedge can handle during the lifetime of the network.
tx Maximum transmission range of a sensor node We make the following assumptions motivated by the uni-
E Total energy budget packed by a sensor node formity of the deployment:
Angle subtended by wedge W
n Number of sensor nodes in wedge W each sensor node in W is equally likely to be the source of
N Number of sector-to-sink paths that W sees during the lifetime a path to the sink;
of the network for 2  i  k, each sensor in sector Ai1 is equally likely
T Total number of transactions that W can handle during the
to serve as the next hop for a path that involves a node
lifetime of the network
in Ai .
By virtue of the first assumption, the expected number of
paths originating at a node in W is
Throughout the remainder of this work we assume a sen-
sor network deployed in a circular area and a co-located sink N
. (7)
node placed at its center. Consider a wedge W subtended by n
an angle of and refer to figure 10. W is partitioned into k Consider sector A1 . Since the N paths have the sink node
sectors A1 , A2 , . . . , Ak by its intersection with k concentric as their destination, the nodes in sector A1 must collectively
circles, centered at the sink node, and of monotonically in- participate in all the N paths. Since A1 contains n1 nodes, the
creasing radii r1 < r2 < < rk . It is important to note that expected number of transmissions per node is N/n1 . Assum-
rk , the deployment radius, is a system parameter and, thus, ing a quadratic power degradation factor, the energy expended
a constant for a particular sensor network. The system para- by a node in A1 per path served is cr12 for some positive con-
meters are summarized in table 2. stant c. Thus, the total energy E1 consumed by a node in A1
For convenience of notation we write r0 = 0 and interpret to fulfill its path-related duties is
A0 as the sink node itself. We assume the following regularity
N 2
condition: E1 = cr
n1 1
ri + ri2
for all i, 2  i  k,
ri1
ri ri2 (4) which, by (6), can be written as
2
which, essentially, specifies the way coronas relate to each N 2 2Nc 2 2Nc
E1 = cr1 = r1 = . (8)
other. Specifically, n1 r12
TRAINING A WIRELESS SENSOR NETWORK 163

2 r2 )

Quite surprisingly, (8) asserts that the total energy ex- r 2 + (r22 r12 ) + (r32 r22 ) + + (ri1
N 1 1 i2
pended by a node in A1 is independent of the value of r1 . Let rk2
T denote the total number of transactions performed by the

entire sensor network (not just wedge W ) during its lifetime r2

= N 1 i1 . (13)
and let N be the corresponding number of node-to-sink paths. rk2
Assuming that the T transactions are uniformly distributed
throughout the sensor network, we can write Recall that sector Ai contains ni nodes. This implies that each
node in Ai must participate in
N N

= . (9) N r2
2 1 i1
ni rk2
By (8) and (9) combined, the total energy needed by a node
in A1 to handle its path-related duties is paths. Using (6), the number of paths handled by each node
in Ai can be written as
2Nc Nc 2

E1 = = . (10) 2N ri1 1
1 2 . (14)
rk ri ri1
2 2

Let E denote the total energy budget of a sensor node. Since

the sensor nodes in A1 must have sufficient energy to handle
their path-related duties, by using (10) we can write
Nc
< E.
Recalling that in our work model there is a one-to-one cor-
respondence between transactions and sector-to-sink paths,
this inequality can be written in its equivalent form
Observe that the width of sector Ai is ri ri1 . It follows
that the transmission range needed to send information be-
tween Ai and Ai1 is ri ri1 . Thus, in a quadratic power
degradation model, we shall assume that the energy expended
by a node in Ai to send information to sensors in Ai1 is
c(ri ri1 )2 .
Let the total amount of energy expended by a node in Ai be
Ei . By (9) and (14), we have

Tc Nc r2 1
< E. (11) Ei = 1 i1 (ri ri1 )2 .
rk2 ri2 ri1
2

Equation (11) tells us that for a given energy budget E, in Simple manipulations show that
order to guarantee a network longevity of T transactions, the 2

deployment density must satisfy the inequality Nc ri1 ri ri1

Ei = 1 2 . (15)
rk ri + ri1
Tc
> . (12) Observing that
E
2
ri1
1 <1
7. Energy constraints rk2
and that
We now turn to the task of evaluating the energy expenditure ri ri1
per node in an arbitrary sector Ai with i
1. Since the case <1
ri + ri1
i = 1 was handled in the previous section, we now assume
i
2. Observe that nodes in a generic sector Ai (2  i  k), equation (15) implies that
are called upon to serve two kinds of paths:
Nc
Ei < . (16)
paths originating in a sector Aj with i < j  k; and
paths originating at a node in Ai . Inequality (16) captures a very important conclusion: the
energy expended by a sensor node in sector Ai to fulfill its
It is easy to confirm that the number of paths involving nodes path-related duties is strictly less than the energy expended
in Ai includes all paths except those originating in one of by a node in sector A1 . More generally, it is instructive to
the sectors A1 , A2 , . . . , Ai1 . Therefore, the total number of compare the energy expended for path-related duties by in-
paths that the nodes in Ai must handle is dividual sensor nodes in sectors Ai and Ai1 . Intuition sug-
N gests that Ei < Ei1 . This is because as we move away
N (n1 + n2 + + ni1 ). from the sink node, by the regularity condition (4) the sec-
n
tors become larger and larger and, as a result, more energy is
By (5) and (6) combined with elementary manipulations, this needed for intra-sector communication and governance tasks.
expression can be written as This implies that less energy should be spent on path-related
164 WADAA ET AL.

duties. The next result shows that, in fact, this intuition is For every j (1  j  i), write aj = rj rj 1 and bj = 1.
correct. Noticing that
i
p=1 ap = Ei /c,
2
Theorem 7.1. Consider a sequence of radii r1 < r2 < i
< rk satisfying the regularity condition. Then for every i, p=1 bp = i,
2

2  i  k, Ei < Ei1 . 
( ip=1 ap )2 = ri2 ,
Proof. Recall that by (15) the energy Ei spent by each node and substituting in Lagranges identity, we obtain
in Ai is  i 2


i

2

iEi
Nc ri1 ri ri1 (ap aq )2 = i ap2 ap = ri2 .
Ei = 1 2 . c
rk ri + ri1 1
p<q
i p=1 p=1

Likewise, the energy Ei1 spent by each node in Ai1 is Thus, we can write



c 2
Nc r2 ri1 ri2 Ei = r + (ap aq )2 . (18)
Ei1 = 1 i2 . i i
rk ri1 + ri2
2 1
p<q
i

Since ri2 < ri1 it is clear that Clearly, the left-hand side of the above equality is minimized
whenever
2
ri1 2
ri2

1 <1 . (ap aq )2 = 0
rk2 rk2 1
p<q
i
Moreover, in order to show that which occurs if and only if
ri ri1 ri1 ri2
 a1 = a2 = a3 = = ai .
ri + ri1 ri1 + ri2
we only need show that Thus, for some positive constant d we have

ri ri1 ri1 ri2 for every j (1  j  i), rj rj 1 = d. (19)

1
1
ri + ri1 ri1 + ri2 It is worth noting that equation (19) satisfies the regularity
or, equivalently, condition (4). It is easy to see that equation (19) implies
ri2

ri1 ri = id (20)
ri1 + ri2 ri + ri1
and so, substituting in (18) we obtain
2
which boils down to ri1
ri ri2 , which is implied directly
by our regularity condition. This completes the proof of the Ei = icd 2 .
theorem.  Let tx be the maximum transmission range of a sensor node.
The way our coronas are organized, suggests setting
8. Crafting the coronas d = tx

The main goal of this section is to show how to select the radii yielding
r1 , r2 , . . . , rk in such a way that the energy spent per sector-  
Ei = i ctx2 . (21)
to-sink path is minimized. For this purpose, let Ei denote the
total amount of energy expended by the nodes along a generic To summarize, in order to minimize the total amount of
path transferring data from sector Ai to the sink. Write r0 = 0 energy spent on a path originating at a sensor node in corona
and assume that A0 is the sink node itself; since in transmit- Ai down to the sink node, all the coronas must have the same
ting from Aj to Aj 1 (2  j  i), the amount of energy width tx and the optimal amount of energy is i times the en-
spent is c(rj rj 1 )2 , it follows that ergy needed to send the desired information between adjacent
coronas.


i
Ei = c (rj rj 1 )2 . (17) 8.1. A synopsis of power expenditure
j =1

Recall the Lagrange identity [17, p. 64] Recall that by equation (10) and theorem 7.1, for every choice
 2 of radii r1 < r2 < < rk subject to the regularity condition,



i

i

i
we have the following energy monotonicity property:
(ap bq aq bp ) =
2
ap2 bp2 ap bp .
1
p<q
i p=1 p=1 p=1 E1 > E2 > > Ek .
TRAINING A WIRELESS SENSOR NETWORK
Table 3
Illustrating various energy ratios.
Corona Energy ratio
2 0.3333 . . .
3 0.2
4 0.1428 . . .
5 0.1111 . . .
6 0.0909 . . .
7 0.0769 . . .
Figure 11. Illustrating the energy ratio Ei /E1 .
It is instructive to get a feel for how fast the energy expen-
ditures drop as we move away from the sink node. For this
purpose, we rewrite equations (10) and (15) as
2
ri1 ri ri1 ri ri1
Ei = E1 1 2 < E1 .
rk ri + ri1 ri + ri1
By using (20), the inequality above can be written as
Ei 1 Acknowledgements
< .
E1 2i 1
Table 3 summarizes the energy ratio Ei /E1 for various values
of i.
As shown by table 3 and also in the companion figure 11,
the energy expended by a sensor node in support of its path-
related obligations drops significantly as we move away from
the sink. For example, in the 6th corona, the power expended
by a node is less than 10% of the energy expended by a node
in the first corona. This leaves plenty of energy for other
tasks.
9. Concluding remarks
In this work we have proposed a virtual infrastructure a dy-
namic coordinate system for a massively-deployed collec-
tion of anonymous sensor nodes. This coordinate system
provides, at no extra cost, an interesting clustering scheme
according to which two nodes are in the same cluster only if
they have the same coordinates. Notice that this clustering
scheme works for anonymous sensor nodes. As a corollary,
165
sensor nodes do not know the identity of the other nodes in the
same cluster. Our second contribution was to show that train-
ing the sensor nodes the process of learning their coordi-
nates can be performed by a protocol that is at the same time
lightweight and secure. Being energy efficient, this training
can be repeated on either a scheduled or ad-hoc basis to pro-
vide robustness and dynamic reorganization. We also showed
that in a trained wireless sensor network the tasks of routing
and data fusion can be performed by very simple and energy-
efficient protocols. Finally, we showed how to design the co-
ordinate system such as to minimize the power expended in
collecting and routing data.
In this paper we addressed the problem of training a sensor
network in a two-dimensional plane. In practice, however, the
network training problem manifests itself in three dimensions,
for example, because of irregularities in a rugged deployment
terrain. To extend our work we have developed solutions for
the three-dimensional training problem where the majority of
the nodes are assumed to reside in one logical base plane,
while the remaining nodes are dispersed over other parallel
planes. The goal was to mimic the case of minor terrain ir-
regularities. However, training a sensor network of nodes ar-
bitrarily dispersed in a three-dimensional space remains an
open problem.
In spite of these encouraging results, more work has to be
done. We are currently looking at security-related problem
in wireless sensor networks. This is an extremely important
problem as the information provided by the sensor network
may be used for decision making in military or civilian en-
vironments where human life is at stake. Looking at various
issues related to securing sensor networks promises to be an
exciting are for future work.
The authors wish the thanks three anonymous referees for
their constructive comments that have lead to an improved
presentation. A preliminary version of this paper [39] has
appeared in Proceedings of 3rd International Workshop on
Wireless, Mobile and Ad Hoc Networks (WMAN03), Nice,
France, April 2003.
Appendix
Our proposed training protocol consists of a call for training
protocol, and a corona and wedge training protocol.
A.1. Call for training protocol
This protocol enables the sink node to issue calls for training
to a random subset of nodes of an arbitrary size post deploy-
ment. Each call communicates the next rendezvous time for
corona and wedge training along with the local time at the
sink node. We use the notations from table 4.
166 WADAA ET AL.

Table 4
used to calibrate the time attributes in the message received by
Parameters of the training pseudocode.
the sensor node to account for propagation and computation
Parameter Description delays.
k Total number of coronas in the coordinate system
w Total number of wedges in the coordinate system
rT The sink commences transmission of training mes-
sages at this time A.2. Corona and wedge training protocol
clock Local clock on board the sensor node
clock.set(t) Set the local clock to time t
Since training for wedges uses the same protocol used in
time() Returns the local clock time in a sensor node or the
sink node training for coronas in an analogous way, we limit our dis-
Timer A local timer in a sensor node or the sink node cussion here to corona training.
Timer.set(t) Set the timer to expire at time t
sleep() Sleep until a timer interrupt occurs
radio The radio device on board a sensor node Corona training (trainer (sink node))
radio.receive() Put radio in receive mode for one message time if a int cTrainThem(int k){
message is being transmitted it return the message for (i = 1; i < k; + + i){
X An ordered tuple (X(i) is the ith component of the m = message((i))
tuple) radius = R(inorder(k, i))
message(X) Construct a message that has ordered tuple X as its radio.transmit(radius, m)
contents
radio.transmit(r, m) Transmit message m, using a transmission radius of r
}

Time interval during which calls for training mes- }
sages are transmitted
The time slot size. For simplicity, we assume that is
also the time required to transmit (receive) a protocol Corona training (trainee (sensor node))
message int cTrainMe(int k){
Timer.interrupt() Blocks until a timer interrupt occurs start = time()
R(i) Radius of concentric circle i, 1  i  k messageNumber = 1
inorder(k, i) Inorder number for the node in a full binary tree of
coronaNumber = 0
height log k whose preorder number is i
bitCounter = 0
offset = k/2

while (bitCounter < log k){

Call for training (trainer (sink node))
if(m = radio.receive()){
CFTtrainer(
){
(coronaNumber  1)+ = 0
deadline = time() +

+ + messageNumber
while (time() < deadline){
offset/ = 2
m = message((time(), rT , k, w))
+ + bitCounter
radio.transmit(rk , m)
Timer.set(start + (messageNumber 1) )
}
Timer.interrupt()
}
}
else{
Call for training (trainee (sensor node)) (coronaNumber  1)+ = 1
CFTtrainee(
){ messageNumber+ = offset
while (true){ offset/ = 2
Timer.set(time() + random()) + + bitCounter
sleep() Timer.set(start + (messageNumber 1) )
if (m = radio.receive()){ sleep()
clock.set(m.X(1) + ) }
Timer.set(m.X(2)) }
sleep() return (coronaNumber)
cTrainMe() }
}
} It should be noted that a sensor node remains awake and
} tunes in to receive the protocol message j + 1, if it succeeds
in receiving message j . If the node tunes in but fails to receive
Note that cTrainMe() invokes the corona training protocol message j , the node sleeps until the time of the next message
(sensor node side) that is described below. The parameter is to which it must listen.
TRAINING A WIRELESS SENSOR NETWORK
References
[1] J. Agre and L. Clare, An integrated architecture for cooperative sensing
networks, IEEE Computer 33(5) (2000) 106108.
[2] F. Akyildiz, W. Su, Y. Sankarasubramanian and E. Cayirci, Wireless
sensor networks: A survey, Computer Networks 38(4) (2002) 393422.
[3] R. Anderson, Security Engineering: A Guide to Building Dependable
Distributed Systems (Wiley, New York, 2001).
[4] R. Anderson and M. Kuhn, Tamper resistance a cautionary note,
in: Proc. 2nd USENIX Workshop on Electronic Commerce, Berkeley,
CA (1996), pp. 111. IEEE Wireless Communications 9(1) (2002) 40
48.
[5] S. Bandyopadhyay and E. Coyle, An efficient hierarchical clustering
algorithm for wireless sensor networks, in: Proc. INFOCOM2003, San
Francisco, CA (April 2003).
[6] D. Braginsky and D. Estrin, Rumor routing algorithm for sensor net-
works, in: Proc. Internat. Conf. on Distributed Computing Systems
(ICDCS-22) (November 2001).
[7] N. Bulusu, J. Heidemann and D. Estrin, GPS-less low cost outdoor lo-
calization for very small devices, IEEE Personal Communications 7(5)
(2000) 2834.
[8] N. Bulusu, J. Heidemann and D. Estrin, Scalable coordination for
wireless sensor networks: self-configuration localization systems, in:
Proc. 6th Internat. Sympos. on Communication Theory and Applica-
tions (ISCTA-2001) (July 2001).
[9] S. Capkun, M. Hamdi and J.-P. Hubeaux, GPS-free positioning in mo-
bile ad-hoc networks, Cluster Computing 5(2) (2002) 157167.
[10] D.W. Carman, P.S. Kruus and B.J. Matt, Constraints and approaches
for distributed sensor network security, Technical Report 00-010, NAI
Labs (2000).
[11] D. Coore, R. Nagpal and R. Weiss, Paradigms for structure in an amor-
phous computer, MIT Artificial Intelligence Laboratory Technical Re-
port AI-1616 (October 1997).
[12] L. Doherty, H.S.J. Pister and L.E. Ghaoui, Convex position estimation
in wireless sensor networks, in: Proc. INFOCOM2001, Anchorage,
AK (April 2001).
[13] A. Ephremides, J. Wieselthier and D. Baker, A design concept for re-
liable mobile radio networks with frequency hopping signaling, Pro-
ceedings of the IEEE 75(1) (1987) 5673.
[14] D. Ganesan, R. Govindan, S. Shenker and D. Estrin, Highly resilient,
energy-efficient multipath routing in wireless sensor networks, ACM
Mobile Computing and Communications Review 5(4) (2001).
[15] S. Ghiasi, A. Srivastava, X. Yang and M. Sarrafzadeh, Optimal energy-
aware clustering in sensor networks, Sensors 2 (2002) 258269.
[16] L. Girod, V. Bychkovskiy, J. Elson and D. Estrin, Locating tiny sensors
in time and space: A case study, in: Proc. International Conference on
Computer Design (ICCD 2002), Freiburg, Germany (September 2002).
[17] R.G. Graham, D.E. Knuth and O. Patashnik, Concrete Mathematics
(Addison-Wesley, New York, 1989).
[18] http://www.darpa.mil/mto/mems/
[19] http://www.stanford.edu/class/ee321/ho/MEMS-14-
sensors.pdf
[20] http://www.xs4all.nl/ganswijk/chipdir/m/sensor.
htm
[21] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion:
A scalable and robust communication paradigm for sensor networks,
in: Proc. MOBICOM00, Boston, MA (August 2000).
[22] K. Jones, A. Wadaa, S. Olariu, L. Wilson and M. Eltoweissy, Towards a
new paradigm for securing wireless sensor networks, in: Proc. New Se-
curity Paradigms Workshop (NSPW2003), Ascona, Switzerland (Au-
gust 2003).
[23] J.M. Kahn, R.H. Katz and K.S.J. Pister, Mobile networking for Smart
Dust, in: Proc. MOBICOM99, Seattle, WA, August 1719 (1999).
[24] J. Kulik, W. Heinzelman and H. Balakrishnan, Negotiation-based pro-
tocols for disseminating information in wireless sensor networks, Wire-
less Networks 8(3) (2002).
[25] K. Langendoen and N. Reijers, Distributed localization in wireless sen-
sor networks, a quantitative comparison, Manuscript (May 2003).
167
[26] K. Nakano and S. Olariu, Randomized initialization protocols for radio
networks, in: Handbook of Wireless Networks and Mobile Computing,
ed. I. Stojmenovic (Wiley, New York, 2002) pp. 195218.
[27] K. Nakano and S. Olariu, Uniform leader election for radio networks,
IEEE Transactions on Parallel and Distributed Systems 13 (2002) 516
526.
[28] National Research Council, Embedded, Everywhere: A Research
Agenda for Systems of Embedded Computers, Committee on Net-
worked Systems of Embedded Computers, for the Computer Science
and Telecommunications Board, Division on Engineering and Physical
Sciences, Washington, DC (2001).
[29] D. Niculescu and B. Nath, Ad-hoc positioning system, in: Proc. IEEE
GLOBECOM, San Antonio, TX (November 2001) pp. 29262931.
[30] A. Perrig, R. Szewczyk, V. Wen, D. Culler and J.D. Tygar, SPINS:
Security protocols for sensor networks, in: Proc. ACM SIGMOBILE,
Rome, Italy (July 2001).
[31] G.J. Pottie and W.J. Kaiser, Wireless integrated sensor networks, Com-
munications of the ACM 43(5) (2000) 5158.
[32] P. Saffo, Sensors, the next wave of innovation, Communications of the
ACM 40(2) (1997) 9397.
[33] C. Savarese, J. Rabaey and K. Langendoen, Robust positioning al-
gorithms for distributed ad-hoc wireless sensor networks, in: Proc.
USENIX Technical Annual Conference, Monterey, CA (June 2002)
pp. 317328.
[34] R.C. Shah and J. Rabaey, Energy aware routing for low energy ad hoc
sensor networks, in: Proc. IEEE WCNC, Orlando, FL (March 2002).
[35] C.-C. Shen, C. Srisathapornphat and C. Jaikaeo, Sensor information
networking architecture and applications, IEEE Personal Communica-
tions (August 2001) 5259.
[36] E. Shih, S. Cho, N. Ickes, R. Min, A. Sinha, A. Wang and A. Chan-
drakasan, A physical layer driven protocol and algorithm design for
energy-efficient wireless sensor networks, in: Proc. MOBICOM2001,
Rome, Italy (July 2001).
[37] K. Sohrabi, J. Gao, V. Ailawadhi and G. Pottie, Protocols for self-
organization of a wireless sensor network, IEEE Personal Communi-
cations (October 2000) 1627.
[38] S. Tilak, N.B. Abu-Ghazaleh and W. Heinzelman, A taxonomy of wire-
less micro-sensor network models, Mobile Computing and Communi-
cations Review 6(2) 2836.
[39] A. Wadaa, S. Olariu, L. Wilson, K. Jones and Q. Xu, On training wire-
less sensor networks, in: Proc. 3rd International Workshop on Wireless,
Mobile and Ad Hoc Networks (WMAN03), Nice, France (April 2003).
[40] B. Warneke, M. Last, B. Leibowitz and K. Pister, SmartDust: com-
municating with a cubic-millimeter computer, IEEE Computer 34(1)
(2001) 4451.
[41] A. Woo and D.E. Culler, A transmission control scheme for media ac-
cess in sensor networks, in: Proc. MOBICOM01 (July 2001).
[42] A.D. Wood and J.A. Stankovic, Denial of service in sensor networks,
IEEE Computer 35(10) (2002) 5462.
[43] W. Ye, J. Heidemann and D. Estrin, An energy-efficient MAC protocol
for wireless sensor networks, in: Proc. INFOCOM 2002, New York
(June 2002).
[44] V.V. Zhirnov and D.J.C. Herr, New frontiers: self-assembly and nano-
electronics, IEEE Computer 34(1) (2001) 3443.
[45] J. Zyren, T. Godfrey and D. Eaton, Does frequency hopping
enhance security? http://www.packetnexus.com/docs/
20010419frequency_Hopping.pdf
Ashraf Wadaa is a Lecturer of Computer Science
and a Ph.D. candidate at Old Dominion University.
Wadaas dissertation is on wireless sensor networks.
His research interests include wireless sensor, and ad
hoc networks, network security, database systems,
and distributed computing. He published in refer-
eed journals and conference proceedings. He also
served on a number of technical committees for con-
ferences, and workshops. His funding record in-
cludes a grant from the Commonwealth Information
168
Security Center (CISC) in Virginia to pursue research in sensor network se-
curity. Wadaa earned a B.S. (1986 with top rank) in Computer Science and
Automatic Control from Alexandria University, Egypt. He was nominated by
Old Dominion University for an outstanding teacher award in 2002. Wadaa
is a member of the honor society of Phi Kappa Phi.
E-mail: wadaa@cs.odu.edu
Stephan Olariu received the M.Sc. and Ph.D. de-
grees in computer science from McGill University,
Montreal, Canada in 1983 and 1986, respectively. In
1986 he joined the Computer Science Department
at Old Dominion University where he is now a full
professor. He has published extensively in various
journals, book chapters, and conference proceedings.
His research interests include wireless networks and
mobile computing, parallel and distributed systems,
performance evaluation, and medical image process-
ing. Prof. Olariu serves on the editorial board of several archival journals
including IEEE Transactions on Parallel and Distributed Systems, Journal
of Parallel and Distributed Computing, International Journal of Foundations
of Computer Science, Journal of Supercomputing, International Journal of
Computer Mathematics, VLSI Design, and Parallel Algorithms and Applica-
tions.
E-mail: olariu@cs.odu.edu
Larry Wilson is an Associate Professor of Computer
Science at Old Dominion University. He received
his Ph.D and M.Sci. degrees from the University of
Texas at Austin. His B.Sci. is from Midwestern Uni-
versity. Dr. Wilson is currently working in the area
of wireless sensor networks. Previous areas of re-
search include design and analysis of parallel algo-
rithms, software reliability and Mikusinski Operator
Functions.
E-mail: wilson@cs.odu.edu
WADAA ET AL.
Mohamed Eltoweissy is a (visiting) Professor of
Computer Science at Virginia Tech. He is also a
Professor of Computer Science at James Madison
University. Eltoweissys research interests include
information security and privacy, wireless sensor
and ad hoc networks, network security, computer-
supported cooperative work, and distributed com-
puting. He published extensively in books, refer-
eed journals and conference proceedings. He also
served on numerous technical committees for con-
ferences, workshops, seminars, and NSF panels. He has an aggressive record
of funding (over $10 million). Eltoweissy founded the Commonwealth In-
formation Security Center (CISC) in Virginia and was a founding member
of the award-winning Virginia Alliance for Secure Computing and Network-
ing (VA SCAN). Eltoweissy earned a Ph.D. in computer science (1993) from
Old Dominion University, and M.S. (1989) and B.S. (1986 with top rank) in
computer science and Automatic Control from Alexandria University, Egypt.
He was nominated by JMU for the Virginia State-wide Outstanding Faculty
Awards in 2003. Eltoweissy is a member of ACM, ACM SIGSAC, and the
honor societies of Phi Kappa Phi and Upsilon Pi Epsilon.
E-mail: toweissy@vt.edu
Kennie Jones received the B.S. degree in Biology
from Christopher Newport College, Newport News,
VA, in 1977 and the M.S. degree in computer sci-
ence from Old Dominion University (ODU), Nor-
folk, VA, in 1981. He is currently pursuing a Ph.D.
degree in computer science from ODU. After teach-
ing in secondary schools for 4 years, he joined Com-
puter Sciences Corporation in 1981 as a contractor
to the National Aeronautics and Space Administra-
tion (NASA). In 1990, he became an employee of
NASA. In that capacity, he has made significant contributions in the areas of
computer graphics and data management. He is currently researching sensor
networks and their application to NASA missions as a member of ODUs
Sensor Network Research Team.
E-mail: k.h.jones@nasa.gov
 Mobile Networks and Applications 10, 169181, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Quorum-Based Asynchronous Power-Saving Protocols for IEEE

802.11 Ad Hoc Networks
JEHN-RUEY JIANG
Department of Information Management, Hsuan-Chuang University, Taiwan

YU-CHEE TSENG
Department of Computer Science and Information Engineering, National Chiao-Tung University, Taiwan

CHIH-SHUN HSU
Department of Computer Science and Information Engineering, National Central University, Taiwan

TEN-HWANG LAI
Department of Computer and Information Science, The Ohio State University Columbus, OH 43210, USA

Abstract. This paper investigates the power mode management problem for an IEEE 802.11-based mobile ad hoc network (MANET) that
allows mobile hosts to tune to the power-saving (PS) mode. There are two major issues that need to be addressed in this problem: (a) wakeup
prediction and (b) neighbor discovery. The former is to deliver buffered packets to a PS host at the right time when its radio is turned on.
The latter is to monitor the environment change under a mobile environment. One costly, and not scalable, solution is to time-synchronize all
hosts. Another possibility is to design asynchronous protocols as proposed by Tseng et al. in [25]. In this paper, we adopt the latter approach
and correlate this problem to the quorum system concept. We identify a rotation closure property for quorum systems. It is shown that
any quorum system that satisfies this property can be translated to an asynchronous power-saving protocol for MANETs. Thus, the result
bridges the classical quorum system design problem in the area of distributed systems to the power mode management problem in the area
of mobile ad hoc networks. We derive a lower bound for quorum sizes for any quorum system that satisfies the rotation closure property.
We identify a group of quorum systems that are optimal or near optimal in terms of quorum sizes, which can be translated to efficient
asynchronous power-saving protocols. We also propose a new e-torus quorum system, which can be translated to an adaptive protocol that
allows designers to trade hosts neighbor sensibility for power efficiency. Simulation experiments are conducted to evaluate and compare
the proposed protocols.
Keywords: IEEE 802.11, distributed system, mobile ad hoc network (MANET), power management, quorum system, wireless communi-
cation

1. Introduction power-aware routing in [6,18,19,24], and low-power mode

The mobile ad hoc network (MANET) has attracted a lot
of attention recently. A MANET consists of a set of mo-
bile hosts, and does not have the support of any base station.
Hosts may communicate in a multi-hop manner. Applications
of MANETs include communications in battlefields, disaster
rescue operations, and outdoor activities.
Power saving is a critical issue for portable devices sup-
ported by batteries. Battery power is a limited resource, and
it is expected that battery technology is not likely to progress
as fast as computing and communication technologies do.
Hence, how to save the energy consumption in a MANET,
which is all supported by batteries, has been intensively stud-
ied recently (e.g., power control is studied in [8,9,17,26,28],
Y.C. Tsengs research is co-sponsored by the MOE Program for Promoting
Academic Excellence of Universities, Taiwan, under grant numbers A-91-
H-FA07-1-4 and 89-E-FA04-1-4, by NSC of Taiwan under grant numbers
NSC92-2213-E009-076 and NSC92-2219-E009-013, and by the Lee and
MTI Center of NCTU. J.R. Jiangs research is sponsored by NCS of Taiwan
under grant number NSC 92-2213-E-364-002.
management in [1,2,7,10,13,20,22,23,27,29]).
This paper investigates the power mode management prob-
lem in an IEEE 802.11-based MANET, which is characterized
by multi-hop communication, unpredictable mobility, and no
plug-in power. IEEE 802.11 [11] has defined its power-saving
(PS) mode for single-hop (fully connected) MANETs based
on periodical transmissions of beacons. The protocol, when
applied to a multi-hop MANET, may encounter several prob-
lems, including costly clock synchronization and even incor-
rect network partitioning [25].
There are two major issues that need to be addressed in the
power mode management problem in a multi-hop MANET:
Wakeup prediction. Since a host entering the PS mode will
reduce its radio activity, other hosts who intend to send
packets to the PS host need to know when the host will
turn its radio on so as to correctly deliver packets to it at
the right time.
Neighbor discovery. Because hosts transmission/recep-
tion activities are reduced under the PS mode, a host may
170
take longer time, or may be even unable, to detect the
arrival and departure of other hosts in its radio covered
range. Thus, hosts may become less sensitive to neigh-
borhood change. Neighbor discovery is essential for route
discovery in a MANET. A host may incorrectly report that
another host is unreachable if the route to this host has to
go through some PS hosts that are not detectable by their
neighbors on the path.
One possible solution to the above problems is to always time-
synchronize all hosts. This approach is adopted by IEEE
802.11 under the ad hoc mode. However, 802.11 only con-
siders single-hop MANETs. Time synchronization in a large-
scale distributed environment is generally very costly. It is
even infeasible in a mobile environment since communica-
tion delays are typically long and, worse, the MANET may
be temporarily partitioned at any time, making time synchro-
nization impossible. Another solution is to develop asyn-
chronous power-saving protocols. This is first investigated
in [25], where three solutions are proposed. Among them,
the quorum-based protocol is probably the most interesting
one. It has the merit of sending the fewest beacon signals
(and is thus very energy-efficient). The central idea in the
quorum-based protocol can be related to the grid quorum sys-
tem [15]. This leads to a more general question: Can we apply
other forms of quorum systems to this asynchronous power-
saving problem? The result can potentially bridge the impor-
tant quorum system concept in traditional distributed systems
to the area of mobile computing, which may in turn gener-
ate more efficient asynchronous power-saving protocols. This
work does confirm such possibility.
In this paper, we correlate the asynchronous power-saving
problem to the concept of quorum systems, which are widely
used in the design of distributed systems [4,12,14,15]. A quo-
rum system is a collection of sets such that the intersection of
any two sets is always non-empty. Not all quorum systems
are applicable to the power-saving problem. We identify a ro-
tation closure property for quorum systems. It is shown that,
through our mechanism, any quorum system satisfying this
property can be translated to an asynchronous power-saving
protocol for MANETs. We derive a lower bound for quorum
sizes for any quorum system satisfying the rotation closure
property. We identify a group of quorum systems that are
optimal or near optimal in terms of quorum sizes (the grid
quorum system [15], the torus quorum system [12], the cyclic
quorum system [14], and the finite projective plane quorum
system [15]), which can be translated to efficient asynchro-
nous power-saving protocols. We also propose a new e-torus
quorum system, which can be translated to an adaptive proto-
col that allows designers to trade hosts neighbor sensibility
for power efficiency. A host can dynamically adjust its bea-
con rate according to its mobility. Simulation experiments are
conducted to evaluate and compare the proposed protocols in
terms of the survival ratio, the route establishment probability,
and the power efficiency.
The rest of this paper is organized as follows. Prelimi-
naries are given in section 2. Section 3 introduces the rota-
JIANG ET AL.
tion closure property. Section 4 shows several quorum sys-
tems that satisfy this property. Section 5 presents our adap-
tive power-saving protocol. Simulation results are presented
in section 6. Conclusions are drawn in section 7.
2. Preliminaries
2.1. Power-saving modes in IEEE 802.11
IEEE 802.11 supports two power modes: active and power-
saving (PS). Under the PS mode, a host can reduce its radio
activity by only monitoring some periodical signals (such as
beacons) in the network. Tuning a host to the PS mode can
save a lot of energy. For example, table 1 summarizes the
power consumption of ORiNOCO IEEE 802.11b PC Gold
Card [21]. However, PS mode should be used cautiously so
that the network throughput and delay do not get hurt.
Under the ad hoc mode, IEEE 802.11 divides the time axis
into equal-length beacon intervals, each of which starts with
an ATIM (Ad hoc Traffic Indication Map) window. The ATIM
window is relatively small compared to the beacon interval.
PS hosts must remain active during the ATIM window so as to
be notified by those intending senders, and may go to doze in
the rest of the beacon interval if no one intends to send packets
to it. It is assumed that the ad hoc network is fully connected,
so time synchronization is not an issue. In the beginning of
a beacon interval, each mobile host will contend to send a
beacon frame. Any successful beacon serves the purpose of
synchronizing mobile hosts clocks as well as inhibiting other
hosts from sending their beacons. To avoid collisions, each
beacon is led by a random backoff between 0 and 2CW min 1
slots.
After the beacon, a host with buffered packets can send
a direct ATIM frame to each of its intended receivers in the
PS mode. ATIMs are transmitted by contention in accor-
dance with the DCF (Distributed Coordination Function) ac-
cess procedure. A receiver, on hearing the ATIM, should re-
ply an ACK and remain active. After the ATIM window, hosts
having neither packets to send nor packets to receive can go
back to the PS mode to save energy. The buffered unicast
packets are then sent based on the DCF access procedure af-
ter the ATIM window. If the sender does not receive an ACK,
it should retry in the next ATIM window. If a mobile host is
unable to transmit its ATIM frame in the current ATIM win-
dow or has extra buffered packets, it should retransmit ATIMs
in the next ATIM window. To protect PS hosts, only RTS (re-
quest to send), CTS (clear to send), ACK, Beacon, and ATIM
frames can be transmitted during the ATIM window. An ex-
ample is illustrated in figure 1.
Table 1
Power consumption of the ORiNOCO IEEE 802.11b PC Gold Card
(11 Mbps).
Mode PS (doze) Transmit Receive Monitor
Power
Consumed 60 mW 1400 mW 950 mW 805 mW
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS
Figure 1. Transmission scenarios for PS hosts in a single-hop 802.11 MANET.
Figure 2. Structures of quorum intervals and non-quorum intervals.
2.2. Review: a quorum-based PS protocol
IEEE 802.11 only considers single-hop MANETs. For multi-
hop MANETs, the following two issues have to be addressed:
wakeup prediction and neighbor discovery. In [25], three so-
lutions are proposed to solve these problems: the dominating-
awake-interval, the periodically-fully-awake-interval, and the
quorum-based protocols. Among them, the quorum-based
one has the merit of sending the fewest beacon signals. Be-
low, we briefly review the quorum-based protocol proposed
in [25]. Still, the time axis is divided evenly into beacon in-
tervals. Hosts can be arbitrarily asynchronous in their clocks.
Beacon intervals are classified into two types (refer to fig-
ure 2):
Quorum interval. It starts with a beacon window followed
by a MTIM window. After the MTIM window, the host
remains active (in monitor mode) for the rest of the beacon
interval.
Non-quorum interval. It starts with a MTIM window. Af-
ter the MTIM window, the host may go to the PS mode if
it has no packets to send or receive.
Similar to IEEE 802.11, the beacon window is for hosts to
compete sending their beacons. The MTIM window is sim-
ilar to the ATIM window a host with buffered packets can
compete to send notifications to intended receivers in the PS
mode to wake them up. It is named so to reflect that it is
used for multi-hop ad hoc networks. We assume that beacon
171
windows are not longer than MTIM windows (the assumption
is practical considering these two windows functionality; the
assumption will also be used in our later proofs). With these
definitions, we say that a PS host is active when it is currently
in a beacon window, a MTIM window, or in a quorum inter-
val.
In [25], it is proposed that each host divides its beacon
intervals into groups such that each group consists of n con-
secutive intervals. Each group is organized as an n n
array in a row-major manner. The host then picks intervals
along an arbitrary row and an arbitrary column from the ar-
ray as quorum intervals, and the remaining intervals as non-
quorum intervals. Thus, there are 2 n 1 quorum intervals.
It is shown that no matter how asynchronous hosts clocks
are, a PS host always has two or more beacon windows that
are fully covered by another PS hosts active period in every
n consecutive beacon intervals. Intuitively, this implies that
two hosts can discover each other at least twice in every n
consecutive beacon intervals, if their beacon frames do not
encounter collisions during transmission.1 Thus, the neigh-
bor discovery problem is resolved. Further, by carrying clock
information in beacon frames, the wake-up prediction prob-
lem is also solved.
Figure 3 shows an example with n = 16. Host A picks in-
tervals along the first row and the second column as its beacon
intervals. Host B, which does not coordinate with A, picks the
third row and the third column. In the middle, we show the
case where As and Bs clocks are perfectly synchronized, in
which case intervals 2 and 9 of A and B are fully covered by
each other. On the bottom, we show the case where A and
B are asynchronous in clocks. The beacon windows of in-
tervals 0 and 13 of A are fully covered by the duration when
B is active. On the contrary, the beacon windows of inter-
vals 2 and 8 of B are fully covered by the duration when A is
active.
1 Collision is inevitable in any kind of contention-based MAC protocols.
172
Figure 3. Arrangement of quorum intervals based on the grid quorum system
in [25].
2.3. Problem statement
The arrangement of quorum intervals in [25] is in fact based
on the grid quorum system [15]. This leads to the follow-
ing interesting question. Can one simply take any quorum
system, which is a collection of pairwise non-disjoint sets,
and apply it to solve the asynchronous power-saving prob-
lem in MANET? The answer is negative, due to the following
counterexample. Lets number each hosts beacon intervals
by 0, 1, and 2 repeatedly, and let {{0}} be the quorum sys-
tem. Hence, each host will pick interval 0 as its quorum in-
terval. It is evident that two hosts whose clocks drift by 1 or 2
beacon intervals will never be able to hear each others bea-
cons. Now, an even more interesting question arises: What
kind of quorum systems is applicable to solve the asynchro-
nous power-saving problem in MANETs?
The quorum-based power-saving (QPS) problem is for-
mally defined as follows. We are given a universal set U =
{0, . . . , n 1}, n
2, which represents a set of consecutive
beacon intervals of mobile hosts. The goal is to determine un-
der U a quorum system Q, which is a collection of pairwise
non-disjoint subsets of U , each called a quorum, such that
each mobile host has freedom to pick any quorum G Q to
contain all its quorum intervals (the beacon intervals not in G
are thus non-quorum intervals). The quorum system Q has
to guarantee that for any two arbitrarily time-asynchronous
hosts A and B, host As beacon windows are fully covered by
host Bs active durations at least once in every n consecutive
beacon intervals, and vice versa.
3. Quorum systems for the QPS problem
Definition 1. Given a universal set U = {0, . . . , n 1}, a
quorum system Q under U is a collection of non-empty sub-
sets of U , each called a quorum, which satisfies the intersec-
tion property:
G, H Q: G H = .
For example, Q = {{0, 1}, {0, 2}, {1, 2}} is a quorum sys-
tem under U = {0, 1, 2}.
JIANG ET AL.
Figure 4. Timing drift of clocks of two asynchronous hosts.
Definition 2. Given a non-negative integer i and a quorum H
in a quorum system Q under U = {0, . . . , n 1}, we define
rotate(H, i) = {(j + i) mod n | j H }.
Definition 3. A quorum system Q under U = {0, . . . , n 1}
is said to have the rotation closure property if
G, H Q, i {0, . . . , n 1}: G rotate(H, i) = .
For instance, the quorum system Q = {{0, 1}, {0, 2}, {1,
2}} under {0, 1, 2} has the rotation closure property. How-
ever, the quorum system Q = {{0, 1}, {0, 2}, {0, 3}, {1, 2, 3}}
under {0, 1, 2, 3} has no rotation closure property because
{0, 1} rotate({0, 3}, 3) = .
Throughout the rest of this paper, we will assume that bea-
con windows are not longer than MTIM windows. The fol-
lowing theorem connects quorum systems to the QPS prob-
lem.
Theorem 1. If Q is a quorum system satisfying the rotation
closure property, Q is a solution to the QPS problem.
Proof. Let A and B be two asynchronous PS hosts in a
MANET which choose G and H Q to represent their quo-
rum intervals, respectively. Without loss of generality, let As
clock lead Bs clock by k BI +
t, where BI is the length
of one beacon interval, k < n is a non-negative integer, and
0 
t < BI. This is illustrated in figure 4. First, we show
that Bs beacon window is fully covered by As active dura-
tions at least once every n beacon intervals. The pattern H
of B is, in fact, rotate(H, k) from As point of view, with
an extra delay of
t. Note that in the following discussion,
time always refers to As clock. By the rotation closure prop-
erty of Q, G rotate(H, k) = . Let e be any element in
G rotate(H, k) and let s be the starting time of As inter-
val e. Also, let BW and MW be the lengths of one beacon
window and one MTIM window, respectively. Taking into
account the next interval e + 1, we know that A is active from
s to s +BI +MW. Since Bs beacon window falls in the range
[s +
t, s +
t +BW] and BW  MW, it is easy to see that for
any value of
t, [s +
t, s +
t + BW] [s, s + BI + MW].
So this part is proved.
Next, we show the reverse direction that As beacon win-
dow is fully covered by Bs active durations at least once
every n beacon intervals. We first observe that if 0 <
t <
BI, the pattern G of A is rotate(G, n k 1) from Bs
point of view, with an extra delay of BI
t (note that
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS
0 < BI
t < BI). We also observe that if
t = 0, the
pattern G is rotate(G, n k) with 0 delay from Bs point of
view. Thus, a proof similar to that in the last paragraph can be
applied to prove the reverse direction by exchanging A and B
and substituting
t with BI
t. 
We comment that the above proof requires the property
that BW  MW; otherwise, the conclusion of [s +
t, s +

t + BW] [s, s + BI + MW] may not be true.
It is important to note that the number of quorum inter-
vals reflects the power consumption of PS hosts since quo-
rum intervals are more energy-consuming (recall that a PS
host needs to send a beacon and remains active in each quo-
rum interval). Given a fixed n, the cost can be measured by
the sizes of quorums in the quorum system. It is desirable
that the quorum sizes are as small as possible. In the follow-
ing theorem, we derive a lower bound on quorum sizes for
any quorum system satisfying the rotation closure property.
A quorum system is said to be optimal if the sizes of all its
quorums meet the lower bound.
Theorem 2. Let Q be a quorum system under {0, . . . , n 1}.
If Q satisfies the rotation closure
property, then any quorum
in Q must have a cardinality
n.
Proof. Let H = {h1 , . . . , hk } be any quorum in Q, where
0 < k < n. There are two cases.
Case 1. H = rotate(H, i) for any i = n (mod n).
Since h1 , h2 , . . . , hk are distinct elements, it is clear that
h1 + i, h2 + i, . . . , hk + i (mod n) are also distinct for
any i = 1..n 1. So, |rotate(H, i)| = k. Lets call
rotate(H, i), i = 1..n 1, the rotating quorums of H . For
each element hj H , it belongs to exactly k 1 rotating
quorums of H , namely rotate(H, (hj hj  ) mod n) for every
hj  = hj . By the rotation closure property, H must contain
at least one element from each of the n 1 rotating quorums
of H . Since each element appears in exactly k 1 rotat-
ing quorums of H and there are k elements in H , we have
k(k 1)
n 1, which implies k > n. Thus, the theorem
holds for case 1.
Case 2. H = rotate(H, i) for some i = n (mod n). Let
d be the smallest integer such that H = rotate(H, d). It is a
simple result in number theory that n is a multiple of d. So it
can be concluded that H = rotate(H, d) = rotate(H, 2d) =
rotate(H, 3d) = = rotate(H, n d). That is, when map-
ping the quorum elements of H onto the time axis, H can be
regarded as n/d equivalent segments, each of length d. In
fact, from H , we can define a smaller quorum
H  = {j mod d | j H }
under the universal set {0, . . . , d 1}. Intuitively, on the time
axis, H can be considered as a concatenation of n/d copies
of H  . Since H rotate(H, i) = , we can conclude that H 
rotate(H  , i) = for any i under modulo-d arithmetic. So
{H  } is also a quorum system satisfying the rotation closure
w/2 elements, each
property under the universal set {0, . . . , d
1}. We can apply
w/2.
the result in case 1 and infer that |H |
d. It follows that
|H | = (n/d)|H  |
(n/d) d > n.
173
Figure 5. Two quorums of the torus quorum system in a 3 6 torus.
4. Quorum systems with the rotation closure property
Although there are volumes of works devoted to quorum sys-
tems, none of them discusses the rotation closure property
to the best of our knowledge. In this section, we prove that
the grid quorum system [15], the torus quorum system [12],
the cyclic quorum system [14], and the finite projective plane
quorum system [15] are all optimal or near optimal quorum
systems (in terms of quorum sizes) satisfying the rotation clo-
sure property.
4.1. The grid quorum system
The grid quorum system [15] arranges
elements
of the univer-
sal set U = {0, . . . , n 1} as a n n array. A quorum
can be any set containing a full column plus a full row of ele-
ments
in the array. Thus, each quorum has a near optimal size
of 2 n 1. As noted above, the work in [25] adopts the grid
quorum system. Below, we prove the rotation closure prop-
erty for the grid quorum system. The theorem, when accom-
panied with theorem 1, can simplify the lengthy correctness
proof of the work in [25], which needs to deal with compli-
cated timing relation between quorum and non-quorum inter-
vals among different asynchronous hosts.
Theorem 3. The grid quorum system satisfies the rotation
closure property.
Proof. Let Q be a grid quorum system. Let H Q, which
contains all elements on the column
c of the array, namely
c, c + n, . . . , c + ( n 1) n, where 0  c < n (note
that we number columns from 0 to n 1). Now observe
that rotate(H, i) must contain all elements on column (c + i)
(mod n). It follows that rotate(H, i) must have intersection
with any quorum G Q because G must contain a full row
in the array. 
4.2. The torus quorum system
Similar to the grid quorum system, the torus quorum sys-
tem [12] also adopts an array structure. The universal set is
arranged as a t w array, where tw = n. Following the con-
cept of torus, the rightmost column (resp., the bottom row) in
the array are regarded as wrapping around back to the leftmost
column (resp., the top row). A quorum is formed by picking
any column c, 0  c  w 1, plus
of which falls in any position of column c + i, i = 1..
Figure 5 illustrates the construction of two torus quorums G
 and H under U = {0, . . . , 17} with t = 3 and w = 6. G is
174
formed by picking the second column plus three elements,
each from one of the third, fourth, and fifth columns. H is
formed by picking the sixth column plus three elements, each
from one of the first, second, and third columns. G and H
intersect at element 7.
As shown in [12], if we let t = w/2, the quorum size will
be 2tw = 2n, which is near optimal. By equating n,
the torus quorum size is about 1/ 2 that of the grid quorum
size. Below, we prove the rotation closure property for the
torus quorum system.
Theorem 4. The torus quorum system satisfies the rotation
closure property.
Proof. Let Q be a torus quorum system formed by a t w
array and H Q be a quorum containing column c. By de-
finition, H also contains another
w/2 elements, each from that a FPP can be constructed when n = k(k 1)+1 and k 1
one of the
w/2 succeeding columns of column c. Clearly, is a prime power. Otherwise, the FPP may or may not exist.
rotate(H, i) still has the torus quorum structure for an arbi-
trary i. It follows that for any G Q, G rotate(H, i)
= .  system can be regarded as a special case of the cyclic quorum
4.3. The cyclic quorum system
The cyclic quorum systems [14] are constructed from the dif-
ference sets as defined below.
Definition 4. A subset D = {d1 , d2 , . . . , dk } of Zn is called a
difference set under Zn if for every e = 0 (mod n) there exist
elements di and dj D such that di dj = e (mod n).
Definition 5. Given any difference set D = {d1 , d2 , . . . , dk }
under Zn , the cyclic quorum system defined by D is Q =
{G1 , G2 , . . . , Gn }, where Gi = {d1 + i, d2 + i, . . . , dk + i}
(mod n), i = 0, . . . , n 1.
For example, D = {0, 1, 2, 4} Z8 is a difference set
under Z8 since each e = 1..7 can be generated by taking the
difference of two elements in D. Given D, Q = {G0 =
{0, 1, 2, 4}, G1 = {1, 2, 3, 5}, G2 = {2, 3, 4, 6}, G3 =
{3, 4, 5, 7}, G4 = {4, 5, 6, 0}, G5 = {5, 6, 7, 1}, G6 =
{6, 7, 0, 2}, G7 = {7, 0, 1, 3}} is a cyclic quorum system un-
der Z8 .
Given any n, a difference set as small as k can be found
when k(k 1) + 1 = n and k 1 is a prime power. Such a
difference set is called the Singer difference set [3]. For ex-
ample, the sets {1, 2, 4} under Z7 and {1, 2, 4, 9, 13, 19} un-
der Z31 are Singer difference sets. Note that in this case the
quorum size k meets the lower bound in theorem 2. So cyclic
quorum systems defined by the Singer difference sets are opti-
mal. Reference [14] had conducted exhausted searches to find
the minimal difference sets under Zn for n = 4..111. The re-
sults are useful here to construct near-optimal cyclic quorum
systems.
Theorem 5. The cyclic quorum system satisfies the rotation
closure property.
JIANG ET AL.
Proof. Let H be a quorum in the cyclic quorum system Q
generated from the difference set D = {d1 , d2 , . . . , dk }. By
definition, rotate(H, i) is also a quorum in Q for any i. Then
by the intersection property, the theorem holds. 
4.4. The finite projective plane quorum system
The finite projective plane (FPP) quorum system [15] arranges
elements of the universal set U = {0, . . . , n 1} as ver-
tices on a hypergraph called the finite projective plane, which
has n vertices and n edges, such that each edge is connected
to k vertices and two edges have exactly one common ver-
tex. (Note that the hypergraph is a generalization of typical
graphs, where each edge is connected to only two vertices.)
A quorum can be formed by the set of all vertices connected
by the edge, and thus has a size of k. It has been shown in [15]
In [14], the FPP construction is associated to the construction
of Singer difference sets, and it is shown that the FPP quorum
system when n = k(k 1) + 1 and k 1 is a prime power. It
follows that FPP quorum systems also own the rotation clo-
sure property, and are optimal, when existing.
4.5. Quorum systems with one quorum
In this subsection, we discuss the rotation closure property for
those quorum systems with only one quorum. The result has
strong connection to the difference sets, and can help identify
the quorum systems that are solution to the QPS problem.
Theorem 6. Let Q = {H } be a quorum system under U =
{0, . . . , n 1}. Q satisfies the rotation closure property if and
only if H is a difference set of Zn .
Proof. For the if part, let H be a difference set of Zn . For
any i, there must exist two elements hx , hy H such that
hx hy = i. It follows that hx = hy + i rotate(H, i) H .
So rotate(H, i) H = for any i.
For the only if part, suppose for contradiction that H is
not a difference set of Zn . Then there exists an i = 0 such
that hx hy = i for all possible combinations of hx and hy
in H . Since rotate(H, i) = {(hy + i) mod n | hy H }, it
follows that H rotate(H, i) = , a contradiction. 
Corollary 1. Let Q be a quorum system under U = {0, . . . ,
n 1}. Q does not satisfy the rotation closure property if at
least one quorum in Q is not a difference set under Zn .
Theorem 6 says that if a quorum system has a difference
set being its sole quorum, it satisfies the rotation closure prop-
erty and is thus a solution to the QPS problem. Such a quo-
rum system has the practical advantage that it is very easy to
maintain since it has only one quorum to keep. For exam-
ple, from each of the minimal difference sets found in [14]
(for n = 4..111), a solution to the QPS problem exists by
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS 175

simply putting the difference set as the single quorum in the

quorum system. On the contrary, when n is too large such
that exhaustive searches (as in [14]) are prohibited, we can
pick any quorum G in the quorum systems with the rotation
closure property. Then G is a difference set by the contra-
position of corollary 1. For example, from the torus quorum
system, we can quickly find a lot of near-optimal difference
sets by arranging numbers from 0 to n 1 as an array. Note
that in situations when n can not be divided into a product of t (a)
and w, we can always add a virtual element on the array, as
proposed in [15], to solve the problem. For example, when
n = 13, we can make a 2 7 array with the last position filled
by 0 as the virtual element.

5. An adaptive QPS protocol

All the quorum systems discussed above ensure that given a

fixed n, two asynchronous mobile hosts picking any two quo-
rums have at least one intersection in their quorums. It would
be desirable to have an adaptive solution in the sense that the
number of intersecting elements can be dynamically adjusted.
One of the main reasons to do so would be to adjust this value
to adapt to host mobility. Intuitively, the number of beacons
that two hosts can hear from each other is proportional to the
number of intersecting elements. Thus, a host with higher
mobility may like to have more intersections with its neigh-
boring hosts so as to be more environment-sensitive. On the
contrary, a host with lower mobility may not need to intersect
in so many elements with its neighbors so as to save more
energy. The proposed solution is adaptive in this sense.
We assume that a host is able to calculate its mobility lev-
els, either through attaching a GPS device, or simply by eval-
uating the number of hosts that are detected to leave/enter the
hosts radio coverage. We leave this as an independent issue,
and only focus on the design of adaptive quorum systems to
meet our goal.
The proposed solution is basically an extension of the
torus quorum system, and is thus called the extended torus
(e-torus) quorum system. An e-torus quorum system is also
defined based on two given integers t and w such that U =
{0, 1, . . . , tw 1} is the universal set. Elements of U are
arranged in a t w array. Below, we use [x, y] as an array
index, 0  x < t and 0  y < w.
Definition 6. On a t w array, a positive half diagonal
starting from position [x, y], where 0  x < t and 0 
y < w, consists of element [x, y] plus
w/2 elements quorum, the theorem holds.
[(x + i) mod t, (y + i) mod w], for i = 1..
w/2. A negative
half diagonal starting from position [x, y] consists of element
[x, y] plus
w/21 elements [(x+i) mod t, (yi) mod w], e-torus(k2) quorum derived from the same array. For any in-
for i = 1..
w/2 1. tegers i and j , |rotate(G, i) rotate(H, j )|

(k1 + k2 )/2.
(b)
Figure 6. (a) the Christmas tree structure of an e-torus(4) quorum, and
(b) the intersection of an e-torus(2) quorum and an e-torus(3) quorum.
in the southwest direction. The diagonal is slightly different
from typical diagonal in matrix algebra in that the array is
not necessarily square and that the torus has the wrap-around
property.
Definition 7. Given any integer k  t, a quorum of an
e-torus(k) quorum system is formed by picking any position
[r, c], where 0  r < t and 0  c < w, such that the quo-
rum contains all elements on column c plus k half diagonals.
These k half diagonals alternate between positive and nega-
tive ones, and start from the following positions:

  
t
r+ i , c , i = 0..k 1.
k
Intuitively, each quorum in the e-torus(k) quorum system
looks like a Christmas tree with a trunk in the middle and k
branches, each as a half diagonal, alternating between posi-
tive and negative ones. Figure 6(a) illustrates the conceptual
structure of an e-torus(4) quorum.
Theorem 7. The e-torus quorum system satisfies the rotation
closure property.
Proof. Since any e-torus quorum is a super set of a torus

Theorem 8. Let G be an e-torus(k1) quorum and H be an

Intuitively, a positive (resp., negative) half diagonal is a Proof. This theorem can be easily observed from the geo-
partial diagonal on the array starting from the array index metric structure of the e-torus quorum system. The value of
[x, y] with a length
w/2+1 (resp.,
w/2). A positive diag- |rotate(G, i) rotate(H, j )| can be observed from how tree
onal goes in the southeast direction, while a negative one goes branches intersect with the trunks of Christmas trees. 
176
Figure 7. Analysis of neighbor sensibility of an e-torus(k1 ) and an e-torus(k2 ) quorum systems under a 714 torus.
For example, figure 6(b) shows how an e-torus(3) quorum
and an e-torus(2) quorum intersect with each other. The in-
tersecting elements are guaranteed to appear in the trunks of
the Christmas trees. Note that two branches from two e-
torus quorums may cross with each other, but intersection
is not necessarily guaranteed (from the geometric structures
of branches, it does look like that they are guaranteed to in-
tersect). The reason is illustrated in the zoomed-in part in
figure 6(b), where the two branches just miss each other on
2t/k1 ), which
the array. Also note that by our arrangement, the intersect-
ing elements of two e-torus quorums are unlikely to concen-
trated in certain areas of the array. Instead, they will be spread
evenly over the trunks. This is a desirable property because it
implies that the quorum intervals that two mobile hosts may
detect each other will be spread evenly over the time axis.
Based on the above features, we propose an adaptive QPS
protocol as follows. We can rank a hosts mobility into k-
levels, where level 1 means the lowest mobility, and level k
means the highest mobility. Whenever a host determines that
its mobility falls within level i (1  i  k), it adjusts its quo-
rum intervals based on any e-torus(i) quorum. Consequently,
a host can dynamically adjust its sensibility to the environ-
ment change in its neighborhood.
6. Performance comparison and simulation results
6.1. Analytical comparison
In this subsection, we compare the proposed quorum-based
protocols analytically. We evaluate the active ratio, which is
defined to be the number of quorum intervals over n (the size
of universal set), and the neighbor sensibility (NS), which is
the worst-case delay for a PS host to detect the existence of
a newly approaching PS host in its neighborhood.
The NS
of the grid quorum system is BI (n n + 1), which
happens when two quorums intersect at indices (i, j ) and
JIANG ET AL.
(i + 1, j 1) of the array. The NS of two e-torus quo-
rum systems e-torus(k1) and e-torus(k2) under a t w torus
is discussed below (refer to figure 7 for illustration). When
(k1 = k2 = 1), (k1 = 2 k2 = 1), or (k1 = 1 k2 = 2),
the two quorums may intersect at only one interval, so NS is
BI n. For (k1 = 3 k2 = 1) or (k1 = 1 k2 = 3), the NS
is BI (n 1), which happens when two quorums intersect
at two consecutive quorum intervals. For (k1 = 1 k2 = 4)
or (k1 = 4 k2 = 1), the NS is BI (n
happens when the intersections fall in one column. Table 2
contains the NS of other cases. Table 2 also summarizes the
active ratio and neighbor sensibility of the proposed quorum-
based protocols. Figure 8 further demonstrates the active ratio
of the proposed protocols for n = 5..100. The cyclic quorum
performs the best in terms of active ratio. The FPP quorum
system, when available, represents the optimal solution.
6.2. Simulation results
In this subsection, we compare the proposed power-saving
protocols through a simulator written in C. An area of size
1000 m 1000 m is simulated. Each host has an antenna
with a transmission rate of 2 Mb/s and a transmission ra-
dius of 250 m, and has an initial battery energy of 100 J. The
MAC part basically follows the IEEE 802.11 standard [11],
except the power management part. Routes with random
sources/destinations are generated, and the AODV (ad-hoc
on-demand distance vector) routing protocol [16] is adopted.
Four parameters are tunable in our simulations:
Mobility. Host mobility follows the random way-point
model, with pause time of 20 sec. When moving, a hosts
speed can range in 020 m/sec.
Traffic load. Routes are generated by a Poisson distribu-
tion with rates between 14 routes/sec. For each route,
10 packets, each of size 1 KB, are sent.
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS 177

Table 2
Active ratios and neighbor sensitivity of quorum-based protocols.

Quorum system Active ratio Neighbor sensibility

Grid 2 BI (n n + 1)
n

Torus 2 BI n
n
Cyclic 1 BI n
n
Finite projective plane k/n, where k(k 1) + 1 = n, BI n
and k 1 is a prime power
e-torus(k), s/n, where s = t + w
k/2 + r
(w 1)/2, For e-torus(k1 ) and e-torus(k2 ):
(under t w torus) r = 0 if k is even and r = 1 if k is odd BI n, for (k1 = k2 = 1), (k1 = 2 k2 = 1), or
(k1 = 1 k2 = 2)
BI (n 1), for (k1 = 3 k2 = 1) or
(k1 = 1 k2 = 3)
BI (n
2t/k1 ), for (k1 = 1 k2 = 4) or
(k1 = 4 k2 = 1)
less than BI (n
2t/k1 ), for (k1 > 4 k2 = 1) or
(k1 = 1 k2 > 4)
BI (n w + 1), for (k1 = 2 k2 = 2)
less than BI (n w + 1), for (k1
2 k2 > 2) or
(k1 > 2 k2
2)

Table 3
Power consumption parameters used in the simulation.

Unicast send 454 + 1.9 L J/packet

Broadcast send 266 + 1.9 L J/packet
Unicast receive 356 + 0.5 L J/packet
Broadcast receive 56 + 0.5 L J/packet
Idle 843 J/ms
Doze 27 J/ms

Table 4
Traffic-related parameters used in the simulation.

Unicast packet size 1024 bytes

Figure 8. Active ratios of different quorum systems for n = 5..100.
Beacon interval. The length of one beacon interval is 100
400 ms.
Number of hosts. The total number of mobile hosts in the
MANET is 50200 hosts.
Three performance metrics are measured in the simula-
tions:
Survival ratio. The number of surviving hosts (with non-
zero energy) over the total number of hosts.
Neighbor discovery time. Average time to discover a
newly approaching neighbor.
Throughput. The average number of MAC-layer data
packets successfully received in the network per second.
Except the survival ratio, the above metrics are evaluated up
to the time when 10% of the hosts run out of energy. A host
can go to the PS mode when it does not serve as a source,
destination, or relay of any route. A broadcast (such as the
AODV route request message) may need to be sent multiple
times if the sending host finds that some of its neighbors are
in the PS mode [25]. This is necessary because these PS hosts
may wake up at different times and we need multiple trans-
Broadcast packet size 32 bytes
Beacon window size 4 ms
MTIM window size 16 ms
missions to cover all of them. However, once a route is estab-
lished (via the notification of a route reply message), all hosts
in the route have to tune to the active mode.
Table 3 summarizes the power consumption parameters
used in our simulations, which are obtained from real exper-
iments using Lucent WaveLAN cards [5]. Sending/receiving
a unicast/broadcast packet of L bytes has a cost Pbase + Pbyte
L, where Pbase is the power consumption independent of
packet length and Pbyte is the power consumption per byte.
Unicast consumes more power than broadcast because it in-
curs extra control frames (RTS, CTS, and ACK). Idle/doze
represents the condition when a host has no send and receive
activity and is in the active/PS mode, respectively. The traffic-
related parameters are summarized in table 4.
Below, we show how mobility, beacon interval length, traf-
fic load, and host density affect the performance of the pro-
posed PS protocols. We mainly compare the cyclic quorum
system (which has the lowest active ratio) and the e-torus quo-
rum system (which is more adaptive). Below, C(n) stands
for the cyclic quorum system under {0, 1, . . . , n 1}, and
178
Figure 9. Host survival ratio vs. mobility (beacon interval = 100 ms,
100 hosts, traffic load = 1 route/sec).
Figure 10. Neighbor discovery time vs. mobility (beacon interval = 100 ms,
100 hosts, traffic load = 1 route/sec).
E(t w) the e-torus quorum system under a t w torus.
For the e-torus quorum system, four speed levels (14) are as-
sumed. A host is said to be at speed level i if its speed is larger
than 5(i 1) m/sec and less than or equal to 5i m/sec. To
make comparison, we also simulate an always-active (AA)
scheme in which all hosts are active all the time.
6.2.1. Impact of mobility
Mobility has a negative impact on survival ratio. Figure 9
compares the cases when all hosts are stationary and when all
hosts moving speed = 20 m/sec. Mobility will incur higher
energy consumption because hosts may spend more energy in
retransmitting packets. On the contrary, mobility has very lit-
tle impact on AA. However, because hosts can tune to the PS
mode, C(98) and E(714) still outperform AA significantly
in terms of survival ratio. The survival ratio of C(98) is better
than that of E(714) because its active ratio is smaller.
Figure 10 shows the impact of mobility on the neighbor
discovery time. Mobility has a negative impact on neighbor
discovery time for C(98). On the contrary, E(714) can bet-
ter adapt itself to mobility. We even see shorter neighbor dis-
covery time when host mobility becomes higher (at the cost
of more beacon intervals).
Figure 11 shows the impact of mobility on throughput.
Mobility has a negative impact on throughput for all schemes
because more retransmissions are incurred as hosts move
faster. The results show that C(98) and E(714) will slightly
degrade throughputs compared to AA when we allow hosts
to tune to the PS mode, which is reasonable. However, the
JIANG ET AL.
Figure 11. Throughput vs. mobility (beacon interval = 100 ms, 100 hosts,
traffic load = 1 route/sec).
Figure 12. Survival ratio vs. beacon interval length (100 hosts, traffic load =
1 route/sec, moving speed = 020 m/sec with mean = 10 m/sec).
Figure 13. Neighbor discovery time vs. beacon interval length (100 hosts,
traffic load = 1 route/sec, moving speed = 020 m/sec with mean =
10 m/sec).
benefit is that the network can be used for much longer time,
as reflected by the axis throughput lifetime, where the
lifetime of a network is counted up to the point when 10% of
hosts runs out of energy.
6.2.2. Impact of beacon interval length
We observe the impact of beacon interval (BI) length on host
survival ratio by varying the beacon interval length between
100400 ms. Figure 12 shows that a longer BI will slightly
shorten the lifetime of the network for C(98) and E(714).
We believe that this is due to a higher transmission cost for
broadcasting route request packets. However, a longer BI
makes hosts conserve more energy, which in turn prolongs
the lifetime of the hosts. This may explain the crossing points
in figure 12.
A longer BI also hurts the neighbor discovery time. As
shown in figure 13, the neighbor discovery time will increase
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS
Figure 14. Throughput vs. beacon interval length (100 hosts, traffic load =
1 route/sec, moving speed = 020 m/sec with mean = 10 m/sec).
Figure 15. Survival ratio vs. traffic load (beacon interval = 100 ms,
100 hosts, mobility = 020 m/sec with mean = 10 m/sec).
linearly as BI increases for both for C(98) and E(714). The
E(714) scheme, which can tune its quorum intervals adap-
tively, has much shorter neighbor discovery time compared to
C(98).
Figure 14 shows the impact of BI on throughput. Longer
beacon intervals do decrease throughputs. This is because it
takes longer time for a host to wake up its neighboring PS
hosts to help relay packets. The result shows that E(714)
slightly outperforms C(98) in terms of throughput due to
its adaptivity, and the gain will enlarge slightly as BI in-
creases. However, C(98) outperforms E(714) in terms of
throughput lifetime.
6.2.3. Impact of traffic load
Next, we observe the effect of traffic load. We vary the traffic
load in the range of 14 routes/sec in the simulations. Fig-
ure 15 shows how traffic load decreases host survival ratios.
Higher traffic loads do reduce host survival ratios of C(98)
and E(714), which is reasonable. On the contrary, the im-
pact of traffic load on AA is insignificant because anyway
hosts have to stay awake all the time. Overall, the proposed
schemes still outperform the AA scheme in terms of survival
ratio significantly. The effect of traffic load on throughput and
accumulated throughput is shown in figure 16. The trend is
similar to the earlier observation. Traffic load does not influ-
ence the neighbor discovery time much, so the related results
are omitted.
6.2.4. Impact of host density
In this experiment, we vary the number of hosts in the range
of 50200. Since the network area is fixed, this parameter
179
Figure 16. Throughput vs. traffic load (beacon interval = 100 ms, 100 hosts,
mobility = 020 m/sec with mean = 10 m/sec).
Figure 17. Survival ratio vs. host density (beacon interval = 100 ms, traffic
load 1 route/sec, mobility = 020 m/sec with mean = 10 m/sec).
Figure 18. Throughput vs. node density (beacon interval = 100 ms, traffic
load 1 route/sec, mobility = 020 m/sec with mean = 10 m/sec).
reflects the host density of the network. Figure 17 shows
that a higher host density will bring down the network life-
time. On the contrary, the AA scheme is almost unaffected.
So a higher host density has a negative effect on survival ra-
tio for our schemes. The reason can be explained as fol-
lows. As the network becomes denser, when a route re-
quest is issued, not only more hosts will help searching for
routes, but also the broadcast cost per individual host will
increase so as to wake up neighboring hosts (note that the
traffic load remains unchanged in this case). In terms of sur-
vival ratio, C(98) outperforms E(714), which in turn out-
performs AA.
As figure 18 shows, a higher node density has a nega-
tive effect on throughput for quorum-based protocols, while
it does not influence the AA scheme much. When the node
density goes higher, broadcast cost will increase because of a
higher wake-up cost. The effect is an increased number of col-
lisions and a lower probability of route establishment. Conse-
quently, the throughput goes down. According to figure 18,
180
the throughput of the AA scheme is the highest, and the
throughputs of the E(714) and the C(98) schemes are very
close. Again, when we consider throughput lifetime, our
quorum-based protocols outperform the AA scheme signifi-
cantly.
7. Conclusions
In this paper, we have addressed the asynchronous power
mode management problem for an IEEE 802.11-based
MANET. We have correlated the problem to the concept
of quorum systems and identified an important rotation clo-
sure property for quorum systems. We have proved that any
quorum system satisfying the rotation closure property can
be translated to an asynchronous power-saving protocol for
MANETs. The purpose of the rotation closure property is
to deal with asynchronism among hosts clocks. Under the
rotation closure property, we have derived a quorum size
lower bound for any quorum system. We have identified a
group of optimal or near optimal quorum systems. Optimal
or near optimal quorum systems are preferable because in
a quorum-based power-saving protocol, the number of bea-
cons sent and the ratio of a host remaining active are both
proportional to the quorum size. We have shown that the
grid quorum system [15], the torus quorum system [12], the
cyclic quorum system [14], and the finite projective plane
quorum system [15] are all optimal or near optimal quo-
rum systems satisfying the rotation closure property. We
have developed theorems to help identify good quorum sys-
tems satisfying the rotation closure property, such as quo-
rum systems with only one member, which are very easy
to maintain. We have further proposed a new e-torus quo-
rum system, which can be translated to an adaptive power-
saving protocol allowing hosts to dynamically tune to differ-
ent quorum systems according to their mobility, so as to trade
neighbor sensibility for power expenditure. Extensive sim-
ulation results have been presented to evaluate these proto-
cols.
References
[1] B. Chen, K. Jamieson, H. Balakrishnan and R. Morris, Span: An
energy-efficient coordination algorithm for topology maintenance in ad
hoc wireless networks, in: Proc. of the International Conference on
Mobile Computing and Networking (2001) pp. 8596.
[2] C.F. Chiasserini and R.R. Rao, A distributed power management policy
for wireless ad hoc networks, in: Proc. of IEEE Wireless Communica-
tion and Networking Conference (2000) pp. 12091213.
[3] C.J. Colbourn and E.J.H. Dinitz, The CRC Handbook of Combinatorial
Designs (CRC Press, 1996).
[4] C.J. Colbourn, J.H. Dinitz and D.R. Stinson, Quorum systems con-
structed from combinatorial designs, Information and Computation
(2001) 160173.
[5] L.M. Feeney and M. Nilsson, Investigating the energy consumption of
wireless network interface in an ad hoc networking environment, in:
IEEE INFOCOM (2001) pp. 15481557.
[6] J. Gomez, A.T. Campbell, M. Naghshineh and C. Bisdikian, A distrib-
uted contention control mechanism for power saving in random-access
JIANG ET AL.
ad-hoc wireless local area networks, in: Proc. of IEEE International
Workshop on Mobile Multimedia Communications (1999) pp. 114123.
[7] J.C. Haartsen, The Bluetooth radio system, IEEE Personal Communi-
cations (February 2000) 2836.
[8] L. Hu, Topology control for multihop packet radio networks, IEEE
Transactions on Communications 41 (October 1993) 14741481.
[9] C.F. Huang, Y.C. Tseng, S.L. Wu and J.P. Sheu, Increasing the through-
put of multihop packet radio networks with power adjustment, in: Proc.
of International Conference on Computer, Communication, and Net-
works (2001).
[10] E.-S. Jung and N.H. Vaidya, An energy Efficient MAC protocol for
wireless LANs, in: Proc. of INFOCOM 2002 (2002).
[11] LAN MAN Standards Committee of the IEEE Computer Society, IEEE
Std 802.11-1999, Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) specifications (IEEE, 1999).
[12] S.D. Lang and L.J. Mao, A torus quorum protocol for distributed mu-
tual exclusion, in: Proc. of the 10th Internat. Conference on Parallel
and Distributed Computing and Systems (1998) pp. 635638.
[13] J.R. Lorch and A.J. Smith, Software strategies for portable computer
energy management, IEEE Personal Communications (June 1998) 60
73.
[14] W.S. Luk and T.T. Wong, Two new quorum based algorithms for dis-
tributed mutual exclusion, in: Proc. of International Conference on
Distributed Computing
Systems (1997) pp. 100106.
[15] M. Maekawa, A N algorithm for mutual exclusion in decentralized
systems, ACM Transactions on Computer Systems (1985) 145159.
[16] C.E. Perkins and E.M. Belding-Royer, Ad-hoc on-demand distance
vector routing, in: Proc. of IEEE Workshop on Mobile Computing Sys-
tems and Applications (1999) pp. 90100.
[17] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
wireless networks using transmit power adjustment, in: Proc. of IEEE
INFOCOM (2000) pp. 404413.
[18] J.H. Ryu and D.H. Cho, A new routing scheme concerning power-
saving in mobile ad-hoc networks, in: Proc. of IEEE International Con-
ference on Communications, Vol. 3 (2000) pp. 17191722.
[19] J.H. Ryu, S. Song and D.H. Cho, A power-saving multicast routing
scheme in 2-tier hierarchical mobile ad-hoc networks, in: Proc. of IEEE
Vehicular Technology Conference, Vol. 4 (2000) pp. 19741978.
[20] A.K. Salkintzis and C. Chamzas, An in-band power-saving protocol
for mobile data networks, IEEE Transactions on Communications 46
(September 1998) 11941205.
[21] E. Shih, P. Bahl and M.J. Sinclair, Wake on wireless: An event
driven energy saving strategy for battery operated devices, in: Proc. of
MOBICOM 2002 (2002).
[22] T. Simunic, H. Vikalo, P. Glynn and G.D. Micheli, Energy efficient
design of portable wireless systems, in: Proc. of the International Sym-
posium on Low Power Electronics and Design (2000) pp. 4954.
[23] S. Singh and C.S. Raghavendra, Power efficient MAC protocol for mul-
tihop radio networks, in: Proc. of IEEE International Personal, Indoor
and Mobile Radio Communications Conference (1998) pp. 153157.
[24] S. Singh, M. Woo and C.S. Raghavendra, Power-aware routing in mo-
bile ad hoc networks, in: Proc. of the International Conference on Mo-
bile Computing and Networking (1998) pp. 181190.
[25] Y.C. Tseng, C.S. Hsu and T.Y. Hsieh, Power-saving protocols for IEEE
802.11-based multi-hop ad hoc networks, in: Proc. of IEEE INFOCOM
(2002).
[26] R. Wattenhofer, L. Li, P. Bahl and Y.M. Wang, Distributed topology
control for power efficient operation in multihop wireless ad hoc net-
works, in: Proc. of IEEE INFOCOM (2001) pp. 13881397.
[27] H. Woesner, J.P. Ebert, M. Schlager and A. Wolisz, Power-saving
mechanisms in emerging standards for wireless LANs: the MAC level
perspective, IEEE Personal Communications (June 1998) 4048.
[28] S.L. Wu, Y.C. Tseng and J.P. Sheu, Intelligent medium access for mo-
bile ad hoc networks with busy tones and power control, IEEE Jour-
nal on Selected Areas in Communications 18 (September 2000) 1647
1657.
[29] Y. Xu, J. Heidemann and D. Estrin, Geography-informed energy con-
servation for ad hoc routing, in: Proc. of the International Conference
on Mobile Computing and Networking (2001) pp. 7084.
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS
Jehn-Ruey Jiang received his Ph.D. degree in com-
puter science in 1995 from National Tsing-Hua Uni-
versity, Taiwan. He joined Chung-Yuan Christian
University as an Associate Professor in 1995. He is
currently an Associate Professor of the Department
of Information Management, Hsuan-Chuang Univer-
sity. He is a recipient of the Best Paper Award in In-
ternational Conference on Parallel Processing, 2003.
His research interests include distributed computing,
mobile computing, distributed fault-tolerance, proto-
cols for mobile ad hoc networks and wireless sensor networks.
Yu-Chee Tseng received his B.S. and M.S. degrees
in computer science from the National Taiwan Uni-
versity and the National Tsing-Hua University in
1985 and 1987, respectively. He worked for the
D-LINK Inc. as an engineer in 1990. He obtained
his Ph.D. in computer and information science from
the Ohio State University in January of 1994. From
1994 to 1996, he was an Associate Professor at the
Department of Computer Science, Chung-Hua Uni-
versity. He joined the Department of Computer Sci-
ence and Information Engineering, National Central University in 1996, and
has become a Full Professor since 1999. Since August 2000, he has become
a Full Professor at the Department of Computer Science and Information En-
gineering, National Chiao-Tung University, Taiwan.
Dr. Tseng served as a Program Chair in the Wireless Networks and Mobile
Computing Workshop, 2000 and 2001, as a Vice Program Chair in the Inter-
national Conference on Distributed Computing Systems (ICDCS), 2004, as
an Associate Editor for The Computer Journal, as a Guest Editor for ACM
Wireless Networks special issue on Advances in Mobile and Wireless Sys-
tems, as a Guest Editor for IEEE Transactions on Computers special on
Wireless Internet, as a Guest Editor for Journal of Internet Technology
special issue on Wireless Internet: Applications and Systems, as a Guest
Editor for Wireless Communications and Mobile Computing special issue on
Research in Ad Hoc Networking, Smart Sensing, and Pervasive Comput-
ing, as an Editor for Journal of Information Science and Engineering, as
a Guest Editor for Telecommunication Systems special issue on Wireless
181
Sensor Networks, and as a Guest Editor for Journal of Information Science
and Engineering special issue on Mobile Computing.
He is a two-time recipient the Outstanding Research Award, National Sci-
ence Council, ROC, in 20012002 and 20032005, and a recipient of the Best
Paper Award in International Conference on Parallel Processing, 2003. His
research interests include mobile computing, wireless communication, net-
work security, and parallel and distributed computing. Dr. Tseng is a Senior
Member of the IEEE.
Chih-Shun Hsu received his B.S. degree in com-
puter education from National Taiwan Normal Uni-
versity, Taiwan, in 1990, and the M.S. degree in
computer science from National Taiwan University,
Taiwan, in 1992. He joined the faculty of the De-
partment of Information Management, Nanya Insti-
tute of Technology, Taiwan, as an instructor in 1996.
He is currently a Ph.D. candidate of the Department
of Computer Science and Information Engineering,
National Central University, Taiwan. His current re-
search interests include wireless communications and mobile computing.
Ten-Hwang Lai received his Ph.D. in computer sci-
ence from University of Minnesota in 1982. Since
then he has been on the faculty of computer and in-
formation science at the Ohio State University, where
he is presently a professor. His research interests
include parallel and distributed computing, mobile
computing, and wireless networking.
Lai was on the editorial board of IEEE Transac-
tions on Parallel and Distributed Systems from 1993
to 1996, and is currently an editor of ACM/Kluwer
Wireless Networks, Journal of Information Science and Engineering, and Wi-
ley Encyclopedia of Computer Science and Engineering. He served as pro-
gram chair and general chair of the 1998 and 2000 International Conference
on Parallel Processing, respectively; and has been designated as program co-
chair and general chair of the 2004 and 2005 IEEE International Conference
on Distributed Computing Systems, respectively.
 Mobile Networks and Applications 10, 183197, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

CROMA An Enhanced Slotted MAC Protocol for MANETs

MARCEAU COUPECHOUX
Institut Eurecom and Alcatel Research & Innovation, France Route de Nozay, 91461 Marcoussis cedex, France

BRUNO BAYNAT
University Pierre et Marie Curie, Paris, France

CHRISTIAN BONNET
Institut Eurecom, Sophia-Antipolis, France

VINOD KUMAR
Alcatel Research & Innovation, Marcoussis, France

Abstract. TDMA based MAC protocols can provide a very good utilization of the shared radio resources, especially at high input loads, in
synchronized mobile ad hoc networks (MANETs). Global positioning systems like GPS or GALLILEO should provide a very good timing
accuracy for synchronization of nodes. This paper presents a new medium access protocol for mobile ad hoc networks, called CROMA.
CROMA is collision-free and receiver-oriented. It operates in a slotted environment, in a dynamic and distributed way. In this protocol,
receivers act as local base stations and can manage one or several communications on a single slot. Thus, sophisticated functions are allowed
at higher layers. Moreover, the hidden terminal as well as the exposed terminal problems are handled by CROMA. A theoretical analysis
and extensive simulations show that CROMA can reach very high throughputs.
Keywords: mobile ad hoc networks, MAC, conflict-free protocol, scheduling, dynamic slot allocation, TDMA

1. Introduction a multi-hop environment. According to [29], 802.11 has still

In recent years a lot of effort has been spent in the design of
protocols for mobile ad hoc networks. Such packet networks
are mobile and multi-hop and operate without any fixed in-
frastructure. This can be a low cost and easily deployable
technology to provide high speed Internet access in a wireless
environment, to organize networks of sensors, or to comple-
ment the coverage of future cellular networks.
In this paper, we pay special attention to the medium ac-
cess control (MAC) sub-layer. It has a lot of impact on the
system performance and its design is a very challenging is-
sue. MAC should control access to the medium and share the
channel between sourcedestination pairs and/or flows of data
in a dynamic and distributed way. Some desirable features of
the access protocol are: to be able to reuse the resources as
efficiently as possible, to avoid congestion and collisions, to
be fair, reliable, and energy efficient.
Many MAC protocols try to address these issues. In the lit-
erature two categories of schemes have been proposed: (i) the
contention based schemes; (ii) the conflict-free schemes.
In the contention based protocols, the channel has to
be acquired by the nodes for each packet to be transmit-
ted. Examples of contention based schemes are CSMA/CA,
MACA [18], MACAW [5], FAMA [14], IEEE 802.11 [1].
The latter seems to be very popular in most of the testbeds
because the IEEE 802.11 family products are available off
the shelf. Although IEEE 802.11 is flexible, robust and sim-
ple, a recent paper [29] claims that it may not do very well in
the hidden terminal problem, does not handle the exposed ter-
minal problem at all and its backoff strategy leads to severe
unfairness. In this family of protocols, MACA-BI [26] was
the first one to be receiver oriented, i.e., the transmission of
a packet is initiated by the receiver that sends a short control
packet in order to reserve the channel and to invite the sender
to transmit. As the receiver does not have the exact knowl-
edge of packet queue at the sender, it must rely on a traffic
prediction algorithm.
On the other hand, conflict-free protocols allow the reser-
vation of the channel for a certain amount of time or data and
transmissions are conflict-free. TDMA deterministic schedul-
ing may be preferred for networks with heavy load, carrying
mixed traffic and realizing sophisticated functions at higher
layers. That is the reason why we propose in this paper a slot
allocation protocol for mobile ad hoc networks.
Unfortunately, most of the scheduling problems are NP-
complete. For example, Arikan [2] has shown that construct-
ing an optimal schedule for the point-to-point scheduling
problem to optimize throughput is NP-complete. And this
is the same for the broadcast scheduling problem based
on throughput optimization, as proved by Ephremides and
Truong [12]. Consequently, MAC designers have focused on
sub-optimal, dynamic and decentralized solutions for the slot
assignment problem.
A first class of scheduling protocols relies on the alloca-
tion of priorities to nodes. A given slot is assigned preferably
to the node with the highest priority according to its offered
184
traffic. Slots can be allocated by using a control channel, e.g.,
in [7]. Priorities of the neighbors are assumed to be known at
each node and are allocated in a pseudo-random way as in [3].
Then different strategies can be applied for the allocation of
the priorities in order to have a fair and efficient share of the
channel (see, e.g., [23]). However, some of these protocols
suffer from a high overhead due to the control channel. Oth-
ers do not address the problem of the distributed and dynamic
assignment of priorities.
On the other hand, time-spread protocols seem to be very
attractive because they are topology-independent (see, e.g.,
[6] or [17]). However, the frame length makes them less
scalable and this class of protocols also faces the problem of
distributed and dynamic code assignment.
At last, the necessity to address the problem of mobil-
ity, topology changes, and scalability, gives rise to a family
of protocols where the reservation of the slots is done via a
random access, most of the time a handshaking, combined
with a carrier sensing mechanism. FPRP [30] proposes a
five-phase handshaking supported by a pseudo-Bayesian al-
gorithm to enable a faster convergence of the reservation
procedure. CATA [27] uses four mini-slots in each time-slot
to enable unicast and multicast transmissions. The protocol
proposed in this paper comes within this family of protocols.
It tries to make use of the advantages of the most popular
contention based protocols to a slotted environment in order
to increase their efficiency. In particular, the aim of CROMA
is to achieve a high slot utilization, i.e., a high capacity, at
high input load thanks to an original reservation and polling
scheme.
The paper is organized as follows. In section 2, we give
a precise description of our proposed MAC protocol. We ex-
amine the correctness of this protocol in section 3. Section 4
gives an analytical study of the protocol in a fully connected
network. At last, section 6 is the conclusion of the paper.
2. Protocol description
The Collision-free Receiver-Oriented MAC (CROMA) is a
medium access protocol for mobile ad hoc networks that
schedules transmissions in a slotted environment. It is a
dynamic and distributed protocol that operates on a single-
frequency channel with omni-directional antennas. CROMA
has been shortly presented in [9] and [8]. The present pa-
per gives a full description of the protocol, integrates new
advanced features, and provides an enhanced performance
analysis.
In CROMA, time is divided into frames, each of them di-
vided into a fixed number L of time-slots. Each slot can be
temporarily and locally attributed to the receiver of a com-
munication link depending on topology changes and traffic
patterns. When a receiver is occupying a slot, it is allowed
to poll several senders among its neighbors. The number of
current communications for each slot is, however, limited by
the protocol to a pre-defined value K.
The polling packet sent by the receiver is used to reserve
the channel and to invite a sender to send a data packet. In that
COUPECHOUX ET AL.
Figure 1. Frame structure of CROMA.
sense, CROMA is a receiver-oriented protocol since a slot in
the frame is associated to a single receiver.
CROMA does not rely on a traffic prediction algorithm
at the receiver. Indeed, a requesting node has to reserve re-
sources at its intended receiver during a random access phase.
This reservation is needed only at the beginning of a packet
train (or message). When a receiver has no longer traffic to
poll, communications are released and the slot is free for an-
other receiver.
2.1. Frame structure
CROMA divides time into frames that are, in turn, divided
into L equal time-slots. All mobile nodes are assumed to be
perfectly synchronized.
Synchronization is a very critical issue for CROMA as for
all distributed TDMA systems. A possible solution, now at
low cost, consists in making use of the GPS (Global Posi-
tioning System) that provides a global synchronization for
all nodes. Also the European satellite navigation system,
GALILEO, will provide a very good timing accuracy [13].
In this case, guard intervals have to be foreseen. Another
way of research is local synchronization, where nodes try
to synchronize themselves by exchanging beacons with their
neighborhood [10,11]. The algorithms proposed in the lit-
erature can be adapted in order to be used with CROMA.
However, as in [30] and [27], this paper focuses on the
protocol description and considers that synchronization is a
realistic assumption.
Throughout this paper, the following terminology has been
chosen. A requesting node is a node that has data packets to
send but has not yet succeeded in the reservation phase, its
intended receiver is the destination node of these data packets.
A sender is a node that succeeded in the reservation phase and
that transmits data packets when it is polled by the receiver.
A receiver is a node that polls senders on a given slot. At
last, we will clearly distinguish the sender/receiver pair of a
communication as defined earlier from the source/destination
pair of a packet, that can be different for control packets.
Each time-slot is divided into three parts: two mini-slots,
called REQ-mini-slot (request) and RTR-mini-slot (ready to
receive) for the signaling, and a data transmission phase,
called DATA-mini-slot (see figure 1).
The REQ-mini-slot is used by requesting nodes during the
random access phase for sending a REQ to its intended re-
ceiver. The RTR-mini-slot is used by their intended receivers
to acknowledge requests as well as previous data transmis-
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS
Figure 2. Example of two parallel connections on a slot with CROMA.
sions, and to poll one of the senders that previously managed
a successful reservation. During the DATA-mini-slot, the
sender that has been polled in the RTR-mini-slot transmits a
data packet. These data packets are of fixed length. Indeed, it
is assumed that a higher layer is responsible for fragmentation
and reassembly.
2.2. CROMA from an example
Before going into more details in the protocol description,
let us illustrate the key feature of CROMA that is to allow
multiple reservations on the same slot. The receiver indeed
maintains a list of senders that managed a successful reserva-
tion and will poll them in the successive frames. This feature
is illustrated in figure 2, which shows two successive reserva-
tions on the same slot i. In frame j , the REQ/RTR dialogue
starts the connection between nodes A and B: A sends a REQ
packet with its address. B sends back a RTR, that contains a
field to acknowledge the reservation (ackreq), and a field to
poll node A (pol). The RTR is also received by node C that is
now aware of a communication on slot i with B as receiver.
During the data phase, A, that has just been polled by B, is
allowed to transmit a packet to B with its address A and a
sequence number (sn) 0. We say that B has got the floor on
slot i. In frame j + 1, C establishes a connection with B.
With the RTR, node B acknowledges the reservation with the
field ackreq, acknowledges the packet transmitted by node A
in frame j , and polls node C. In frame j + 2, B now polls A.
With the RTR, it also acknowledges the data packet of C with
sequence number 0. In frame j + 3, node B polls node C and
acknowledges the data packet of A with sequence number 1.
2.3. The choice of a receiver-oriented protocol
The choice of a receiver-oriented protocol is justified by the
following arguments:
185
Figure 3. Packet formats of CROMA.
(i) this is a natural choice since only the zone that has to
be secured with respect to collisions is the zone around
the receiver, and thus, the spatial reuse of the radio re-
sources is favored;
(ii) this choice allows the multiplexing of several communi-
cations on a single slot. That implies finer flow control
and QoS negotiation. If a slot is associated to a sender, it
cannot easily multiplex communications with different
receivers since they may not be available because of a
hidden terminal;
(iii) if a slot is associated to a receiver, a current communi-
cation on a given slot does not prevent a random access
on this slot. More bandwidth for the contention for the
channel implies less collisions and interference. If a slot
is associated to a sender, it has to send at each frame a
control packet (RTS) to give the address of its intended
receiver. Moreover, the receiver has to respond with an-
other control packet (CTS) in order to avoid the hidden
terminal problem. In CROMA, once the reservation has
been done, the REQ is not used any more for the dura-
tion of the communication, and the REQ-mini-slot can
be used for new reservations.
2.4. Packet formats
This section describes the different packet formats and the
MAC header of the data packets. It gives also the definition
of all the MAC fields. Their signification will become clearer
in the protocol description (sections 2.52.7).
2.4.1. Common parts
In figure 3, the control packet formats and the MAC header of
the data packets are shown. In all packets, generic informa-
tion not described in this paper, like the protocol version, are
given in the field fc that stands for frame control. The field fcs
(frame check sequence) contains a CRC (cyclic redundancy
code) calculated on all the fields of the MAC header and on
186
the frame body. The field source.ad gives the Ethernet address
of the source of the packet.
Note that all packets, including data packets, have a fixed
size, and each mini-slot is just long enough to allow the trans-
mission of the associated packet. For example, the time to
transmit a REQ including additional bits from the physical
layer, the transmit-to-receive turn around time, as well as a
small time interval to take into account the propagation de-
lays equal the time of the REQ-mini-slot. Note also that it is
preferred that the size of the control packets are short com-
pared to the length of the data packets (e.g., 512 bytes).
2.4.2. REQ control packet
In a REQ, the field dest.ad gives the ethernet address of the
destination of the packet (the intended receiver). The field
qs is used by a requesting node to indicate to the intended re-
ceiver the requested quality of service for the communication.
This field may be used by higher layers to negotiate the QoS.
It will be used in future versions of the protocol.
2.4.3. RTR control packet
A RTR has three different functions, as illustrated in sec-
tion 2.2 and in figure 2: respond to a REQ, poll the different
senders on the current slot and acknowledge data packets.
In the RTR, the fields req.ad and r are used to reply to the
requests sent on the same slot (during the REQ-mini-slot). If
a request is correctly received and accepted, it is acknowl-
edged by putting the address of the requesting node in the
field req.ad and the value ACK in the field r. If a request has
been correctly received, but the communication cannot be es-
tablished, the field r is set to NACK. This situation is possible
if the requested QoS is not allowed or if the number of current
communications has reached its maximum, K. If the receiver
detects a collision of REQs, r is set to COL. If the receiver did
not received any request, or if the request cannot be decoded
because of the channel conditions, r is set to NOTRECVD.
The values NACK, COL, and NOTRECVD are useful infor-
mation for the requesting nodes to reschedule their requests.
The field polled.ad is used by a receiver to poll a sender
that previously managed to establish a connection on this slot.
If a sender reads its address in the field polled.ad, it is allowed
to send a data packet during the DATA-mini-slot of the same
slot, just after receiving the RTR.
The acknowledgement of data packets is done thanks to
the field sn that stands for sequence number. Each node main-
tains a counter that is incremented for each new data packet.
Receivers keep the last received sequence number. If in time-
slot i of frame j , a receiver has received a data packet with
sequence number m, it sets the field sn to m in the RTR of the
slot i of frame j + 1 and so, acknowledges the previous data
packet.
The byte n of a RTR gives information about the slot uti-
lization. It is decomposed into seven bits that indicate the
number k of current communications, and one bit t to inform
that the receiver will not accept requests on this slot anymore.
More details on the use of the bit t for fairness are given in
COUPECHOUX ET AL.
section 2.8. If k has reached the maximum K or if the bit t is
set to 1, no more request can be done on this slot.
2.4.4. Data packets
In data packets, the field dest.ad gives the address of the des-
tination of the packet.
As previously explained, each sender maintains a counter
that is incremented for each new packet. This sequence
number is put in the field sn and is used by the receiver to
acknowledge the reception of the packet. Let us recall that
data packets have a fixed size, that results of a higher layer
segmentation or aggregation.
2.5. Reservation
Any communication between two nodes must be preceded by
a preliminary reservation phase. In the reservation phase, re-
questing nodes contend to get access to a receiver. This access
is done in a random way during the REQ-mini-slots and con-
sists of five sub-phases: listening of an entire frame, choice
of a time-slot, transmission of the REQ on the chosen slot,
listening of the RTR, and retry of a new reservation phase in
case of failure (with or without random backoff). These five
sub-phases are now detailed.
2.5.1. Frame listening
The first phase of the reservation consists in listening to the
RTR-mini-slots during an entire frame, and maintaining for
each slot in the frame the state of the slot. This listening
process starts at the beginning of the reservation phase and
lasts until the reservation has succeeded.
A slot can be in several states:
FREE: no activity has been sensed during the RTR-mini-slot,
i.e., no receiver has got the floor on this slot. A request will
be possible on this type of slot.
OCC-NA: i.e., occupied and not available. This is the case
if a RTR has a source.ad different from the address of the
intended receiver or if the requesting node has detected a
collision during the RTR-mini-slot, or if it did not man-
aged to decode the field source.ad in the RTR, or if the
requesting node is itself a receiver on this slot. This is also
the case if the field k of byte n has reached the maximum
number of communications on a slot or if the bit t of byte
n is equal to 1. Note that a RTR collision detected on a slot
does not necessary mean that the slot is free in a multi-hop
situation. A request will not be possible on this slot.
OCC-A-COL-k: i.e., occupied, available, collision, and k
communications. In this case, the source.ad of the RTR
is the address of the intended receiver, a collision has
been detected by the receiver during the REQ-mini-slot
(r = COL in the RTR), and there are currently k < K
communications on the slot. A request will be possible on
this slot.
OCC-A-NCOL-k: i.e., occupied, available, no collision, and
k communications. In this case, the source.ad of the RTR
is the address of the intended receiver, no collision has
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 187

Table 1
Decision of a requesting node after listening to the RTR-mini-slot.

Reception req.ad r Decision

RTR decoded my_address ACK Enter the transmission phase

my_address NACK Retry on next frame
not my_address Retry on next frame
broadcast_address NOTRECVD Retry on next frame
broadcast_address COL Start backoff algorithm
RTR not received or decoded Retry on next frame

been detected by the receiver during the REQ-mini-slot replies with an RTR whether it can decode or not the
(r
= COL in the RTR), and there are currently k < K REQ.
communications on the slot. A request will be possible on
this slot. 2.5.4. Listening of the RTR and decision
A requesting node that has sent a REQ during the first mini-
It is important to emphasize that the slot states are up-
slot of the chosen slot listens to the following RTR-mini-slot.
dated continuously during the whole reservation phase. In
Table 1 gives a summary of the decisions of the requesting
order to reduce the energy consumption, slot states updates
node after the RTR-mini-slot.
can be, however, limited to a few frames before the reserva-
If the field req.ad has been set to its address and r to ACK,
tion process.
the requesting node enters the transmission phase. If r indi-
cates a collision, the random backoff algorithm is started. In
2.5.2. Choosing a time-slot
all other cases, the requesting node is allowed to restart the
The choice of the time-slot depends on the chosen scheduling
reservation phase at the next frame. The random backoff al-
policy. This policy may have several objectives. For example,
gorithm is thus only used when a high load is detected for the
it may maximize the slot utilization, limit the amount of inter-
intended receiver.
ference in the network, establish connections that are robust
to mobility. The impact of this choice is not detailed in this
2.5.5. Backoff algorithm
paper. We present here a simple policy that favors free slots
The backoff algorithm starts when a requesting node has been
first and therefore, aims at maximizing the slot utilization:
informed that a collision occurred. An integer BO is randomly
chosen between 1 and BACKOFFWND. This is a timer that is
1. If there is at least one slot in state FREE,
decremented at the beginning of each frame and each time
choose one randomly and exit, otherwise go to step 2.
the requesting node senses a slot in state OCC-A or FREE.
2. If there is at least one slot in state OCC-A, As soon as BO reaches 0, a slot is chosen on the forthcoming
select the slots having the lowest value of k. Among slots in this set:
frame according to the scheduling policy for a new request.
(a) if there is at least one slot in state OCC-A-NCOL, With this algorithm, the load on the available slots is taken
choose one randomly and exit, otherwise go to step 2(b); into account.
(b) otherwise, choose one slot in state OCC-A-COL randomly and exit; The parameter BACKOFFWND is increased by a multi-
plicative factor (1.5) at each successive retransmission and
Otherwise restart the reservation phase at the next frame.
decreased by one at each success. However, there are a lower
and an upper bound for it, called BOmin and BOmax, e.g., 2
and 32.
2.5.3. Transmission of the request and RTR generation
On the chosen slot, the reservation is done by sending a REQ 2.6. Transmission
during the REQ-mini-slot. Two cases must now be consid-
ered: A sender whose request has been successful on a given slot
starts its transmission phase. During a transmission phase, re-
(i) The sender has chosen a free slot. If the intended re-
ceivers of which resource has been reserved in the reservation
ceiver can decode the REQ, it replies to the request by
phase, do a polling among their associated senders. When
sending an RTR in the same slot and by using the fields
a sender recognizes its address in the field polled.ad of the
req.ad and r of this packet, as explained in section 2.4.3.
RTR, it sends in the same slot a data packet during the DATA-
Otherwise, the intended receiver does not reply. (Note,
mini-slot.
however, that the intended receiver may be aware that the
Each sender maintains a counter of its transmissions that
slot is occupied, which can happen in a hidden terminal
is incremented at each new packet. This sequence number is
configuration. In this case, the receiver does not answer
copied in the field sn of the packet header. With this method,
to the request. See section 3 for more details.)
the receiver is able to acknowledge the last correctly received
(ii) The sender has chosen a slot that is already occupied by data packet. For that, a receiver copies in the field sn the se-
the intended receiver. In this case, the intended receiver quence number of the last received packet. At the sender side,
188
Figure 4. Polling during the transmission phase.
a sent data packet is stored until the receipt of the acknow-
ledgement. If the next RTR is not received or if this RTR does
not acknowledge the stored packet, a retransmission is nec-
essary. After M retransmissions the stored packet is thrown
away. This loss can be treated by an upper layer.
Figure 4 shows an example of a transmission phase with
a receiver and three senders. It only shows slots i of suc-
cessive frames. On the upper part of the figure, the RTRs of
the receiver are represented with the fields polled.ad and sn.
A cyclic polling is pictured for the scheduling of the senders
and data packets are shown with their field sn.
It is clear that each receiver acts on a given slot as a local
base-station with respect to its associated senders. Thus, the
polling mechanism allows a high flexibility for the schedul-
ing of different flows by higher layers and is a base for the
implementation of QoS algorithms. Moreover, several paral-
lel communications are possible on a given time-slot.
2.7. Release
An established communication can be interrupted in three
cases.
(i) The sender informs the receiver that it sends the last
packet of the communication by setting the field sn of
the packets header to the value EOT (end of transmis-
sion). If the last packet is correctly received, the receiver
does not re-schedule the sender any more. However, it
acknowledges the last packet with its next RTR, and this,
even if it has no sender to poll.
(ii) If a receiver has polled a sender and does not receive
any packet from the sender, a counter set to W is
decremented. When this counter reaches 0, the commu-
nication is released, and the receiver does not poll the
sender any more. If after a poll, a packet is received, the
counter is set again to W . After each polling, a sender
starts a timer. If it does not receive any polling from the
receiver when the timer expires, the connection is con-
sidered to be broken.
(iii) During a communication, a sender may receive several
RTRs, i.e., there is a collision of RTRs. In this case,
the sender considers that the current communication on
this slot is released. Indeed, sending a data packet could
imply a collision during the DATA-mini-slot. More pre-
cisions about this specific aspect are given in section 3.
COUPECHOUX ET AL.
2.8. Fairness issue
CROMA includes a mechanism to ensure a local fairness
among data flows. On a given time-slot, fairness among in-
coming flows is assured by the receiver of the slot by means
of the RTRs. By using different polling strategy, a receiver
can easily give a fair allocation of the slot to incoming flows.
However, if the number of slots in the frame is small
compared to the number of potential receivers, situations of
unfairness can arise and flows can be completely starved. The
bit t included in the RTRs is used in order to avoid such situ-
ations.
A receiver having the floor on a given slot counts the num-
ber of consecutive full frames. A frame is full from the point
of view of a receiver, if it senses activity at each slot of the
frame. In this case, it detects a potential blocking situation
for pair of nodes that cannot communicate because there are
no free slots any more. If the number of monitored full frames
reaches MAX_FULLFRAMES, the receiver sets the bit t to 1
indicating that it will not accept new requests and that the cur-
rent communications have to be released.
A sender detecting a bit t set to 1, sets the field sn of its
next packet header to EOT and stops sending packets to the
receiver. This release is done even if the sender have still
packets to transmit. A requesting node detecting a bit t set to
1 in a RTR update the slot state to OCC-NA.
This strategy aims at avoiding blocking situations that can
lead to unfairness. Indeed, these cases are detected by the
receivers that have to free their slot if the situation lasts.
3. Correctness
In this section, we will show that CROMA is correct, i.e., that
it is collision-free in both fixed and mobile environment. The
capture effect is not considered here, so this section shows
that CROMA is collision-free in the common case provided
that a sender releases its communication as soon as it detects
a collision of RTRs.
Let us first consider a fixed and multi-hop topology. We
now prove that two data packets cannot collide.
We suppose that a collision of two data packets occurs at
a receiver R1 . These packets have been sent by two differ-
ent senders, namely S1 and S2 . During the RTR-mini-slot, R1
specified the MAC address of the sender, say S1 , that was al-
lowed to send its data in the current slot. As the MAC address
is unique, a single colliding data packet is destined to R1 .
Therefore, we know that the data packet of S2 was destined to
another receiver, R2 .
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS
Figure 5. Interference between two communications sharing the same slot.
Now, as R1 has received a data packet from S2 and links
are bi-directional, S2 has received the RTR of R1 . More-
over, S2 has also received a RTR from R2 , since it sent a data
packet destined to R2 . Thus, S2 has detected a collision of
RTRs in the current slot without interrupting its communica-
tion with R2 . This is impossible. As a conclusion, no data
collision can occur in a fixed topology.
Let us now consider the case of a dynamic topology. Two
concurrent communications on a slot are shown on the top of
figure 5, from node 1 to node 2 and from node 3 to node 4.
These communications are sharing the same slot in frame j
and they are far away enough, so that they do not interfere. In
case of mobility and at the next frame j + 1, node 3 can either
stay out of range of nodes 1 and 2, enter the communication
range of 1, 2, or both 1 and 2. Same alternatives can occur for
node 4. Thus, after mobility, a total of 16 relative new posi-
tions are possible. Because of the symmetry of the problem,
only 10 cases are shown in figure 5.
The left-hand side of figure 5 shows situations, where
a single communication is interrupted because the sender
detected a collision of RTRs on the considered slot. For ex-
ample, in case (b), node 4 moved in the transmission range of
nodes 1 and 2. In frame j + 1, nodes 2 and 4 send simulta-
neously an RTR. Node 3 receives correctly the polling of 4,
whereas node 1 senses a collision during the RTR-mini-slot.
Node 1 decides to interrupt the communication with node 2
and does not send any data packet on this slot. If node 1 has
still packet in its buffer, it has to enter a new reservation phase.
The central part of figure 5 shows exposed-terminal
topologies, where both communications can still share the
same slot. In case (e), node 4 moved in the transmission range
of node 2. In frame j + 1, node 1 (resp. 3) decodes the RTR
of node 2 (resp. 4) because it is out of the transmission range
of node 4 (resp. 2). Both nodes 1 and 3 can send data packet
during the DATA-mini-slot.
189
The righthand side of figure 5 shows topologies, where
communications are released because both senders detected
a collision of RTRs. Case (j) shows a configuration where
the network of nodes is fully connected after mobility. Here,
RTRs of nodes 2 and 4 collide at nodes 1 and 3. On detecting
the collision, they decide to interrupt their communication.
So, in the common case, in both fixed and mobile environ-
ment, CROMA is collision-free. As in all protocols that rely
on the exchange of short control packets, the capture effect
may, however, affect this conclusion.
4. Analytical study
In this section we calculate the approximate throughput, i.e.,
the slot utilization of the protocol CROMA in a fully con-
nected network. Following [27], we claim that this topology
is the worst case in terms of interference, contention, and spa-
tial reuse because CROMA guarantees a collision-free trans-
mission of data after reservation in a multi-hop environment.
4.1. Model for the slot utilization analysis
First of all, we describe our analytical model for the slotted
MAC protocol CROMA. From this model will be derived the
slot utilization of CROMA as a function of the probability
p to send a REQ for a given sourcedestination pair. Lets
enumerate the assumptions of our model.
1. We consider a fully-connected network of N synchro-
nized nodes.
2. All packets are of constant length and are transmitted
over an assumed noiseless channel.
3. There are L slots per frame.
4. The maximum number of connections on a slot is K, i.e.,
when a receiver is already polling K different senders on
a slot, no new REQ is allowed.
190 COUPECHOUX ET AL.

5. A receiver can only be associated with a single slot. This

hypothesis can be in practice relaxed, but for the sake
of tractability of the model, we limit the analysis to this
case.
6. A node can be a sender on several slots of the frame.
While being in communication on a slot, a node can send Figure 6. Discrete time Markov chain representing the state of the slot, for
K
N.
a REQ on another slot of the frame to start another con-
nection.
Results will show that this assumption is a good approxima-
7. The traffic between any two nodes s and d is a ON/OFF tion.
traffic.
8. The ON periods are modeled by bursts of packets follow- 4.2. One slot analysis
ing a geometrical distribution. The length of a message
follows a geometrical law with parameter q. Thus, the In this section L = 1. In this simple case, we can derive a
average message length (AML) is 1/(1 q). closed-form formula for the slot utilization.
The system is described by the number of parallel con-
9. The OFF periods are modeled by series of slots without nections on the considered slot at the end of the frame (the
transmission following a geometrical distribution. If a DTMC is shown in figure 6). Lets now compute the transi-
source s does not communicate with a destination d, there tion probabilities ri,j of this Markov chain. Remember that
is a probability p that s wants to communicate with d at
the probability for a sourcedestination pair to enter a ON pe-
the next frame.
riod is p. Thus, the probability that a node sends a request on
10. A non-persistent policy is assumed for retransmissions a free slot is the probability that this node has a request for at
after a failure. This hypothesis explains that we can con- least one of the destinations:
sider a fixed probability p to start a communication.
p = 1 (1 p)N1 . (2)
The system is described by the number of parallel connections
on the slots at the end of the frame, (a0 , a1 , . . . , aL1 ), where: Thus, on a free slot, a successful reservation occurs iff only
one single node among N is sending a request during the
ai is the number of current connections on slot i;
REQ-mini-slot. Consequently the probability to have a suc-
0
ai
MIN(K, N 1) (see assumptions 1 and 4); cessful reservation on a free slot is

L1
S =  
i=0 1{ai >0}
MIN(N, L), (see assumptions 3 N  N1
and 5). (0) = p 1 p . (3)
1
For the sake of simplicity, the states describe neither the
receiver associated to each slot, nor the list of associated On an occupied slot with n connections, a receiver has
senders. The vector (a0 , a1 , . . . , aL1 ) is a discrete-time sto- got the floor on the slot and successively polls n senders that
chastic process, whose state space is also discrete. Moreover, managed to reserve resources. Here, a successful reservation
this process is independent of its history because the geomet- occurs iff only a single node among the N (n + 1) nodes
ric law is memoryless. Consequently, this process is a discrete not currently in connection is sending a request. Therefore,
time Markov chain (DTMC). Since the state space is aperi- the probability to have a successful reservation on an occu-
odic and finite, the chain is always ergodic. pied slot is
From a frame to another, we can have the following tran-  
N (n + 1)
sitions on slot i: (n) = p(1 p)N(n+1)1 . (4)
1
ai ai + 1 (ai < K): a reservation has been successful
In state 0
n < K, there is a transition to state n + 1
on slot i AND no communication has come to the end;
iff a successful request is received and this is not the end of
ai ai : (there is a successful reservation AND this is the current communication. The transition state rn,n+1 is thus
the end of a communication) OR (there is no successful given by
reservation AND no message is ending);
rn,n+1 = (n)q. (5)
ai ai 1 (ai > 0): there is no successful reservation
AND this is the end of a communication. In state 0 < n < K, there is a transition to state n
A transition probability between the two states (a0 , a1 , . . . , 1 iff there is no successful request and this is the end of a
aL1 ) and (b0 , b1 , . . . , bL1 ) is assumed to be the product of communication, so
the transition probabilities associated to each slot:  
rn,n1 = 1 (n) (1 q). (6)
 
P (a0 , a1 , . . . , aL1 ) (b0 , b1 , . . . , bL1 )
From these two equations, we obtain directly rn,n for 0 < n

L1 < K:
= P (ai bi ). (1)
i=0 rn,n = 1 rn,n+1 rn,n1 . (7)
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 191

Figure 7. Slot utilization vs. input load, L = 1, N = 5, K = 3. Figure 8. Average number of connections vs. input load, L = 1, N = 5,
K = 3.

In state 0, the slot is free and so r0,1 = (0) and r0,0 =

1 r0,1 . In state K, rK,K = 1 rK,K1 . The transition 4.3. Multi-slot analysis
matrix is given by
In this section, we extend the previous result to the general
P = {ri,j }0
i,j
K . (8) case with L slots. We first compute the transition probabil-
The steady state probabilities are obtained by solving the ities, while distinguishing an occupied slot, a free slot and a
steady state equations  = P  , that enable to express all full slot. For the sake of readability, we only consider the case
the probabilities in function of 0 : K
N.
Lets consider a slot i occupied by the receiver d (this is
 n1 n1
 the case, where 0 < ai < K).
0 q (k)
n = , (9) The number of nodes that are likely to send a REQ to d are
1q 1q 1 (k + 1)
k=0 nodes that are currently not in communication with d, their
number is N 1 ai . The probability for such a node s
for all n {1, . . . , K}. The

system is totally described with to send a REQ on slot i is p (see assumption 9). Thus, the
the following equation: K n=0 n = 1. At last, the slot uti-
probability of a successful reservation is:
lization of the protocol is given by U = 1 0 :
 
1 N 1 ai
U =1
K q
n1 n1 . (10) i = p(1 p)(N1ai )1 . (12)
(k) 1
1+ 1
n=1 1q 1q k=0 1(k+1)
Note that if ai = N 1, all nodes have a connection with
Figure 7 shows the slot utilization of CROMA, U , as a func-
the considered receiver, so that there is no REQ on this slot,
tion of the probability p for K = 3, N = 5 and different
and i = 0. Now the probability that a message is ending is
average message length (AML = 2, 10 and 100 packets).
(see assumption 8) 1 q. We can now derive the transition
Dotted curves have been obtained by simulations. These sim-
probabilities for slot i:
ulations reproduce the assumptions of our model. We can see
on the one hand that the approximations of the analysis have
P (ai ai + 1) = i q, (13)
a small impact on the performance evaluation. On the other
hand, it is clear that CROMA can achieve a very high slot P (ai ai ) = i (1 q) + q(1 i ), (14)
utilization provided that the average message length is high. P (ai ai 1) = (1 i )(1 q). (15)
From the DTMC, the average number of connections,
Nc on the slot can also be derived: Lets now consider a free slot i (ai = 0). There are

L1
S = i=0 1{ai >0} occupied slots in the frame, i.e., S re-

K
Nc = nn .
n=0
Figure 8 shows the average number of connections for dif-
ferent AML values. This mean number is clearly related to
the delay of transmission of a burst because the higher the
number of connections on a slot is, the smaller is the resource
allocated to a single connection. Thus, a trade-off has to be
made between slot utilization and delay.
ceivers, since a receiver is associated to a single slot (see
(11)
assumption 5).
On the considered free slot i, N senders are likely to send a
REQ for N S possible receivers. Indeed, a node is allowed to
send traffic to several receivers in parallel on different slots,
so all nodes are likely to start a new communication on i.
Moreover, requests on i can be addressed to any of the N S
nodes that are not receivers on another slot because i is not
attributed.
192 COUPECHOUX ET AL.

Lets consider a node s. The probability that s has n REQ

for the N S possible receivers is
 
N S
p1 (n) = pn (1 p)NSn (16)
n

if s also belongs to the S receivers, and

 
N S1
p2 (n) = pn (1 p)NSn1 , (17)
n

otherwise. Thus, the probability that s has n requests is:

S N S
p(n) = p1 (n) + p2 (n) . (18)
N N Figure 9. Slot utilization vs. input load, L = 3, N = 5, K = 3.

Now, the probability that s sends a REQ on the free slot i is:


NS
= Pr[s sends a REQ on i | s sends n REQ]p(n)
n=1

NS  
n
= min , 1 p(n). (19)
LS
n=1

At last, there are N possible senders like s, so the transi-

tions probabilities for i are:
 
N
P (0 1) = (1 )N1 , (20)
1
P (0 0) = 1 P (0 1). (21)
Figure 10. Slot utilization vs. input load, influence of K, L = 3, N = 5,
Lets at last consider a full slot (ai = K). The transition AML = 10.
probabilities are obvious:

P (K K) = i (1 q) + q(1 i ), (22)
P (K K 1) = 1 P (K K). (23)

The steady state equations  = P are solved using any

numerical method, e.g., the iterative method of GaussSeidel
(see [4] or [25]).
Figure 9 shows the slot utilization of CROMA as a func-
tion of p for different average message lengths. Analysis and
simulations (dotted lines) are compared and the figure shows
a good adequation of the two methods. As for L = 1, we can
see that CROMA can achieve very high slot utilization pro-
vided that the AML is high. Note that values of p near 1 are
not realistic in a real implementation because of the backoff
algorithm. Simulations show that the point of operation of Figure 11. Average number of connections vs. input load, influence of K,
L = 3, N = 5, AML = 10.
a highly loaded CROMA network with backoff is always for
p < 0.5. Figure 10 shows the influence of K on the system
performance. There is a clear gain of channel utilization as
K increases. However, this is obtained at the cost of higher 5. Performance analysis in a multi-hop environment
delays. This is shown in figure 11, where the average num-
ber of connections per slot is plotted. A higher number of In this section, we provide simulation results and the perfor-
connections per slot implies a higher delay for the burst trans- mance of CROMA and of the standard IEEE 802.11 (DCF
missions. mode) are compared.
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 193

Table 2
5.1. Methodology
Main parameter values for simulations.

Studying MAC protocols in a multi-hop environment leads Parameter Value

to the problem of choosing an appropriate node topology. DATA Packet size 512 bytes
Literature on ad hoc networks has solved the problem by BOmin 2
considering on the one hand typical networks, like the string BOmax 64
network, or the grid network, and on the other hand randomly K 3
W 3
generated networks. In this paper, we adopted part of the two M 7
approaches by running CROMA over a classical and chal- MAX_FULLFRAMES 30
lenging network and over a random network. Inter-mini-slot time 10 s
We will now describe the metrics used to evaluate the per- PHY overhead 24 bytes
formance of the MAC protocols. PHY Data Rate 2 Mbps
ON distribution Exponential
End-to-end delay. This is the average time spent by a packet OFF distribution Exponential
Peak Rate 256 Kbps
from the traffic generator of a source to the reception mo-
Mean OFF time 0.5 s
dule of the destination. Simulation time 200 s
End-to-end delay jitter. This is the standard deviation of the Number of simulations per point 10
end-to-end packet delay.
Aggregate throughput. This is the average number of bits
successfully received by all nodes in the network per
second. The input load is the average number of bits trans-
mitted by all nodes per second.
Fairness index. This is the widely used index, f , defined
in [16]. If a system allocates resources to n contending
entities, such that the ith entity receives an allocation xi ,
then: Figure 12. A multihop topology, the squares topology.


( n x i )2
f (x) =
i=1 . (24) arrow between two nodes means that they are in the commu-
n ni=1 xi2 nication range of each other, i.e., the transmissions from one
If all entities get the same amount, i.e., xi s are all equal, of them can be successfully decoded by the other one. A solid
then the fairness index is 1 and the system is 100% fair. line with arrow means that at least one flow of data is using
The choice of the metric depends upon the application. In this link.
our case, we will consider that the entities are the flows of This configuration is interesting for several reasons:
data between sourcedestination pairs (i, j ) and the metric (i) it exhibits a lot of hidden terminal situations. For exam-
is their throughput, Ti,j . ple, nodes 6 and 2 are hidden from node 0, nodes 7 and
3 are hidden from node 5;
5.2. Performance in a challenging environment
(ii) spatial reuse is possible and there are situations of ex-
5.2.1. Throughput and delay analysis posed terminal. For example, nodes 1 and 2 are exposed.
In order to evaluate the performance of CROMA in ad hoc Several flows can share the same slot, e.g., 14 and 27,
networks, we considered a very simple multi-hop situation or 40 and 73;
that has been used in the literature for the evaluation of MAC
(iii) nodes and flows experience different contention situa-
protocols, e.g., in [15]. Nodes are assumed to be static,
tions, nodes 0, 3, 4, and 7 have three neighbors, while
the traffic is ON/OFF with exponential distributions, and the
nodes 1, 2, 5, and 6 have five neighbors.
packet size is set to 512 bytes. Moreover, the channel is sup-
posed to be perfect with a physical data rate of 2 Mbps. The Figures 13 and 14 show the end-to-end packet delay and
transmission area of a node is a disk of radius R. Outside of jitter as function of the input load for IEEE 802.11 and
the transmission area no communication is possible. Simula- CROMA. The different curves for CROMA assume different
tions have been done using the Network Simulator v2 (ns2, number of slots per frame.
see [21]). The simulation parameter values are presented in In the case of low input load, IEEE 802.11 outperforms
table 2. Note that the mean OFF time is fixed and that the CROMA because the low level of contention implies a small
mean ON time will vary in simulations. number of collisions and small backoff windows. At this level
In this configuration, eight nodes form a regular topol- of load, the network cannot fully take advantage of the reser-
ogy, flows of data are shown in figure 12. Four end-to-end vation scheme because trains of packets are small.
communications are running in parallel: 0123, 0527, In the case of higher input load, IEEE 802.11 nodes expe-
7654, and 3614, so that several nodes have to receive rience more contention, and thus more collisions and wider
and/or to relay several flows of data. A solid line without backoff windows: access delay increases drastically. On the
194
Figure 13. End-to-end delay vs. input load, squares topology.
Figure 14. End-to-end delay jitter vs. input load, squares topology.
other side CROMA takes advantage of packet bursts to re-
duce the number of requests per transmitted packet. If a
flow has made a successful reservation, long trains of pack-
ets can be transmitted without contention. Delays and jitters
of CROMA L = 8 remains, however, always above IEEE
802.11 performance. It is clear that CROMA L = 8 is not
well dimensioned for the topology. Actually, the number of
slots is too high and the resource is not fully exploited, as it is
shown in figure 15. To overcome this problem, a higher layer
can split a link layer connection into two separate CROMA
connections. The slot utilization of CROMA L = 8 does
not exceed 0.75. This is much less than CROMA L = 6 that
reaches 0.97. CROMA L = 3 and 4 fully exploit spatial reuse
and exceed 1.1.
The reservation scheme, the synchronization, and the abil-
ity of CROMA to handle the exposed terminal problem allow
the network to achieve high throughputs. Figure 16 shows
aggregate throughput as a function of the input load. IEEE
802.11 saturates at a throughput of 300 Kbps. In compari-
son, CROMA L = 8 achieves a maximum throughput 350
Kbps, although we have seen that it is obviously badly dimen-
sioned for the topology. CROMA L = 6 reaches a maximum
throughput of 425 Kbps. For less slots per frame, a pro-
blem of stability of the throughput arises. Although CROMA
L = 3 and 4 achieve resp. 475 and 510 Kbps, the through-
COUPECHOUX ET AL.
Figure 15. Slot utilization vs. input load, squares topology.
Figure 16. Throughput vs. input load, squares topology.
put decreases for input loads higher than 525 Kbps. Indeed,
the small number of slots implies a slight instability with the
considered topology. However, curves show a slow decrease
leading to acceptable values even at high input load.
5.2.2. Fairness analysis
Without any fairness strategy and without the use of the bit t,
blocking situations can lead to severe unfairness. This is par-
ticularly the case when the input load is high and the number
of slots per frame is small for the considered topology/traffic
pattern. For example, in the topology of figure 12 with L = 4,
if node 1 hears the RTRs of node 2 on slot 0, node 5 on slot 2,
node 6 on slot 3, and sends RTRs on slot 1, 1 cannot send any
REQ since the frame is full. In case of low input load, this
situation is transient and has a low impact on the long term
fairness. In case of high input load, however, the connection
34 is completely starved leading to severe unfairness.
Figures 17 and 18 shows the benefit of use of the bit t
with the aforementioned fairness strategy. The fairness index
of CROMA L = 3 and L = 4 are compared to the index of
IEEE 802.11. For the IEEE standard and for CROMA without
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 195

Figure 19. Random topology with 30 nodes in a 1000 m 1000 m area.

Figure 17. Fairness index vs. input load, L = 3, squares topology.

Figure 20. Throughput vs. input load, random topology.

Figure 18. Fairness index vs. input load, L = 4, squares topology.

the use of the t, the index is close to 1 for low to moderate

input load. After a threshold, the increase of input load leads
to a drop of the index. This threshold is 350 Kbps for IEEE
802.11, and approximately 700 Kbps for CROMA. With the
use of the bit t, the fairness index of CROMA remains always
above 0.95 for both L = 3 and L = 4.

5.3. Performance in a random network

In the previous section, we compared IEEE 802.11 and

CROMA over a simple and pre-defined multi-hop topology.
In this section, we consider a random connex network. 30
nodes are drawn at random in a 1000 m1000 m square area,
each node having a transmission range of 250 m. This net-
work is shown in figure 19. 10 connections are established
between 10 random pairs of nodes. The traffic is assumed to
be exponential ON/OFF with the same parameters as in the
previous section. Figure 20 shows the aggregate throughput
of the network as a function of the input load. While IEEE
802.11 and CROMA L = 8 saturate at a load of approxi-
mately 500 Kbps, CROMA L = 6, L = 4, and L = 3 reach
resp. 600, 700, and 750 Kbps.
Figure 21 shows the mean end-to-end delay of the data
packets as a function of the input load. It is clear that the bet-
ter performance of CROMA in term of throughput is obtained
Figure 21. End-to-end delay vs. input load, random topology.
at the expense of higher packet delays and jitters at low input
load (see figure 22). In this case, IEEE 802.11 outperforms
CROMA. However, CROMA allows to extend the area of ac-
ceptable delay and jitter by one third. For example, CROMA
L = 6 still exhibits delays under 600 ms at an input load of
700 Kbps. Note also that at low input load, the frame length
of CROMA has little influence on the end-to-end delay.
In term of fairness, CROMA still outperforms IEEE
802.11 in a random topology as shown in figure 23. Note that
it is very difficult to get statistically satisfying results over
196
Figure 22. End-to-end delay jitter vs. input load, random topology.
Figure 23. Fairness index vs. input load, random topology.
random topologies because of the simulation time. However,
ten different random connex networks (not shown here) have
been simulated and provide similar results.
6. Conclusion
In this paper, a new MAC protocol, called CROMA has been
proposed for mobile ad hoc networks. CROMA operates in a
slotted environment, it is collision-free and receiver-oriented.
The reservation of resources is made through a random access
phase on each slot of the frame. The transmission is done
thanks to a polling by the receivers. Thus, receivers of a con-
nection act as local base-stations and sophisticated functions
at higher layers can be easily implemented.
The correctness of CROMA has been proven. Even with a
dynamic topology, CROMA handles both the hidden and the
exposed terminal problems.
Theoretical performance analysis and extensive simula-
tions show that CROMA can reach very high throughputs in
a fully connected network provided that the average message
length is large. Moreover, CROMA outperforms IEEE 802.11
at high input loads thanks to a better channel utilization.
COUPECHOUX ET AL.
References
[1] ANSI/IEEE Std 802.11, Part 11: Wireless LAN Medium Access Con-
trol (MAC) and Physical Layer (PHY) Specifications (IEEE, 1999).
[2] E. Arikan, Some complexity results about packet radio networks, IEEE
Trans. Inform. Theory 30(4) (1984) 681685.
[3] L. Bao and J.J. Garcia-Luna-Aceves, A new approach to channel access
scheduling for ad hoc networks, in: Proc. of ACM/IEEE MOBICOM01
(July 2001) pp. 210221.
[4] B. Baynat, Thorie des files dattente, des chanes de Markov aux
rseaux forme produit (Hermes Science Publications, Paris, 2000).
[5] V. Bharghavan, A. Demers, S. Shenker and L. Zhang, MACAW: A me-
dia access protocol for wireless LANs, in: Proc. of ACM SIGCOMM
(August 1994) pp. 212225.
[6] I. Chlamtac, A. Farag and H. Zhang, Time-Spread Multiple-Access
(TSMA) protocols for multihop mobile radio networks, IEEE/ACM
Trans. Networking 5(6) (1997) 804812.
[7] I. Cidon and M. Sidi, Distributed assignment algorithms for multihop
packet radio networks, IEEE Trans. Comput. 38(10) (1989) 1353
1361.
[8] M. Coupechoux, B. Baynat, C. Bonnet and V. Kumar, Modeling of a
slotted MAC protocol for MANETs, in: Proc. of MADNET03 (March
2003) pp. 2124.
[9] M. Coupechoux, C. Bonnet and V. Kumar, CROMA a new medium
access protocol for mobile ad hoc networks, in: Proc. of WTC02 (Sep-
tember 2002), CD ROM.
[10] J. Elson and D. Estrin, Time synchronization for wireless sensor
networks, in: Proc. of the 15th Internat. Parallel and Distributed
Processing Symposium (April 2001) pp. 19651970.
[11] J. Elson, L. Girod and D. Estrin, Fine-grained time synchronization us-
ing reference broadcasts, in: Proc. of the 5th Symposium on Operating
System Design and Implementation (December 2002) pp. 147163.
[12] A. Ephremides and T.V. Truong, Scheduling broadcasts in multihop
radio networks, IEEE Trans. Commun. 38(4) (1990) 456460.
[13] European Commission and ESA, Galileo Mission high level definition,
v3.0, (September 2002), available on http://europa.eu.int/
comm/dgs/energy_transport/galileo/index_en.html
[14] C.L. Fullmer and J.J. Garcia-Luna-Aceves, Solutions to hidden termi-
nal problems in wireless networks, in: Proc. of ACM SIGCOMM97
(September 1997) pp. 3949.
[15] J.J. Garcia-Luna-Aceves and C.L. Fullmer, Floor Acquisition Multiple
Access (FAMA) in single-channel wireless networks, Mobile Networks
and Applications 4(3) (1999) 157174.
[16] R. Jain, D. Chiu and W. Hawe, A quantitative measure of fairness
and discrimination for resource allocation in shared computer systems,
DEC Research Report TR-301 (September 1984).
[17] J.-H. Ju and V.O.K. Li, An optimal topology-transparent scheduling
method in multihop packet radio networks, IEEE Trans. Networking
6(3) (June 1998) 298306.
[18] P. Karn, MACA a new channel access method for packet radio, in:
Proc. of ARRL/CRRL (April 1990) pp. 134140.
[19] Y.H. Kwon and D.C. Lee, An uplink packet relay protocol for CDMA
cellular-like systems, in: Proc. of MILCOM02, Vol. 2 (October 2002)
pp. 940945.
[20] H. Luo, S. Lu and V. Bharghavan, A new model for packet scheduling
in multihop wireless networks, in: Proc. of ACM/IEEE MOBICOM00
(August 2000) pp. 7686.
[21] ns2, http://www.isi.edu/nsnam/ns
[22] L.C. Pond and V.O.K. Li, A distributed time-slot assignment proto-
col for mobile multi-hop broadcast packet radio networks, in: Proc. of
IEEE MILCOM89, Vol. 1 (October 1989) pp. 7074.
[23] S. Ramanathan, A unified framework and algorithm for (T/F/C)DMA
channel assignment in wireless networks, in: Proc. of IEEE INFO-
COM97, Vol. 2 (April 1997) pp. 900907.
[24] R. Ramaswami and K.K. Parhi, Distributed scheduling of broadcasts in
a radio network, in: Proc. of IEEE INFOCOM89, Vol. 2 (April 1989)
pp. 497504.
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 197

[25] W.J. Stewart, An Introduction to the Numerical Solution of Markov
Chains (Princeton University Press, NJ, 1994).
[26] F. Talucci, M. Gerla and L. Fratta, MACA-BI (MACA by invitation)
a receiver oriented access protocol for wireless multihop networks, in:
Proc. of IEEE PIMRC97, Vol. 2 (September 1997) pp. 435439.
[27] Z. Tang and J.J. Garcia-Luna-Aceves, A protocol for topology-
dependent transmission scheduling in wireless networks, in: Proc. of
IEEE WCNC99, Vol. 3 (September 1999) pp. 13331337.
[28] N.H. Vaidya, P. Bahl and S. Gupta, Distributed fair scheduling in a
wireless LAN, in: Proc. of ACM/IEEE MOBICOM00 (August 2000)
pp. 167178.
[29] S. Xu and T. Saadawi, Does the IEEE 802.11 MAC protocol work well
in multihop wireless ad hoc networks?, IEEE Comm. Magazine 39(6)
(2001) 130137.
[30] C. Zhu and M.S. Corson, A Five-Phase Reservation Protocol (FPRP)
for mobile ad hoc networks, in: Proc. of IEEE INFOCOM98, Vol. 1
(March 1998) pp. 322331.
Marceau Coupechoux received M.Sc. in mathe-
matics from the University Pierre et Marie Curie
(Paris) in 1998 and a double Eng. Degree from the
Ecole Nationale Suprieure des Tlcommunications
(ENST), Paris, in 1999, jointly with the University
of Stuttgart in 2000. He joined Alcatel Research &
Innovation in 2000 and is Ph.D. student since 2001
at the Institut Eurecom, Sophia-Antipolis. He is the
author or co-author of 10 conference papers and 5
patents.
E-mail: marceau.coupechoux@alcatel.fr
Christian Bonnet joined Institut EURECOM as an
Associate Professor in 1992. Since 1998 he has
been at the head of the Mobile Communications De-
partment of EURECOM. He teaches distributed and
real-time systems, mobile communication systems,
wireless LANs and protocols for mobility manage-
ment. His main areas of research are Mobility
Management protocols, wireless access to IP Net-
works and data communications in mobile networks
including Mobile Ad Hoc networks. He is currently
participating in research projects related to UMTS in the field of QoS and
Ipv6.
E-mail: christian.bonnet@eurecom.fr
Vinod Kumar received M.Sc., and Ph.D. from
Ecole Nationale Suprieure des Tlcommunications
(ENST), Paris, France, in 1977 and 1980, respec-
tively. He is Director in Alcatel Research & Innova-
tion for projects related to wireless communications.
He has more than 20 years of experience in digital
mobile communications, signal processing and per-
formance evaluation of higher protocol layers, and
he has more than 20 patents, and more than 30 pub-
lications in technical journals, books and conference
proceedings in above mentioned areas. He is Associate Professor in Univer-
sity of Marne La Valle, and Visiting Professor in several other institutes in
France. He is Officiating Secretary of the Wireless World Research Forum
(WWRF).
E-mail: vinod.kumar@alcatel.fr

Bruno Baynat received the M.Sc. degree from the

Institut National Polytechnique de Grenoble in 1988
and the Ph.D. degree from the University Pierre et
Marie Curie in 1991. Presently, he is Matre de Con-
frence (Associate Professor) at the University Pierre
et Marie Curie. His research interests are presently in
the development of models for the performance eval-
uation of communication systems, with applications
to wired networks (Multicast, QoS) and wireless net-
works (Ad-Hoc, GPRS/EDGE/UMTS, Wi-Fi).
E-mail: bruno.baynat@lipg.fr
 Mobile Networks and Applications 10, 199217, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Dynamic Bandwidth Management in Single-Hop Ad Hoc Wireless

Networks
SAMARTH H. SHAH, KAI CHEN and KLARA NAHRSTEDT
Department of Computer Science, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA

Abstract. Distributed weighted fair scheduling schemes for Quality of Service (QoS) support in wireless local area networks have not
yet become standard. Therefore, we propose an Admission Control and Dynamic Bandwidth Management scheme that provides fairness
and a soft rate guarantee in the absence of distributed MAC-layer weighted fair scheduling. This scheme is especially suitable for smart-
rooms where peer-to-peer multimedia transmissions need to adapt their transmission rates co-operatively. We present a mapping scheme to
translate the bandwidth requirements of an application into its channel time requirements. The center piece of our scheme is a Bandwidth
Manager, which allots each flow a share of the channel, depending on the flows requirements relative to the requirements of other flows
in the network. Admitted flows control their transmission rates so they only occupy the channel for the fraction of time allotted to them.
Thus co-operation between flows is achieved and the channel time is fair shared. As the available channel capacity changes and the traffic
characteristics of various flows change, the Bandwidth Manager dynamically re-allocates the channel access time to the individual flows.
Our simulation experiments show that, at a very low cost and with high probability, every admitted flow in the network will receive at least
its minimum requested share of the network bandwidth. We also present extensive testbed experiments with our scheme using a real-time
audio streaming application running between Linux laptops equipped with standard IEEE 802.11 network cards.
Keywords: wireless single-hop ad hoc network, distributed weighted fair scheduling, bandwidth manager, maxmin fairness

1. Introduction and motivation
In recent times, much effort has gone into solving the prob-
lem of transmitting multimedia data over wireless networks.
Three mutually orthogonal factors make this problem chal-
lenging: (a) stringent QoS requirements of multimedia ap-
plications, (b) bursty nature of some multimedia traffic, and
(c) unreliable and dynamic nature of the wireless network.
Network-specific QoS requirements of multimedia applica-
tions include minimum throughput, maximum delay and
maximum jitter.
In a wireless network, the minimum throughput require-
ment is more difficult to achieve than in a wireline net-
work, because (a) this requires distributed co-operation be-
tween nodes sharing a wireless channel, and (b) the flows in
the wireless network are exposed to various physical chan-
nel errors. In smart-rooms and hot-spot networks, wireless
access-enabled nodes in a small area share limited channel
bandwidth. Since the area is small, the wireless hosts per-
vade through the entire network and are all within each others
transmission range. There are a large number of hosts and
hence connections. So, channeling all data through a sin-
gle intermediate node, such as a base-station, is inefficient.
Communication is pervasive, i.e., there are many source
destination pairs distributed throughout the network. The
sources must not all rely on a single entity, the base-station, to
This work was supported by the DoD Multi-disciplinary University Re-
search Initiative (MURI) program administered by the Office of Naval Re-
search under Grant NAVY CU 37515-6281, and the NSF EIA 99-72884
grant. Any opinions, findings and conclusions are those of the authors and
do not necessarily reflect the views of the above agencies.
relay their data to their respective destinations. They should
be able to directly communicate with their destinations. If a
base-station is used as an intermediary, direct one-hop trans-
missions are needlessly made two-hop. (The base-station
must only serve as an access point to the wired Internet, not as
a relay for peer-to-peer transmissions between mobile nodes
within the wireless network.) Furthermore, in military and
disaster rescue environments, a group of people carrying mo-
bile handheld devices should be able to communicate with
each other, with no time for planning and building a support
infrastructure such as a base-station. The single-hop ad hoc
wireless network, without a base-station, thus accurately rep-
resents the network used in smart-rooms, hot-spot networks,
emergency environments, and in-home networking.
IEEE 802.11 has recently become the de facto Medium
Access Control (MAC) standard in connecting mobile hosts
in an ad hoc network environment. It relies on the Distrib-
uted Co-ordination Function (DCF) to resolve channel access
contention in a distributed way. However, the IEEE 802.11
DCF does not currently have any provision to guarantee QoS,
such as minimum throughput, to flows accessing the channel.
Without any co-ordination, if the sum of transmission rates
of all the hosts (or flows) is greater than the channel capac-
ity, heavy channel contention will occur and thus QoS cannot
be guaranteed for any flow. Much research has been done
in the area of distributed weighted fair scheduling (DWFS)
[3,13,18,19,24] for IEEE 802.11 networks operating in the
DCF mode. In DWFS, each flow is assumed to have a weight
which defines its importance relative to other flows. A sched-
uler combined with the MAC-layer IEEE 802.11 protocol
then schedules the flows packets on the channel such that
200
the throughput they receive is proportional to their weights.
However, DWFS is not yet a standard part of the IEEE 802.11
MAC protocol.
In the absence of distributed MAC-layer weighted fair
scheduling in the current IEEE 802.11 standard, we propose
a scheme at the higher layers of the OSI protocol stack to co-
ordinate individual flows channel access in the single-hop
ad hoc network scenario, in order to promote co-operation
between flows and provide minimum throughput guarantee
for each of them. To this end, we first determine the flows
weights based on their relative channel access requirements.
The flow weights, in turn, determine the transmission rate
of each flow. The flows transmission rates are controlled at
the application or middleware layers, without any MAC-layer
scheduling support. Therefore, our scheme can be used over
the standard IEEE 802.11 protocol and is easily deployable
using todays off-the-shelf 802.11 products. In case DWFS
becomes available at the MAC-layer in the future, our scheme
is still required in order to provide the MAC-layer scheduler
with the flow weights, but enforcing the flow weights will be
left to the MAC-layer scheduler.
The exact share of network bandwidth allotted to a flow de-
pends on its requirements relative to the requirements of other
flows. Each flow maps its minimum and maximum bandwidth
requirements to its minimum and maximum channel time pro-
portion (CTP) requirements, respectively. We propose the use
of a centralized Bandwidth Manager (BM), which obtains
from each flow its CTP requirements, at the start of its ses-
sion. It uses this information to gauge what proportion of unit
channel time (CTP) each flow should be allotted. The CTP
allotted by the BM to each flow (i.e., its flow weight) lies
somewhere between the flows minimum and maximum re-
quirements. The term channel time proportion is defined as
the fraction of unit time for which a flow can have the channel
to itself for its transmissions. Since our network model allows
only one node to transmit on the channel at a time, there is a
direct correspondence between the channel time a flow uses
and the share of the network bandwidth it receives. The BM
may also refuse to admit a flow, i.e., allot 0% channel time.
This can happen if the flows minimum CTP requirement is
so large that the network cannot support it, without violating
some other flows minimum CTP requirement.
The problem with the admission control solution described
above, however, is that it is a one-time procedure performed
before the flow starts. It does not take into account the
changes in the wireless network over the duration of the flows
operation. Not only can the perceived channel capacity vary
over time due to varying contention [6] as flows arrive and
depart, but the channel capacity as perceived by different net-
work nodes at the same time can also be different. The latter
phenomenon is due to location-dependent fading errors and
location-dependent interference from external objects.
When a new flow arrives and demands a share of the chan-
nel, the respective CTPs allotted to already existing flows may
have to be reduced in order to accommodate it. This revo-
cation of channel time should not, however, result in these
existing flows ceasing to meet their minimum CTP require-
SHAH ET AL.
ment. Similarly, when a flow ends, its CTP must be suitably
redistributed among the still existing flows so they can hope
to achieve a better QoS.
The BM must therefore not just perform one-time admis-
sion control and teardown, but also perform dynamic band-
width management. The BM must re-negotiate with each flow
its CTP as its channel characteristics change, and as the num-
ber of active flows in the network varies. The detection of
change in channel characteristics, and adaptation of the flow
to this change, happen continuously through the course of the
session. Bandwidth re-negotiation must also occur before a
flow changes its packet transmission rate, as in the case of
bursty VBR traffic.
The rest of the paper is structured as follows. The next
section describes the overall network topology, the architec-
ture of the bandwidth management system and the bandwidth
management protocol. Section 3 presents our experimental
results. Section 4 discusses some related work in the field.
Finally, section 5 concludes the paper.
2. Bandwidth management system design and
implementation
In the previous section, we motivated the need for admission
control coupled with dynamic bandwidth management in a
single-hop ad hoc wireless network. In this section, we de-
scribe the characteristics of the network we are concerned
with, the architecture of the bandwidth management system
and the communication protocol.
2.1. Network model
We design and implement our bandwidth management scheme
for a wireless network consisting of heterogeneous computers
and devices connected together over the IEEE 802.11 MAC
layer. The network in our prototype testbed implementation
consists of handheld PCs and laptop computers with their
802.11 interfaces configured in peer-to-peer ad hoc mode. We
assume that each node in the network is within the transmis-
sion range of every other node. Hence, only one node can
transmit at a time over the channel. Since every node is within
the transmission radius of every other node, routing is single-
hop.
Unlike in [5], where a base-station determines the sched-
ule of transmission for the entire network and all communi-
cation is via the base-station, in our network, transmission is
distributed and peer-to-peer. The IEEE 802.11 MAC proto-
cols DCF, which is the one relevant to our network model,
does not have a provision for a fixed transmission schedule.
A node can send when it senses that the channel is not busy.
A binary exponential backoff mechanism resolves collisions
that might occur as a result of nodes transmitting at random
times. Moreover, any node in the network can transmit to
any other node directly without using the base-station as an
intermediary hop. Figure 1 illustrates our network model as
compared to the base-station model. The distributed, peer-to-
peer and ad hoc nature of our wireless network model makes
DYNAMIC BANDWIDTH MANAGEMENT
Figure 1. Comparison of network models: (a) base-station model, (b) single-
hop ad hoc network model.
the bandwidth management problem significantly harder to
solve than in the case of a base-station co-ordinated wireless
network where the base-station has full control of the con-
tending flows.
The wireless network has one system selected to host the
Bandwidth Manager (BM) program. In our prototype imple-
mentation, we choose one of the more resource-rich nodes in
the network, i.e., one of the laptops, as the host system for
the BM. We assume that the BM program resides on a well-
known port in a system whose IP address is well-known in
the wireless network. A service discovery mechanism such
as the ones described in [8,12] can be used to obtain the IP
address and port number of the BM service. The BM has to
register with the service discovery system upon startup. If the
BM suddenly becomes unavailable, due to a crash or due to
mobility, an election algorithm can be run to elect a new one
after a time-out.
Note that the base-station network is merely a special case
of the single-hop ad hoc network, but with no peer-to-peer
communication between mobile nodes. (All communica-
tion, as mentioned before, is between the base-station and
the mobile nodes.) Most current wireless LANs, which adopt
the base-station network model, also use IEEE 802.11 DCF.
Hence the contention characteristics are identical to those in
a single-hop ad hoc wireless network. Our solution, which
is basically designed for the single-hop ad hoc network, thus
also works for the base-station network. Uplink and downlink
traffic between a particular mobile node and the base-station
can simply be considered as two separate single-hop flows,
and their respective channel time requirements can be allotted
accordingly by the BM. The BM in the base-station network
can be situated at the base-station itself. In this paper, for
brevity, we focus only on the single-hop ad hoc peer-to-peer
network model.
We assume a network has a set of flows F . Each flow
g F is uniquely identified by its source IP address, source
port number, destination IP address and destination port num-
ber. We call this unique identifier the flow-id of the flow.
A new flow f registers with the BM before beginning its
transmission. The application initiating flow f has a mini-
mum bandwidth requirement Bmin (f ) and a maximum band-
201
Figure 2. Bandwidth management system architecture.
width requirement Bmax (f ). The flow f also has an esti-
mate of the total network bandwidth Bp (f ). At the time of
registration, it specifies its minimum and maximum CTP re-
quirements, pmin (f ) and pmax (f ), to the BM. Section 2.3 dis-
cusses how pmin (f ) and pmax (f ) are obtained from Bmin (f )
and Bmax (f ), respectively. In response, the BM adds flow f
to set F and allots it a certain channel time pa (f ), when the
flow is admitted. Flow f then uses this allotted CTP pa (f )
to calculate its transmission rate. It transmits using this trans-
mission rate until either it stops or until a new pa (f ) value is
allotted to it. A new pa (f ) could be allotted to it when there
is a change in the channel characteristics or in the network
traffic characteristics.
We assume that the flows in the wireless network are well-
behaved and co-operative, i.e., they will refrain from exceed-
ing their allotted channel share (eating into other flows share)
and will release any channel share allotted to them when they
stop. If the flows are not well-behaved and co-operative, then
a policing mechanism (see section 2.7) can be used to detect
the rogue flows and eliminate them from the system.
2.2. Bandwidth management system architecture
The architecture of the bandwidth management system con-
sists of three major components as shown in figure 2: (a) the
Rate Adaptor (RA) at the application or middleware layer,
(b) the per-node Total Bandwidth Estimator (TBE) at the
MAC-layer and (c) the Bandwidth Manager (BM), which is
unique in the entire single-hop wireless network. Our system
takes advantage of cross-layer interaction between the appli-
cation/middleware and link layers.
Rate Adaptor (RA). In our design, we assume the absence
of DWFS at the MAC layer. Hence, a flows bandwidth con-
sumption in accordance with its allotted CTP is regulated
only by the Rate Adaptor (RA). The RA converts a flows
bandwidth requirements into CTP requirements, communi-
cates this to the BM, and obtains an allotted CTP for this flow
202
from the BM. It then controls the transmission rate of each
flow depending on its allotted CTP. For the sake of simplicity,
in our UDP simulation experiments and testbed experiments,
the RA is built into the UDP application itself, to adapt its data
generation rate. Ideally, however, to avoid changing the ap-
plication, we recommend that the RA be implemented sepa-
rately as a module and be linked to the application at run-time.
It would thus function as middleware, just below the applica-
tion layer, and shape the applications traffic. Various queue-
based rate controllers are available for this purpose [16]. Our
interest is in the design of the overall bandwidth management
architecture, rather than the implementation of individual rate
control mechanisms. For our TCP simulation experiments,
we simulate queue-based rate control by having an RA per-
node at the network interface queue, rather than within the
application.
Note that, in case DWFS is present at the MAC layer, shap-
ing the traffic and enforcing flow rates can be left to it. The
RAs function is deprecated to merely communicating with
the BM and determining the flow rate.
Total Bandwidth Estimator (TBE). The per-node Total
Bandwidth Estimator is co-located with the IEEE 802.11 pro-
tocol at the MAC layer. It estimates the total network band-
width Bp (f ) for each flow f sourced at the node it resides
on.1 Bp (f ) is what flow f perceives to be the total band-
width of the network at a particular time. In other words, at
a particular instant in time, Bp (f ) is equal to the theoretical
maximum capacity of the channel (1, 2, 5.5 or 11 Mbps for
IEEE 802.11) minus the bandwidth lost due to channel errors,
caused by fading, interference and contention experienced by
flow f s packets, at that instant. The physical channel errors
and contention at a particular instant in time is estimated from
the errors and contention experienced in recent history. De-
tails of the estimation method of Bp (f ) are in section 2.4.
Note that the TBE is per-node whereas it performs total band-
width estimation per-flow sourced at the node it resides on.
The TBE continuously measures the total perceived band-
width for each flow. It periodically passes this up to the RA
of the flow at the higher layers. The RA of a flow f uses
it in the translation of flow f s bandwidth requirements to
its CTP requirements. When the total bandwidth Bp (f ) per-
ceived by flow f changes, the channel time requirements cal-
culated using Bp (f ) also change. The TBE informs the RA
of the new Bp (f ). The RA may now need to re-negotiate on
behalf of flow f with the BM, using flow f s new CTP re-
quirements that are calculated with the new Bp (f ) estimate.
Since CTP allotted to flow f is directly related to its share of
total network bandwidth, if a flow perceives the total network
bandwidth as having decreased, its share of the bandwidth
1 In a single-hop peer-to-peer wireless network, we perform bandwidth man-
agement per-flow, since each flow can have a different destination. In a
base-station environment, we can perform bandwidth management per-
node since every node only communicates with the base-station. In the
base-station scenario, each node, rather than application, specifies its band-
width requirements to its RA, and bandwidth estimation is done only for
links between mobile nodes and the base-station.
SHAH ET AL.
will also decrease. This may cause it to fall substantially
below its minimum bandwidth requirements. Hence the re-
negotiation. We do not wish to re-negotiate for small changes
in Bp (f ), however, in order to keep re-negotiation overhead
small. The RAs not reacting to small changes in Bp (f ) may
thus cause small violations of the minimum bandwidth re-
quirements. (But not minimum CTP requirements.) The mo-
ment a large violation occurs, the RA immediately reacts and
re-negotiates. The parameter that defines small and large
is tunable. It trades off the hardness of the bandwidth guaran-
tee with re-negotiation overhead.
Example. Assume a flow f in a 2 Mbps wireless network has
minimum bandwidth requirement 300 Kbps and perceives to-
tal network bandwidth of 1.5 Mbps. (That is, the flow f per-
ceives this to be the total capacity of the 2 Mbps channel.)
Assume further that the CTP allotted to it is 20%, thus en-
suring it just meets its minimum bandwidth requirement. If
the total network bandwidth, as perceived by f , decreases to
1.2 Mbps due to an increase in physical channel errors or con-
tention, then the 20% channel time is no longer sufficient for
the flow to meet its minimum bandwidth requirement. Its RA
must then re-negotiate for at least a 25% of the channel time.
Similarly, if a flow perceives the total network bandwidth to
have increased, it must release any excess share of the channel
it has been allotted, so that some other flow can use it.
Bandwidth Manager (BM). The Bandwidth Manager per-
forms admission control at the time of flow establishment
and bandwidth redistribution at the time of flow teardown.
Admission control involves revocation of some channel time
from existing flows and re-allocation of this portion to the new
flow. The BM also performs re-negotiation either when some
flow detects a change in its perceived bandwidth or when its
traffic characteristics change.2
The BM admits a flow only if it can allot at least its mini-
mum CTP requirement. Otherwise, the flow is rejected. The
remaining channel time as yet unallotted after all the admit-
ted flows minimum channel time requirements are satisfied,
is allotted on a maxmin fair basis. We therefore deem our
channel time allocation scheme at the BM maxmin fair with
minimum guarantee. Each flow receives whatever CTP is
allotted to it by the maxmin fair algorithm, in addition to
its minimum CTP request, which is automatically guaranteed
when it is admitted. A detailed description of the maxmin
fairness algorithm can be found in section 2.5 of the paper.
2.3. Bandwidth management protocol
This section describes the protocol used in the interactions be-
tween the various components of the bandwidth management
architecture and the details of the BMs operation. The BM is
2 The centralized BM does not take on the onus of channel bandwidth esti-
mation, and leaves this to the individual per-node TBEs, because the avail-
able channel capacity is different for different peer-to-peer flows, due to
location-dependent physical errors.
DYNAMIC BANDWIDTH MANAGEMENT
Figure 3. Bandwidth management protocol.
Table 1
Explanation of notation used in Bandwidth Management protocol.
Notation Meaning
F Set of flows admitted by the BM
gF All individual flows previously admitted by the BM
f New flow requesting admission
Bmin (f ) Minimum bandwidth requirement of flow f
Bmax (f ) Maximum bandwidth requirement of flow f
Bp (f ) Total network bandwidth as perceived by flow f
pmin (f ) Minimum channel time proportion required by flow f
pmax (f ) Maximum

channel time proportion required by flow f
prem 1 gF pmin (g): channel time remaining after
pmin (g), g F is met
pnewmax (f ) pmax (f ) pmin (f ): maximum channel time proportion
requirement for f that is input to maxmin algorithm
because pmin (f ) is already allotted
pmm (f ) Channel time proportion allotted to flow f by maxmin
algorithm. This is in addition to pmin (f ) which was already
allotted before maxmin algorithm began
pa (f ) Total channel time proportion allotted to flow f , i.e.,
pmin (f ) + pmm (f )
invoked at the time of flow establishment, flow teardown, sig-
nificant change in a flows perception of total bandwidth, or
significant change in a flows traffic pattern. Figure 3 shows
the actions that occur when these events happen. Table 1 is an
explanation of the notation used in the protocol description.
Flow establishment. At the time of initiating a flow f ,
the application specifies its required minimum bandwidth
Bmin (f ) and maximum bandwidth Bmax (f ), both in bits per
second, to its RA. The dRSVP [21] scheme also uses max-
imum and minimum bandwidth requirements as the specifi-
cation of utility. These values have to be each divided by
the flow f s perceived total network bandwidth Bp (f ) to ob-
tain its requested minimum and maximum CTPs, pmin (f ) and
pmax (f ), respectively. The total network bandwidth Bp (f )
perceived by a flow f is estimated by the TBE at the local
node. A best-effort flow will have Bmin (f ) = 0. Figure 4
shows the shape of the utility curve of the application.
203
Figure 4. Utility curve of users.
Note that both the CTP consumed by the flow f s data
packets in the forward direction as well as CTP consumed by
the acknowledgements in the reverse direction, if any, must
be included in f s CTP requirement. Still, it is sufficient to
do bandwidth estimation at only one of the end-points of the
link. This is because both types of packets traverse the same
wireless link, and hence face the same level of contention
and physical errors. The TBE simply quantifies the effect of
these phenomena. We perform bandwidth estimation, using
the TBE, at the source. Of course, the data and acknowl-
edgements may be of different sizes and packets of different
sizes are affected differently by the same level of physical er-
ror. Hence Bp (f ) is different for different packets of the same
flow. The TBE returns a single bandwidth estimate Bp (f ), for
the link flow f traverses, normalized to a standard packet size.
(See section 2.4.) It must be appropriately scaled for different
flow packet sizes, using the reverse of the normalization pro-
cedure, at the time of flow establishment and re-negotiation.
For VBRUDP flows, either the mean packet size can be used
or the VBR flow can be split into CBR components, as de-
scribed later in this section. For TCP flows, separate Bp (f )
values can be derived for data and acknowledgement packets
from the single normalized value returned by the TBE.
It must be kept in mind that the TBE of flow f measures
the perceived bandwidth Bp (f ) using MAC layer frames.
These MAC layer frames include protocol headers from the
intermediate layers of the protocol stack between the appli-
cation and the link layers. The Bp (f ) scaling operation must
take into account the fact that the lower layers of the proto-
col stack will add their respective headers to each packet, and
thus consume some of the channel capacity. The size of the
lower-layer headers must be added to the application packet
size in the scaling operation.
The RA of a node registers a new flow with the nodes
TBE. Initially, the TBE has no estimate of the total network
bandwidth as perceived by this newly beginning flow. This is
because it has to use the flows packets themselves for obtain-
ing an estimate of the total network bandwidth, based on the
physical channel errors and contention these packets experi-
ence. But the flow has not sent out any packets yet and is still
in the process of establishment. So, when initially computing
the flows requested minimum and maximum CTPs, the RA
has to use a hardcoded initial total bandwidth estimate.3 Once
the flow begins, a more accurate total bandwidth estimate will
be available from the TBE. The requested minimum and max-
imum CTPs can then be modified using this new, more ac-
3 In our prototype testbed implementation, we use a 2 Mbps network and we
set this hardcoded value to 1.5 Mbps.
204
curate estimate, and re-negotiation done with these modified
values.
Alternatively, in the case of a connection-oriented flow, the
first few flow-establishing packets can be used in the total
bandwidth estimation instead of the hardcoded estimate. For
example, the physical channel errors and contention faced by
TCPs three-way handshake messages can be used in the ini-
tial measurement. If the application involves some other con-
trol messages (e.g., client asking server if file exists or not),
then these can be used. A current estimate being used by other
flows between the same end-points can also be used initially.
A fourth option is to have the BM maintain a list of current
total bandwidth estimates for all flows. Then, a new flow can
query the BM for an initial estimate. The BM simply returns
the average of the list of total bandwidth estimates.
Let the initial total bandwidth estimate, how ever it
may be obtained, for a new flow f be Bp (f ). The CTP
pmin (f ), required to satisfy the new flow f s minimum band-
width requirement Bmin (f ), is pmin (f ) = Bmin (f )/Bp (f ).
pmin (f ) = 0 for best-effort flows. Similarly, the CTP
pmax (f ), required to satisfy flow f s maximum bandwidth
requirement, is pmax (f ) = Bmax (f )/Bp (f ). The RA of the
new flow f sends the BM a request message containing
the flow-id of f , pmin (f ), pmax (f ) and a timestamp for or-
dering.
The BM checks whether, for
all flows g in the set F of
previously registered flows, 1 gF pmin (g)
pmin (f ).
If this is true, the new flow f is admitted (F = F {f }), else
it is rejected and a reply message offering it zero CTP is
returned to its Rate Adaptor. Note that a best-effort flow with
pmin (f ) = 0 is always admitted. A rejected flow may attempt
again later to gain access to the channel. Flows are admitted
strictly in the order they arrive, to alleviate starvation of pre-
viously rejected real-time flows. The problem of starvation of
a best-effort flow after admission is dealt with in section 2.6.
Once the new flow f is admitted, the BM must redistribute
channel time within the new set of existing flows F . Since the
original admission test was passed by flow f , accommodat-
ing it will not cause the CTP allotted to any flow g F to
fall below its minimum CTP request. Hence, the BM initially
sets allotted CTP pa (g) = pmin
(g), g F . The remain-
ing channel time, prem = 1 gF pmin (g), is distributed
among the flows g F in maxmin fair fashion. Our chan-
nel time allocation policy is thus called maxmin fair with
minimum guarantee. The maximum CTP requirement for
each flow g F in the maxmin fair computation is set to
pnewmax (g) = pmax (g) pmin (g). This is because pmin (g)
has already been allotted to it and it only needs pnewmax (g)
more to fulfill its maximum CTP requirement. Thus, know-
ing prem and pnewmax (g) g F , the maxmin algorithm can
now proceed. Details of the maxmin fairness algorithm can
be found in section 2.5.
Suppose that out of the remaining channel time prem , the
amount allotted to any flow g F by the maxmin algorithm
is denoted by pmm (g). Now, 0  pmm (g)  pnewmax (g) and

flows using the maxmin fair algorithm with minimum guar-
gF pmm (g) = prem . Then, the total CTP allotted to each
flow g F is pa (g) = pmin (g) + pmm (g). Note that for best-
SHAH ET AL.
effort flows, since pmin (g) = 0, pa (g) = pmm (g). In other
words, channel time is allotted to best-effort flows only after
all the higher priority real-time flows are all allotted at least
their minimum share.
After the new flow f is admitted, the BM registers an entry
pertaining to it in its flow table. This entry consists of: (a) the
new flow f s flow-id, (b) the socket descriptor of the socket
used by the BM for communication with f s RA, (c) pmin (f ),
(d) pmax (f ) and (e) pa (f ). The socket descriptor is stored in
the table so that if any re-negotiation needs to be done later
with flow f s RA (for example, when newer flows arrive in
future or existing flows depart), this socket can be used. In
addition, a timestamp indicating the freshness of the latest
request message is also maintained for each flow. This
timestamp is used for two purposes: (a) timing out stale reser-
vations, and (b) proper ordering of multiple outstanding re-
negotiation requests from the same flow. Since reservations
can time-out, the entries in the flow table are soft-state en-
tries. If, for some reason, a flows reservation has timed-out
but the flow is still transmitting, this can be detected using a
policing mechanism. (See section 2.7.)
Finally, for every flow g F , the allotted CTP pa (g) is
then sent to flow gs RA using a reply message. (Note that
the name of the message is a misnomer in the case of all flows
g F except the new flow f because, in their case, the re-
ply is gratuitous, not a response to any message they sent.)
It may be the case that all flows g F do not need to be sent
a reply message. No reply message needs to be sent to a
flow in F whose allotted CTP has not changed due to the ar-
rival of the new flow f . Although we implement the reply
message as multiple unicast messages to individual RAs for
reliability, it can also be implemented for efficiency as a sub-
net broadcast message, containing flow-id and pa (g), g F .
A flow f is rejected using a unicast reply with pa (f ) = 0.
Other existing flows allotted CTPs are not affected.
The RA of every flow that receives a reply message, gra-
tuitous or otherwise, from the BM sets its transmission rate
respectively to pa (g) Bp (g) bits per second (bps). The new
flow f can now begin operation whereas the older flows sim-
ply resume operation with their respective new rates.
The timestamp in the reply to flow g indicates the last
request received from g by the BM. The value of Bp (g)
used to compute pmin (g) and pmax (g) for this request must
then be used in the transmission rate formula above, since it
is based on this value of Bp (g) that pa (g) was calculated by
the BM. As a new Bp (g) is returned by the TBE periodically,
a new rate is also used periodically. If the Bp (g) change is
large since the last period, re-negotiation must occur, as ex-
plained below.
Flow teardown. When a flow f terminates, its RA sends
a teardown message to the BM. The BM removes flow f
from the set of existing flows F i.e., F = F {f }. It then
redistributes flow f s allotted CTP pa (f ) among the other
antees. The RA of each flow g F (the new set F ) is
told of its newly allotted CTP by the BM. The socket de-
DYNAMIC BANDWIDTH MANAGEMENT
scriptors in the flow table are used to send gratuitous re-
ply messages for this purpose. The entry for the terminating
flow f in the BMs flow table is expunged. A teardown-
acknowledgement message is sent to f s RA.
Change in a flows perception of total network bandwidth.
The RA of every flow periodically obtains from the TBE the
flows current perceived total bandwidth. The TBE updates
the RA with the mean of the perceived total network band-
width measured for each packet successfully transmitted by
the flow in recent history. The inter-update period could be
in terms of number of packets transmitted or in terms of time.
We recommend using a hybrid scheme for determining update
period: it should be based on time when the transmission rate
of the flow is low and based on number of packets transmitted
when it is high. In our experiments, we use high transmission
rates in order to determine the performance of our scheme un-
der high network loads. Therefore, we use a perceived band-
width update interval based on number of packets. We use
a default interval of 100 transmitted packets in our experi-
ments, but we also measure how various other intervals affect
the performance of the system.
In case a newly obtained perceived bandwidth value
NEWBp (f ) differs significantly from Bp (f ), the RA must
re-negotiate its flows CTP with the BM, as indicated in the
example in the previous section. It must also set the value
of perceived bandwidth Bp (f ) to the newly obtained value
NEWBp (f ). Note that the RA only sets Bp (f ) to NEWBp (f )
and re-negotiates with the BM using this new value when
there is a significant change, not with every update. A new
rate using the previously allotted CTP is, however, calculated
with every update. In our experiments, we assume a deviation
= 15% of NEWBp (f ) from Bp (f ) as significant enough to
warrant re-negotiation. We also measure how other perceived
bandwidth deviation tolerance () percentages affect system
performance.
If re-negotiation has to be done, the RA of flow f sends
a request message to the BM with flow-id, pmin (f ) and
pmax (f ). The values of pmin (f ) and pmax (f ) sent in the
request message are re-calculated using the new value of
Bp (f ). The rest of the re-negotiation procedure is almost
identical to the one used for flow establishment, both at the
BM as well as at the RA. (See figure 3.) The only difference
is that the BM does not have to add a new entry in its flow
table for f ; it only updates the already existing one.
Note that a flow f s re-negotiation request can be rejected
by the BM, i.e., it can receive pa (f ) = 0, in response to
the requested CTP. This means that the flow has been cut-
off in mid-operation. Unfortunately, the nature of the wire-
less network is inherently unreliable and as network resources
decrease, some flows will necessarily have to be cut-off in
mid-operation so that others can be supported. Our scheme
guarantees that each flow will obtain at least its minimum re-
quested CTP for almost 100% of its active duration. If the
system cannot guarantee the flow at least this level of QoS, it
will drop it altogether. In other words, a flow will either re-
ceive (for nearly 100% of its active duration) at least its min-
205
imum requested CTP pmin (f ), or it will receive no channel
time at all. The guarantee in terms of bandwidth is that the
allotted bandwidth never falls more than a factor of below
the minimum requested bandwidth Bmin (f ), since if Bp (f )
changes by a factor of , re-negotiation occurs.
Currently, we do not use any priority scheme to cut-off
particular flows. If perceived bandwidth decreases for all
flows, the first flow initiating re-negotiation is cut-off. Al-
ternate strategies to pick flows to cut-off in mid-operation are
discussed briefly in section 3.1.1.
Change in a flows traffic characteristics. When a VBR
UDP flow f (e.g., MPEG video stream) needs to send a burst
of traffic at a rate different from its normal rate, it must inform
its RA. The RA will re-negotiate for a larger CTP for flow f
depending on the bandwidth of the burst. The re-negotiation
procedure is the same as in the case of change in perceived
bandwidth. At the end of the burst duration, the RA will
again re-negotiate to release the excess CTP. This solution is
equivalent to splitting up a VBR stream in the time domain
into multiple CBR streams. There exists previous literature
in the context of ATM networks [11] in which VBR streams
are split into multiple CBR streams in the time domain. Since
this scheme only involves re-organizing the traffic rather than
the network, it can be directly applied from ATM networks to
wireless networks.
Figure 5 is an MPEG-4 trace of an hour-long, 25 frames
per second, medium-quality, clip of the movie Silence of the
Lambs. The trace was taken from [10] and the references
therein. On the x-axis is a running count of the frame number.
On the y-axis is the frame size averaged over non-overlapping
blocks of 50 frames. One possible way to split up this VBR
flow into multiple CBR components is shown in figure 5 as
the contour of the plot. The CBR bandwidth component thus
obtained is then used as the minimum bandwidth requirement
Bmin (f ) in negotiating with the BM.
Frequent bursts could result in an explosion in re-negotia-
tion overhead. We deal with the problem of frequent bursts
in one of two ways: (a) setting Bmin (f ), at the time of burst-
induced re-negotiation, large enough to engulf multiple bursts
and (b) having large buffering at the receiver to deal with the
burst.
Figure 5. MPEG-4 trace of Silence of the Lambs clip with corresponding
CBR components.
206
Figure 6. IEEE 802.11 unicast packet transmission sequence.
2.4. Total bandwidth estimation procedure
To determine pmin (f ) and pmax (f ), the RA of a flow f needs
to have an estimate of the total bandwidth over the wireless
link being used by the flow. To this end, we introduce a band-
width measurement mechanism based on IEEE 802.11 DCF
MAC layer, and demonstrate its robustness.
IEEE 802.11 relies on the DCF method to coordinate the
transmission of packets based on CSMA/CA without any cen-
tral control unit. The packet transmission sequence is illus-
trated in figure 6. Before transmitting a packet, a node senses
the channel to make sure that the channel is idle; otherwise it
backs off by a random interval and senses the channel again.
If the channel is idle, it transmits a RTS (Request-to-Send)
packet to signal its intention to send a packet.4 On receiv-
ing the RTS packet, the destination node replies with a CTS
(Clear-to-Send) packet to give the sender a go-ahead signal,
and to silence the destination nodes neighboring nodes. After
receiving the CTS packet, the sender sends the DATA packet,
and it is then acknowledged by an ACK packet from the re-
ceiver.
Similar to [14], we measure the throughput of transmitting
a packet as TP = S/(tr ts ), where S is the size of the packet,
ts is the time-stamp that the packet is ready at the MAC layer,
and tr is the time-stamp that an ACK has been received. Note
that the time interval tr ts includes the channel busy and
contention time. We keep separate throughput estimates to
different neighboring nodes because the channel conditions
may be very different. We only keep an estimate for active
links, since we do not have any packets to measure tr ts over
inactive ones.
This MAC layer measurement mechanism captures the ef-
fect of contention on a flows perceived channel bandwidth.
If contention is high, tr ts will increase and the throughput
TP will decrease. This mechanism also captures the effect of
physical errors because if the RTS or DATA packets are af-
fected by channel errors, they have to be re-transmitted, upto
the re-transmission limit. This increases tr ts and corre-
spondingly decreases the flows perceived bandwidth. Since
our MAC layer measurement of perceived bandwidth takes
into account the effects of both contention and physical errors
due to fading and interference on a flow, we can have the flow
react suitably to these factors by monitoring the change in per-
ceived bandwidth. It should be noted that the perceived band-
width is measured only using successful MAC layer transmis-
sions.
4 For very small packets, the sender may skip the RTS packet and directly
send out the DATA packet.
SHAH ET AL.
Figure 7. Raw throughput and normalized throughput at MAC layer.
It is clear that the measured throughput of a packet depends
on the size of the packet. Larger packet has higher measured
throughput because it sends more data once it grabs the chan-
nel. To make the throughput measurement independent of
the packet size, we normalize the throughput of a packet to
a pre-defined packet size. Before being used by a flow of a
particular packet size, it must be scaled to that packet size. In
figure 6, Td = S/BW ch is the actual time for the channel to
transmit the data packet, where BW ch is the channels bit-rate.
Here we assume channels bit-rate is a pre-defined value. The
transmission times of two packets should differ only in their
times to transmit the DATA packets. Therefore, we have:
S1 S2
(tr1 ts1 ) = (tr2 ts2 ) (1)
BW ch BW ch
S2 S2
= , (2)
TP2 BW ch
where S1 is the actual data packet size, and S2 is a pre-defined
standard packet size. By equation (2), we can calculate the
normalized throughput TP2 for the standard size packet. To
verify the validity of this equation, we simulated a group
of mobile nodes within a single-hop ad hoc network using
the ns-2 network simulator [23]. We sent CBR traffic from
one node to another, and varied the packet size from small
(64 bytes) to large (640 bytes) during the course of the simu-
lation. The measured raw throughput is normalized against a
standard size (picked as 512 bytes). Figure 7 shows the result
of the measured raw throughput and its corresponding nor-
malized throughput. Obviously, the raw throughput depends
on the packet size; larger packet size leads to higher measured
throughput. The normalized throughput, on the other hand,
does not depend on the data packet size. Hence, we use the
normalized throughput to represent the bandwidth of a wire-
less link, to filter out the noise introduced by the measured
raw throughput from packets of different sizes.
Another important issue is the robustness of the MAC
layer bandwidth measurement. We measure the bandwidth of
a link in discrete time intervals by averaging the throughputs
of the recent packets in the past time window, and use it to es-
timate the bandwidth in the current time window. Obviously,
this estimation may not be accurate because the channel con-
dition may have changed. To evaluate the estimation error,
we run a CBR flow over UDP with data rate 160 Kbps from
a node to another in a 10 node one-hop environment. Back-
ground traffic consists of 1 greedy TCP flow in the light chan-
nel contention case, and 7 TCP flows in the heavy contention
DYNAMIC BANDWIDTH MANAGEMENT
case. Here we use TCP only to generate bursty cross-traffic
to the UDP flow. We measure and normalize the throughput
of the CBR flow every 2 seconds using the average of packet
throughputs in the past time window. Our results show that
under light channel contention, over 97% of the estimates are
within 20% of error; under heavy contention, still over 80% of
the estimates are within 20% of error. We thus conclude that
using average throughput of past packets to estimate current
bandwidth is feasible and robust.
It should be noted that the bandwidth estimation mecha-
nism in no way alters the IEEE 802.11 protocol. Our band-
width estimation mechanism, with the normalization exten-
sion, was satisfactorily accurate for the scenarios in our
simulation and testbed experiments. However, the theory
behind the normalization may not be applicable for arbitrar-
ily large packet sizes or arbitrarily high bit-error rates. In
such cases, the TBE could keep an indexed table of sepa-
rate estimates for different packet size ranges per active link,
rather than maintaining a single normalized estimate per ac-
tive link and scaling it to various packet sizes at the time
of flow establishment/re-negotiation. If the indexed table
method is used, the source and destination must both perform
total bandwidth estimation, for data and acknowledgements,
respectively. The destination must periodically communicate
its bandwidth estimate for acknowledgement packets with the
source using an in-band signaling mechanism. (The signal-
ing itself consumes negligible bandwidth.) In the single nor-
malized estimate method, the source alone does the estima-
tion and appropriately scales the normalized estimate for both
data and acknowledgement packet sizes. Thus, although the
indexed table estimation method improves accuracy of the es-
timate in certain special cases, it also incurs a small storage
space and in-band signaling overhead.
2.5. Maxmin fairness
Fairness is an important issue in designing our Bandwidth
Manager. In this paper, we adopt a maxmin fairness algo-
rithm with minimum guarantee in allotting channel time to
the flows. This section describes the maxmin algorithm to
calculate how much channel time each flow gets beyond its
guaranteed minimum requested channel time, after the flow
is admitted.
In maxmin fairness [4], flows with small channel time
requests are granted their requests first; the remaining chan-
nel capacity is then evenly divided among the more demand-
ing flows. As described in section 2.3, pa (f ) is first set to
pmin (f ) for all the flows. The channel time that remains,
prem , after satisfying the flows minimum requirements, is
allotted to the flows in maxmin fashion. The new maxi-
mum requirement for each flow in the maxmin algorithm
is pnewmax (f ) = pmax (f ) pmin (f ), because pmin (f ) has
already been allotted to it and must be subtracted from the
original maximum requirement. We denote the channel time
allotted to flow f by the maxmin algorithm as pmm (f ). This
is in addition to pmin (f ) allotted before the maxmin algo-
rithm is even invoked.
207
Input. Channel time: p_rem; set of requests: p_newmax[f ]
Output. Set of allocations: p_mm[f ]
proc Maxmin(p_rem, p_newmax[f ])
R := {}; //set of satisfied flows
N := size_of (p_newmax[f ]);
p_mm[f ] := 0;
while (true) do
total_satisfied = 0;
foreach f R do
total_satisfied+ = p_mm[f ];
od
CA := (p_rem total_satisfied)/(N size_of (R));
stop := true;
foreach f / R do
if (p_newmax[f ] < CA) then
R := R + {f };
p_mm[f ] := p_newmax[f ];
stop := false;
fi
od
if (stop) then
foreach f / R do
p_mm[f ] := CA;
od
break;
fi
od
Figure 8. Maxmin fair resource allocation algorithm.
The computation of the maxmin allocation is as follows.
Initially, the set of flows f , whose new maximum channel
time requirement pnewmax (f ) has been satisfied, is empty:
R = . Then, we compute the first-level allotment as
CA0 = prem /N, where N is the total number of flows. Now
we include all flows f with pnewmax (f ) < CA0 in set R, and
allot each of them p
mm (f ) = pnewmax (f ). Next, we com-
pute CA1 = (prem f R pnewmax (f ))/(N R). If for
all flows g / R, pnewmax (g)
CA1 , then we allot each of
them pmm (g) = CA1 and stop. Otherwise, we include those
flows g with pnewmax (g) < CA1 in set R, allot each of them
pmm (g) = pnewmax (g), and re-compute the next level CA2 .
When the algorithm terminates, the allocation pmm (f ) for
all the flows is maxmin fair. The pseudo-code for the al-
gorithm is shown in figure 8. It is clear that the computational
complexity of this algorithm is O(N 2 ). As mentioned earlier,
after every flow f s pmm (f ) has been determined using the
maxmin algorithm, the BM sets pa (f ) = pmin (f )+pmm (f )
and returns this value to flow f s RA.
2.6. Alternate channel time allocation strategies
Although we use the maxmin fairness with minimum guar-
antee policy for bandwidth allocation in our implementation,
a different fairness policy or even a biased, priority-based
scheme could also be used.
In our policy, as mentioned earlier, best-effort flows are
only given access to the channel after all the real-time
flows minimum requirements are satisfied. This could lead
208
to

starvation of the best-effort flows, in the rare case that
gF pmin (g) 100%. One way to eliminate this prob-
lem would be to partition channel time into a large minimum-
guarantee portion and a small maxmin fair portion, similar
to the bandwidth partitioning in [1]. The minimum require-
ments of the real-time flows, i.e., all pmin (g) > 0, are al-
lotted only from the minimum-guarantee portion. The max
min fair portion, along with any left over minimum-guarantee
portion, is used to allot the flows extra CTP pmm (g), using
just a maxmin scheme. Both real-time as well as best-effort
flows, i.e., all flows with pnewmax (g) > 0, can vie for this
portion. The presence of a separate maxmin fair portion en-
sures that, however large the minimum requirements of the
real-time flows, some channel time is always available for
best-effort flows to vie for, so they are never starved. The dis-
advantage of having a separate maxmin fair portion is that
the channel time available to satisfy minimum guarantees of
real-time flows (the minimum-guarantee portion) is reduced,
which could lead to more real-time flows being dropped.
Another alternate scheme involves pricing of channel time
and enforcing priorities based on flow budgets. The max
min fair policy with minimum guarantee lends itself to an ele-
gant two-tier pricing scheme. The guaranteed minimum CTP
pmin (g) is valued at a substantial price, whereas any chan-
nel time pmm (g) in excess of this is relatively very cheap.
Under this two-tier pricing scheme, users would be inclined
to request as little minimum guaranteed bandwidth as possi-
ble, in order to save cost. High minimum requirements are
thus punished while high maximum requirements carry no
penalty. The BM adjusts the price so as to trade-off blocking
probability of the flows with its revenue. If the price is too
high, too few flows can afford it and hence blocking probabil-
ity is high. If the price is low, blocking probability is low, but
revenue may suffer. Pricing for wireless networks has been
studied previously [17,20,22,26], but our two-tier approach is
especially suitable for our bandwidth allocation policy.
2.7. Policing
In our bandwidth management scheme, policing refers to the
task of monitoring the users, to make sure that they conform
to their allocated bandwidth. The bandwidth manager oper-
ates in two modes: normal and policing. When operating in
policing mode, the bandwidth manager listens promiscuously
to the network traffic, and checks whether a flow, identified
by the source and destination addresses and port numbers in
its packet headers, is sending out packets faster than its al-
lotted rate. Additionally, it can also catch those flows who
have not registered with the bandwidth manager. This can be
some type of denial of service attack by a malicious users,
or caused by some unmanaged applications.
Operating in policing mode is expensive. Therefore, the
bandwidth manager should operate in this mode only when
necessary. To this end, the bandwidth manager relies on the
sudden, sharp decrease of channel bandwidth as an indica-
tion, in the re-negotiation process. If there is a sudden flock of
re-negotiation requests due to reduction in Bp (g), it is likely
SHAH ET AL.
that abnormally high channel contention has occurred. Subse-
quently, the bandwidth manager switches into policing mode
to monitor the activity of the network. It may be that the chan-
nel contention is due to a sudden increase in physical errors
or it may be that it is due to a malicious or unmanaged flow.
The policing scheme can identify which of the above is the
cause. It could also happen that the unreliable subnet broad-
cast reply message did not reach a particular RA, so a flow
is continuing to transmit packets faster than its re-allotted rate.
3. Experimental results
We evaluate the performance of our Admission Control and
Dynamic Bandwidth Management system using both a proto-
type testbed as well as simulations using the ns-2 simulator.
We used our testbed when evaluating the performance of a
flow in the presence of both physical channel errors caused by
fading and interference effects as well as medium contention
from two other active stations, because there is no way to set
up physical obstacles such as walls, ceilings and doors that
cause signal weakening in ns-2. We used ns-2 simulations to
evaluate the performance of the system when there is heavy
medium contention due to the presence of a large number of
active stations.
3.1. Simulation experiments
For experiments with large numbers of nodes (
5 nodes)
and flows, we used the ns-2 simulator. We compared
the performance of an Admission Control and Bandwidth
Management-enhanced IEEE 802.11 network (henceforth
called enhanced IEEE 802.11 scheme) with an IEEE 802.11
network without bandwidth management (henceforth called
base IEEE 802.11 scheme). We used a 170 m 170 m net-
work area and the transmission range of each node was 250 m.
Hence, the entire network area falls within every nodes trans-
mission range. The maximum theoretical channel capacity
was 2 Mbps. We used the random waypoint mobility model
with moderate node speeds in our simulations.
3.1.1. UDP throughput performance
Our first simulation scenario consisted of a 20-node network
with 10 flows. Each flow had a minimum bandwidth require-
ment of 100 Kbps and a maximum bandwidth requirement
of 200 Kbps, which are typical of an audio streaming appli-
cation. All the 10 flows used 512 byte packets. The sim-
ulation ran for 600 seconds. The transmission rate used by
our scheme at any instant was determined using the method
described in section 2.3. The transmission rate used in the
base IEEE 802.11 scheme was a constant set to the maximum
requested rate of the CBR flow, as would be the case in an
unmanaged application. The RAs inter-update interval was
100 packets and its perceived bandwidth variation-tolerance
threshold = 15%, by default.
Figure 9(a) is a plot of number of packets successfully
transmitted over every 1 second interval for each of the 10
DYNAMIC BANDWIDTH MANAGEMENT
(a) Base IEEE 802.11.
(b) Enhanced IEEE 802.11.
Figure 9. Comparative throughput performance of base and enhanced IEEE
802.11 for 10-flow scenario.
flows using the base IEEE 802.11 scheme. Figure 9(b) is the
same plot using the enhanced IEEE 802.11 scheme. Note that
in our scheme two flows needed to be cut-off in mid-operation
so that other flows minimum CTP requirements are not vio-
lated. One of these is cut-off at time 149 seconds and the other
at time 264 seconds. These times indicate the respective first
occasions when the flows in question requested a minimum
CTP that could not be supported. When a new flow is admit-
ted, contention increases for all the existing flows. In general,
the flow that notices an unacceptably poor channel qual-
ity and complains first is dropped. Alternate flow dropping
strategies can also be employed, such as dropping the flow
last admitted. Pricing could also pay a role here: the flow
paying the least can be dropped.
It is clearly evident from the plots that our protocol dra-
matically improves throughput fairness. In the base IEEE
802.11 scheme, flows often fall far below their minimum
bandwidth requirement over the 1 second measurement in-
terval, resulting in a chaotic plot. Using our scheme, flows
almost never fall below their minimum bandwidth require-
ment shown with the horizontal line at 24 packets per second.
(100 Kbps/4096 bit packets is approximately 24 packets per
second.) Even when they do, it is only by a small amount.
Our scheme thus ensures that the minimum bandwidth re-
quirements of the flows are met with a far higher probability
than the base IEEE 802.11 scheme. Figure 10 is a 100-second
snapshot from the combined plot of figures 9(a) and 9(b) that
shows the comparative behavior of a single flow (flow 1).
209
Figure 10. Comparative behavior of a single flow over base 802.11 versus
enhanced 802.11.
(a) Without smoothing.
(b) With smoothing.
Figure 11. Perceived bandwidth and re-negotiations corresponding to its vari-
ation.
Figure 11(a) shows the variation of perceived bandwidth
for one of the flows in the above experiment as measured by
its TBE at the MAC layer. The superimposed stepwise curve
shows the bandwidth last used for re-negotiation in the above
experiment. Recall that = 15%. We also experimented with
smoothed perceived bandwidth estimates, which reduced the
overhead of re-negotiation frequency. Figure 11(b) is a plot
of a running average of the measured perceived bandwidth
with exponential decay, which is used for smoothing of the
estimate. The smoothed estimate falls as contention increases
and rises when the two flows are dropped and contention de-
creases. Other methods to reduce re-negotiation overhead are
described in the next subsection.
In section 1, we mentioned that improving fairness is es-
sential for providing minimum throughput guarantees to wire-
less multimedia applications. The key factor enabling our
210 SHAH ET AL.

scheme to provide minimum bandwidth requirement guar-

antees with a high probability, is its improved fairness. No
flow takes up excess bandwidth during a particular interval
thereby depriving another flow of bandwidth and resulting in
a large throughput discrepancy (i.e., poor fairness) between
the flows. Our scheme also reduces jitter in throughput as
compared to base IEEE 802.11. Throughput jitter is the dif-
ference in throughput observed over two consecutive same-
sized time intervals. It should be as low as possible for a CBR
flow. We use 1 second time intervals. We thus designate fair-
ness and throughput jitter as the key performance measures (a) Base IEEE 802.11.
that characterize the performance of our system. The better
these measures, the higher the probability of the flows meet-
ing their minimum bandwidth requirements.
While our scheme focuses on ensuring that flows receive
their minimum throughput, the delay and delay jitter are
also improved as a by-product of our bandwidth management
scheme. Since we co-operatively control the sending rate of
the flows, we observe a negligible packet loss rate when using
our scheme. Due to the rate control, queue length is uniformly
short, queuing delay is small, and congestion loss is avoided.
Since contention is uniformly low, delay jitter is also im-
proved. With base IEEE 802.11, however, since the transmis- (b) Enhanced IEEE 802.11.
sion rate is set to the maximum, a 33% packet loss rate results
due to congestion and the resultant queue overflow. When Figure 12. Comparative throughput performance of base and enhanced IEEE
802.11 for 3-flow scenario with identical bandwidth requirements.
using our scheme without perceived bandwidth smoothing,
each flow re-negotiates its allotted CTP once every 14 sec- tion in which all flows are active together. Assume that the
onds on average. In section 3.2.2, we determine that each of number of seconds for which the measurement is done is n.
these re-negotiations can take upto 60 ms in the presence of
f
We define a fairness metric FM = f A |Ni Ni |/A.
contention. This does not affect the flow too much because it We also define a throughput jitter metric for a flow f , JM f =
continues sending at a rate dictated by the previously allotted
n1 f f
CTP and current value of Bp (f ) during this interval. It does i=1 |Ni Ni+1 |/(n 1). The
overall jitter metric JM is
the mean of the JM f s, i.e., JM = f A JM f /A.
however represent a small amount of network traffic over-
For the experiments in this subsection, we use a differ-
head. The mean throughput of an active flow for our scheme
ent network scenario in which there are 6 nodes in the ns-
in the above scenario is 8% lower than that of an active flow in
2-simulated wireless network and 3 flows. The flows each re-
base IEEE 802.11. We believe that this lower mean through-
quire a minimum throughput of 200 Kbps (approximately 48
put is a small price to pay for the vastly improved stability packets/sec.) and a maximum throughput of 600 Kbps. We
in throughput. The latter property is essential for multimedia ran this simulation scenario for a duration of 300 seconds. All
applications. In the next subsection, we will discuss the rea- other simulation parameters exactly remain the same from the
sons for throughput deterioration and present mechanisms to previous subsection. We used the period when all three flows
reduce the flow-initiated re-negotiation overhead. are active for all our measurements.
Figures 12(a) and 12(b) show the number of packets trans-
3.1.2. Overhead for UDP experiments mitted over every 1 second for base IEEE 802.11 and en-
There exists a trade-off between network traffic overhead and hanced 802.11, respectively. Once again, it is evident from
performance in terms of fairness and jitter. We need to be able the plots that our scheme performs better in terms of both fair-
to quantify the fairness and throughput jitter so that we can ness and throughput jitter. However, we apply our metric to
measure how much they are affected when we try to reduce determine exactly how much our scheme improves these per-
overhead. formance measures. We obtained a value of FM = 6.72 pack-
In our simulations, we measure the number of packets of ets for base IEEE 802.11 versus FM = 4.06 packets for our
each flow transmitted over each 1 second interval in the 600 scheme. (Lower FM means better fairness.) We also obtained
second run. Let us denote the number of packets transmit- JM = 8.80 packets for base IEEE 802.11 vs. a JM = 4.93
f
ted by flow f over second i as Ni . Let the average over all packets for our scheme. (Lower JM means lower through-
flows of number of packets transmitted in second i be denoted put jitter.) We conclude that for this particular scenario, our
as Ni . Let the set of active flows, i.e., flows that have been es- scheme results in a 6080% improvement in performance.
tablished but not yet torn down or cut-off, during second i Each flow in our scheme requests a re-negotiation of CTP
be A. We only measure throughput per second for the dura- once every 7 seconds, without perceived bandwidth smooth-
DYNAMIC BANDWIDTH MANAGEMENT
Table 2
Effect of Bp (f ) inter-update period on performance and overhead.
Inter-update period FM JM Overhead
(pkts.) (pkts.) (pkts.) (requests/flow/sec.)
50 3.62 4.37 0.5
100 4.06 4.66 0.143
150 4.15 4.93 0.059
200 4.18 5.10 0.019
ing. This is lower than the 14 seconds for the scenario in the
previous section because the transmission rate is higher and
hence the 100-packet inter-update interval is reached faster.
As in the case of the scenario in the previous subsection,
there is a 28% packet drop rate in the case of base IEEE
802.11, but negligible drop rate using our scheme. Also as
in the previous scenario, the mean throughput of base IEEE
802.11 is 15% higher during the period under measurement
(all 3 flows are active) than our scheme. This is because of
three reasons: (a) the flows are pumping data into the net-
work as fast as possible in order to get as much throughput as
they can in the base IEEE 802.11 scheme while we are using
rate control, (b) our TBE is configured to return a conserva-
tive estimate for Bp (f ), and (c) in the Dynamic Bandwidth
Management scheme, the re-negotiation messages between
the various RAs and the BM consume some network band-
width.
The conservative Bp (f ) estimate was used to minimize
packet drop rate. The cost of using such a conservative esti-
mate is that our enhanced IEEE 802.11 scheme under-utilizes
the network. Mean throughput is less than it would be under
full network utilization. However, the TBEs estimate can be
suitably tuned so that throughput of our scheme approaches
that of the base IEEE 802.11 scheme and network utilization
increases. On the other hand, this will also increase the packet
drop rate of our scheme and thereby degrade performance as
packets are dropped randomly from flows. So, there exists a
trade-off between throughput and packet drop rate.
In addition to the perceived bandwidth smoothing de-
scribed in the previous section, we now discuss two other
methods to minimize re-negotiation overhead and hence the
network bandwidth re-negotiation consumes. One method is
to increase the inter-update period between successive per-
ceived bandwidth updates from the TBE to the RA. Recall
that we use 100 packets as the default inter-update interval
in our experiments. Table 2 shows how overhead and perfor-
mance vary with different inter-update intervals. As the inter-
update interval increases, some changes in perceived band-
width go undetected and cannot be responded to. Hence, the
fairness and throughput jitter worsen while the overhead im-
proves. The overhead is measured as the frequency of re-
negotiation requests per flow. The threshold tolerance to per-
ceived bandwidth changes was set at the default of = 15%
for this experiment.
The other method to reduce re-negotiation overhead is
to increase the tolerance to changes in perceived bandwidth
Bp (f ). Recall that we define significant change as a = 15%
change in perceived bandwidth. If we define significant
211
Table 3
Effect of various Bp (f ) variation tolerance levels on performance and over-
head.
Tolerance level FM JM Overhead
(%) (pkts.) (pkts.) (requests/flow/sec.)
10 3.22 4.36 0.333
15 4.06 4.66 0.143
20 4.89 5.19 0.056
25 5.77 5.37 0.026
change as, say, a = 25% change, then we can reduce
re-negotiation overhead because the RA now waits longer
and tolerates more Bp (f ) fluctuation before initiating re-
negotiation. Again, this worsens the performance of the sys-
tem because fidelity to bandwidth variations is reduced. Ta-
ble 3 shows how overhead and performance vary with differ-
ent levels of tolerance to Bp (f ) variation. The inter-update
interval was set to 100 packets for this experiment. Tables 2
and 3 both show that for a small price in terms of perfor-
mance, we can obtain large gains in overhead reduction.
3.1.3. Additional UDP performance results
In this section, we present results for two additional scenar-
ios: (a) when the flows have different minimum bandwidth
requirements and (b) when the arrival time of the flows is
staggered. We use the 6-node, 3-flow scenario used in the
previous section, with the default perceived bandwidth toler-
ance of = 15% and the default inter-update interval of 100
packets.
Figure 13 shows the comparative base IEEE 802.11 and
enhanced IEEE 802.11 throughput performance when the 3
flows each have different minimum bandwidth requirements.
The minimum requirements of the 3 flows are 100 Kbps,
200 Kbps and 400 Kbps, respectively. The maximum band-
width requirement, 600 Kbps, is the same for all 3 flows.
The plots show that while no guarantee can be made with
base IEEE 802.11, we can make coarse guarantees with our
scheme.
While in all our previous scenarios, all participating flows
started at around the same time, figure 14 shows the through-
put performance of the enhanced IEEE 802.11 scheme when
the start times are staggered. All simulation parameters are
identical to those in section 3.1.2, except the staggered start
times and the length of the simulation run, which is set to
200 seconds. The bandwidth requirements are identical for
all 3 flows, as in section 3.1.2. This plot is similar to figure 1
from [5] and figure 11 from [2], which were for a base-station
network with centralized scheduling. We have produced a
similar effect for a single-hop ad hoc network that uses the
IEEE 802.11 protocols DCF.
3.1.4. TCP experiments
So far our simulation experiments have focused on multime-
dia applications and UDP flows. In this section we investi-
gate the behavior of TCP flows and their interactions with the
BM scheme. To this end, we simulate three TCP flows, each
running between different nodes, in a single-hop ad hoc net-
work managed by a BM, i.e., using enhanced IEEE 802.11.
212
(a) Base IEEE 802.11.
(b) Enhanced IEEE 802.11.
Figure 13. Comparative throughput performance of base and enhanced IEEE
802.11 for 3-flow scenario with different minimum bandwidth requirements.
Figure 14. Enhanced IEEE 802.11 performance for 3-flow scenario with
staggered start times.
TCP traffic is best-effort and elastic, so pmin (f ) is set to zero
and pmax (f ) to 100%. As mentioned in section 2.3, different
Bp (f ) values derived from the same normalized bandwidth
estimate are used for data and acknowledgements, due to their
different packet sizes, when obtaining their respective CTP
requirements. The size of the network interface queue is 50
packets, and the maximum congestion window size for a TCP
flow is 128 packets. The experiment lasts 200 seconds. While
for the UDP experiments, rate-control using the RA is done
in the UDP application, in the TCP experiments, queue-based
rate control is done per-node at the network interface queue.
The interface queue only releases packets at the rate allotted
by the BM.
SHAH ET AL.
(a) Base IEEE 802.11.
(b) Enhanced IEEE 802.11.
Figure 15. TCP congestion window behavior when interface queue size is
smaller than congestion window limit.
Figure 15(b) shows the congestion window sizes of the
three TCP flows, in the enhanced IEEE 802.11 case. They
each expose the same behavior: the window size increases
each time to 50 packets, cuts back and the cycle repeats. This
behavior is due to TCPs additive-increase multiplicative-
decrease (AIMD) congestion control algorithm, where the
congestion window size will decrease only when a packet loss
event is encountered. Packet loss occurs only when the queue
overflows, because of co-ordinated channel access ensured by
the RA. Queue overflow occurs only when congestion win-
dow exceeds the maximum queue size. A TCP flow will keep
increasing its congestion window size up to the queuing limit.
In fact, this probing of congestion window size is TCPs
way of aligning itself to the available bandwidth of the net-
work. Without knowing the BMs allocated rate for this node,
a TCP flow has to fill the router queue before it cuts back
its congestion window size, which incurs unnecessary long
queuing delay for the packets. However, this behavior does
not forfeit its allocated bandwidth, as TCP always keeps the
queue non-empty.
As comparison, we run the same TCP experiments over
a single-hop ad hoc network without the bandwidth manage-
ment, i.e., using base IEEE 802.11. Figure 15(a) shows that
the congestion window sizes of the three flows follow the
same saw-tooth pattern as in figure 15(b). But the maxi-
mum window size that each flow can reach may not be ex-
actly the same, because of the unfairness in the channel ac-
cess, and hence in time to first packet loss, for each queue.
DYNAMIC BANDWIDTH MANAGEMENT
Table 4
Performance and throughput loss comparison using TCP with interface queue
size smaller than congestion window limit.
Scheme FM JM Pkts. Tput
(pkts.) (pkts.) dropped (total acks recvd.)
Base IEEE 802.11 6.70 9.40 565 45065
Enhanced IEEE 802.11 2.12 2.39 33 35698
Figure 16. TCP congestion window behavior when interface queue size is
larger than congestion window limit.
Unmanaged release of packets from the queue results in un-
equal congestion window growth and causes unfairness. As
a result, the fairness metric (FM) and jitter metric (JM) of
the flows deteriorates, and the number of dropped packets are
significantly larger than that in the BM managed scheme, as
shown in table 4. The total number of dropped packets is
greater in the base IEEE 802.11 case because an entire win-
dow of packets may be dropped at a time before TCP resets
its congestion window size, whereas in the enhanced IEEE
802.11 case, a single packet loss results in window reset. The
overall throughput of the TCP flows in the enhanced IEEE
802.11 case, however, is smaller than that in the base IEEE
802.11 scenario. This is similar to the result for UDP flows
as shown in section 3.1.2. We also experimented with less
conservative Bp (f ) estimates, which resulted in a decrease
in throughput disparity between the base and enhanced IEEE
802.11 cases, at the cost of some performance deterioration.
Thus the Bp (f ) values can be used to trade-off performance
(as measured by the FM and JM) and throughput loss, as with
the UDP experiments.
Another scenario of running TCP over BM is setting each
nodes interface queuing limit to be larger (150 packets) than
the congestion window limit (128 packets) of a TCP flow. We
run the experiments for this scenario for 150 seconds. TCPs
congestion window size can never reach the maximum inter-
face queue size, and hence there is no packet loss as result of
queue overflow. In this case, we can expect TCPs congestion
window size to stay at its maximum limit without fluctuat-
ing, because there is no packet loss at the MAC layer either.
Figure 16 shows this behavior. Note that the slow conver-
gence speed of TCPs congestion window size does not im-
pact its throughput efficiency, as the interface queue is kept
non-empty at all times. However, in order to minimize queu-
ing delay, it is advisable to set TCPs congestion window
limit to a small value when running over a bandwidth man-
213
Table 5
Performance and throughput loss comparison using TCP with interface queue
size larger than congestion window limit.
Scheme FM JM Pkts. Tput
(pkts.) (pkts.) dropped (total acks recvd.)
Base IEEE 802.11 6.53 8.50 0 33804
Enhanced IEEE 802.11 2.51 2.72 0 26577
Figure 17. Single-hop ad hoc network testbed.
aged network. Table 5 compares the fairness performance and
throughput loss for the base and enhanced IEEE 802.11 sce-
narios for the case where congestion window limit is less than
the interface queue size. From the plot in figure 16, it is ob-
vious that the throughput disparity, as a percentage, between
the base and enhanced IEEE 802.11 cases in this scenario,
decreases with time.
3.2. Testbed experiments
We used our testbed experiments to evaluate the throughput
performance and the request-reply delay overhead in the pres-
ence of both physical channel errors as well as contention
from a limited number of active stations. Our testbed (see fig-
ure 17) consisted of 3 IBM ThinkPad laptops, each equipped
with an ORiNOCO PCMCIA 802.11b wireless card config-
ured in peer-to-peer ad hoc mode.
We used a rate-adaptive CBR audio streaming application
over UDP in our testbed experiments. The audio streaming
application could operate at 5 different QoS levels between
32 Kbps and 256 Kbps depending on the available channel
capacity perceived by the TBE. At the maximum QoS (256
Kbps), all audio samples were transmitted while at lower lev-
els fewer samples were sent, and the audio was reconstructed
through interpolation at the receiver. The purpose of the test-
bed experiments was to study the feasibility of our scheme in
a testbed with a realistic single-hop ad hoc network environ-
ment. The RA in the application and the TBE communicated
via the /proc interface.
214
Figure 18. Indoor testbed experiment plot.
3.2.1. Throughput performance
We conducted two throughput experiments, one indoors and
one outdoors. In each case, we started some unmanaged ping
sessions, as shown in figure 17, to bring about contention.
The ping ICMP packet transmission on the channel also ar-
tificially reduced its bandwidth so that the bandwidth per-
ceived by the audio streaming application actually fluctuated
between 32 and 256 Kbps depending on the physical errors.
In the absence of the pings, the reduction in perceived band-
width brought about by the physical errors alone was not suf-
ficient to cause the audio streaming application to adapt its
quality. The physical errors, at their worst, reduced the per-
ceived bandwidth by a few hundreds of Kbps. Given a 2 Mbps
channel and an application with a peak rate of 256 Kbps,
these errors thus had no effect on the application. Its qual-
ity level did not fluctuate. To bring about adaptation on the
part of the application, the physical errors had to vary the
available channel capacity for the application between 32 and
256 Kbps. Hence, we used the pings to contend with the ap-
plication for the channel and thus artificially reduce the avail-
able channel capacity it perceives to the necessary range. The
pings brought down the available channel capacity to around
500 Kbps so that fading and interference errors could then re-
duce it further below the 256 Kbps threshold needed for the
application to adapt.
SHAH ET AL.
Figure 18 shows the throughput performance for the in-
doors scenario. On the x-axis is time in 45 second units. The
y-axis shows the adaptation of the audio streaming applica-
tion, between 32 and 256 Kbps, to the change in available
channel capacity. The channel bit-rate was fixed at 2 Mbps
at the network cards. The perceived bandwidth variation-
tolerance was set at = 15% and the inter-update interval
was 100 packets. The BM was located on the same machine
as the sender, 12.0.0.11.
The flurry of re-negotiations with the BM on the left-
hand side of the plot corresponds to our moving the sender
(12.0.0.11) down to a secluded portion of the basement of the
building while the receiver (12.0.0.12) and the third laptop
(12.0.0.10) remained in the lab on the second floor. While
in the basement, the sender moved around, down narrow cor-
ridors, over staircases and through fire doors. As the level
of fading and interference changed drastically, the perceived
channel capacity also changed drastically and hence the flurry
of channel time re-negotiations. The contending pings also
were affected by the physical errors and produced variable
contention, thus inducing even greater instability in the appli-
cation QoS.
We then brought the sender back to the second floor, the
perceived bandwidth returned to around 500 Kbps, and the
quality of the audio returned to its maximum. We then placed
DYNAMIC BANDWIDTH MANAGEMENT
Figure 19. Outdoor testbed experiment plot.
the sender and receiver next to each other so that physical er-
rors were rare. The 3 dips in the graph on the right-hand side
correspond to experiments with no physical errors, but 3 dif-
ferent levels of contention due to 3 different ping rates. All
3 of these ping rates were greater than those used for the first
part of this experiment. In the first part, the pings reduced the
available channel capacity to around 500 Kbps and the phys-
ical errors dragged it further down. In this part there were
no physical errors, but the larger ping rates themselves took
the available channel capacity below 256 Kbps, causing re-
negotiation from the application. The purpose of this experi-
ment with no physical errors was to demonstrate the effect of
the contending ping sessions: they produce a reduction in the
perceived available channel capacity of the managed audio
streaming application, in a controlled fashion, and the reduc-
tion is a constant one.
Next, we performed another set of experiments outdoors.
The channel bit-rate was set at 5.5 Mbps for this experiment.
As before, we had pings produce contention to artificially re-
duce available channel capacity for the audio streaming flow.
Other parameters such as the value of and the inter-update
interval were the same as in the indoor experiment. In the
outdoor scenario, we used only two of the laptops. The BM
was once again co-located with the sender, 12.0.0.11. At
215
the start of the experiment, the sender 12.0.0.11 and the re-
ceiver 12.0.0.12 were next to each other on the sidewalk of
a street. Then, keeping the receiver 12.0.0.12 stationary on
the sidewalk, the sender 12.0.0.11 was moved away by a per-
son walking at a normal pace down the street on the sidewalk.
When the sender was around 150 meters away, the available
channel capacity perceived by the audio flow began fluctu-
ating due to signal fading effects. This resulted in a flurry of
re-negotiations shown in figure 19. The sender then wandered
for a while around the point 150 meters away before return-
ing to the starting position. As the sender moved closer to the
receiver, at one point, the available channel capacity returned
to its ping-induced constant level and the application returned
to its highest quality level.
We repeated our experiments using ARS (auto rate selec-
tion) feature of the wireless card, instead of using constant
rates 2 Mbps and 5.5 Mbps mentioned above. Our results
were very similar when using ARS as compared to when
using fixed rates. We also experimented with the BM at
the destination node, with no change in performance. The
request-reply delay overhead for re-negotiation requests does
not affect performance much because the application paral-
lelly continues transmitting at its previously allotted CTP un-
til the re-negotiation reply arrives, a few milliseconds later.
216
3.2.2. Request-reply delay
The request-reply delay is the time delay between the send-
ing of a request message and the receipt of a reply mes-
sage. All our control messages had a 32-byte payload. This
exchange of messages occurs both during flow establishment
as well as when perceived bandwidth changes significantly.
We set the bandwidth of the network to be 5.5 Mbps, as in
the case of the outdoor experiment. We used all 3 laptops for
the request-reply delay experiments, with 12.0.0.11 being the
sender, 12.0.0.12 being the receiver and the BM being located
on 12.0.0.10. We found that, if there is no contention, each
request-reply round-trip took 23 ms on average. In the pres-
ence of ping-induced contention, each request-reply round-
trip took 61 ms on average. Flow establishment occurs only
once per flow, obviously, and if the perceived bandwidth does
not change much, then the 2060 ms request-reply delay is a
small one.
4. Related work
In this section we discuss two areas of related work: (a) cen-
tralized channel allocation, and (b) distributed fair scheduling
in single-hop and multi-hop wireless networks.
In wireless network environment, past research has fo-
cused on flow scheduling at the access-point to achieve cer-
tain fairness criteria between flows competing for the wireless
channel [5,7,15]. Bianchi et al. [5] proposed the utility fair
criteria in bandwidth allocation, where each users bandwidth
is allocated in such a way that their individual utility is equal-
ized. It assumes that the central manager at the base station
has exact knowledge of the asymptotic utility curves of all the
applications, which might be difficult to obtain. The flows in
our scheme can specify a simple linear utility curve using just
two points. In our scheme, the BM guarantees a minimum
bandwidth for each flow, and allots the rest of the channel
capacity in a maxmin fashion to each flow up to its maxi-
mum request. We believe our approach is simple yet effective
in a smart-room where random users walk up to the room and
share the wireless channel. Another difference is that we use a
distributed peer-to-peer transmission (details in section 2.1),
rather than an access-point model, in allocating the channel
resources.
Another wireless network channel allocation scheme is the
effort-limited fair scheduling by Eckhardt and Steenkiste [9].
It adjusts the air time of a flow to meet its minimum band-
width requirement in response to channel error rates, only up
to a certain factor (called the power factor), to avoid starv-
ing other best-effort flows. The usage of air time to measure
the bandwidth requirement of a flow is similar to the chan-
nel time in our scheme. However, it is unclear how the power
factor can be chosen for different flows because this will give
preferential treatment to certain users. In our scheme, when
a flows minimum requirement cannot be satisfied, the flow
is simply rejected. This creates incentive for the users to set
a minimum channel time requirement as small as possible to
reduce the possibility of being denied access to the channel.
SHAH ET AL.
In [25], the authors propose an admission control scheme
for a peer-to-peer, single-hop, ad hoc wireless network model
similar to the one we have used. Their scheme requires the use
of special probe packets to obtain the service curve, which is
an estimate of network load. Using the service curve, one-
time admission control is performed. In contrast, our scheme
estimates network load using the data packets of the connec-
tion itself. Moreover, we perform dynamic bandwidth re-
negotiation over the course of the connection, in addition to
admission control at flow startup.
Another area of related work is the distributed weighted
fair scheduling (DWFS) schemes in single-hop and multi-
hop wireless networks [3,13,18,19,24]. As mentioned before,
our bandwidth management scheme is required to assist the
DWFS scheme when it is available. At the same time, as
shown in our experiments, our scheme also works well with-
out any underlying DWFS schemes. This is a very important
feature because todays IEEE 802.11 network interface card
only implements the standard DCF MAC protocol without
any DWFS extensions. Therefore, our bandwidth manage-
ment scheme, of which we already have a working prototype,
is highly deployable in todays smart-rooms.
5. Conclusion
In this paper, we presented an Admission Control scheme
to determine what fraction of channel time each flow in
a single-hop ad hoc wireless network receives. To this
end, we mapped the bandwidth requirement at the applica-
tion/middleware layer to a channel time proportion (CTP)
requirement at the MAC layer. We presented an applica-
tion/middleware layer rate control mechanism to ensure that
flows conform to their respective CTPs. Since one-time ad-
mission control is not sufficient to handle the changes in net-
work and flow characteristics, we also presented a Dynamic
Bandwidth Management system that adapts the flows respec-
tive CTPs during the course of their operation. The adaptation
can be a response to change in the network environment or
change in a particular flows traffic characteristics. The sim-
plicity and robustness of our system enables the incorporation
of elegant pricing and security features into it. We have devel-
oped a prototype implementation of the system and we have
used this implementation in a testbed, in addition to extensive
simulations, to demonstrate the feasibility and utility of our
scheme.
References
[1] G. Ahn, A. Campbell, A. Veres and L. Sun, Swan: Service differen-
tiation in stateless wireless ad hoc networks, in: Proceedings of IEEE
InfoCom, New York (June 2002).
[2] O. Angin, A. Campbell, M. Kounavis and R. Liao, The mobiware
toolkit: Programmable support for adaptive mobile networking, IEEE
Personal Communications, Special Issue on Adapting to Network and
Client Variability 5(4) (1998) 3244.
[3] B. Bensaou, Y. Wang and C. Ko, Fair medium access in 802.11 based
wireless ad hoc networks, in: Proceedings of IEEE MobiHoc, Boston,
MA (August 2000).
DYNAMIC BANDWIDTH MANAGEMENT
[4] D. Bertsekas and R. Gallager, Data Networks, 2nd ed., chapter 6
(Prentice-Hall, 1992).
[5] G. Bianchi, A. Campbell and R. Liao, On utility-fair adaptive services
in wireless networks, in: Proceedings of IEEE/IFIP IWQoS, Napa, CA
(May 1998).
[6] F. Cali, M. Conti and E. Gregori, Dynamic tuning of IEEE 802.11 pro-
tocol to achieve a theoretical throughput limit, IEEE/ACM Transactions
on Networking 8(6) (2000) 785799.
[7] T. Chen, P. Krzyzanowski, M. Lyu, C. Sreenan and J. Trotter, A sum-
mary of Qos support in SWAN, in: IEEE/IFIP IWQoS 1998, Napa, CA
(May 1998).
[8] S. Czerwinski, B. Zhao, T. Hodes, A. Joseph and R. Katz, An archi-
tecture for a secure service discovery service, in: Proceedings of ACM
MobiCom, Seattle, WA (August 1999).
[9] D. Eckhardt and P. Steenkiste, Effort-limited fair (ELF) scheduling for
wireless networks, in: Proceedings of IEEE InfoCom, Tel Aviv, Israel
(March 2000).
[10] F. Fitzek and M. Reisslen, Mpeg-4 and h.263 video traces
for network performance evaluation, IEEE Network Magazine
15(6) (2001) 4054, http://www-tkn.ee.tu-berlin.de/
research/trace/stat.html
[11] M. Grossglauser, S. Keshav and D. Tse, Rcbr: A simple and efficient
service for multiple time-scale traffic, in: Proceedings of ACM Sig-
Comm, Cambridge, MA (August 1995).
[12] E. Guttman, C. Perkins, J. Veizades and M. Day, Service location pro-
tocol, version 2, RFC 2608 (June 1999).
[13] V. Kanodia, C. Li, A. Sabharwal, B. Sadeghi and E. Knightly, Distrib-
uted multi-hop scheduling and medium access with delay and through-
put constraints, in: Proceedings of ACM MobiCom, Rome, Italy (July
2001).
[14] M. Kazantzidis, M. Gerla and S. Lee, Permissible throughput network
feedback for adaptive multimedia in AODV Manets, in: Proceedings of
IEEE ICC, Helsinki, Finland (June 2001).
[15] J. Kim and M. Krunz, Bandwidth allocation in wireless networks with
guaranteed packet-loss performance, IEEE/ACM Transactions on Net-
working 8(3) (2000) 337349.
[16] A. Kuznetsov, Linux traffic control (tc), http://www.sparre.
dk/pub/linux/tc
[17] R. Liao, R. Wouhaybi and A. Campbell, Incentive engineering in wire-
less LAN based access networks, in: Proceedings of IEEE ICNP, Paris,
France (November 2002).
[18] H. Luo, S. Lu and V. Bharghavan, A new model for packet schedul-
ing in multihop wireless networks, in: Proceedings of IEEE MobiCom,
Boston, MA (August 2000).
[19] H. Luo, P. Medvedev, J. Cheng and S. Lu, A self-coordinating approach
to distributed fair queuing in ad hoc wireless networks, in: Proceedings
of IEEE InfoCom, Anchorage, AK (April 2001).
[20] P. Marbach and R. Berry, Downlink resource allocation and pricing for
wireless networks, in: Proceedings of IEEE InfoCom, New York (June
2002).
[21] M. Mirhakkak, N. Schult and D. Thomson, Dynamic bandwidth man-
agement and adaptive applications for a variable bandwidth wireless
environment, IEEE JSAC 19(10) (2001) 19841997.
[22] Y. Qiu and P. Marbach, Bandwidth allocation in ad-hoc networks:
A price-based approach, in: Proceedings of IEEE InfoCom, San Fran-
cisco, CA (MarchApril 2003).
[23] The network simulator ns-2, http://www.isi.edu/nsnam/ns/,
updated October 2001.
217
[24] N. Vaidya, P. Bahl and S. Gupta, Distributed fair scheduling in a wire-
less LAN, in: Proceedings of ACM MobiCom, Boston, MA (August
2000).
[25] S. Valaee and B. Li, Distributed call admission control in wireless ad
hoc networks, in: Proceedings of IEEE VTC, Vancouver, Canada (Sep-
tember 2002).
[26] Y. Xue, B. Li and K. Nahrstedt, Price-based resource allocation in wire-
less ad hoc networks, in: Proceedings of IEEE IWQoS, Monterey, CA
(June 2003).
Samarth Shah received his B.E. degree in com-
puter science and engineering from the University of
Madras, India, in 1998. He is currently a Ph.D. can-
didate in the Department of Computer Science at the
University of Illinois at Urbana-Champaign. His in-
terests include quality of service (QoS) in wireless
networks.
E-mail: shshah@cs.uiuc.edu
Kai Chen received his B.Eng. degree in computer
science from Tsinghua University, Beijing, China, in
1995, and M.S. degree in computer science from the
University of Delaware, Newark, Delaware, USA, in
1998. Currently he is a Ph.D. candidate in the De-
partment of Computer Science at the University of
Illinois at Urbana-Champaign. From 1998 to 2000,
he worked as a research programmer at the National
Center for Supercomputing Applications (NCSA) at
Urbana, IL, USA. His research interests include mo-
bile ad hoc networks, transport layer issues in mobile networks, quality of
service, incentive engineering, and pervasive computing.
E-mail: kaichen@cs.uiuc.edu
Klara Nahrstedt is an Associate Professor at the
University of Illinois at Urbana-Champaign, Com-
puter Science Department. Her research interests are
directed towards multimedia middleware systems,
Quality of Service (QoS), QoS routing, QoS-aware
resource management in distributed multimedia sys-
tems, and multimedia security. She is the coauthor of
the widely used multimedia book Multimedia: Com-
puting, Communications and Applications published
by Prentice Hall, the recipient of the Early NSF Ca-
reer Award, the Junior Xerox Award, and the IEEE Communication Soci-
ety Leonard Abraham Award for Research Achievements. She is the editor-
in-chief of ACM/Springer Multimedia Systems Journal, and the Ralph and
Catherine Fisher Associate Professor. Klara Nahrstedt received her B.A. in
mathematics from Humboldt University, Berlin, in 1984, and M.Sc. degree
in numerical analysis from the same university in 1985. She was a research
scientist in the Institute for Informatik in Berlin until 1990. In 1995 she re-
ceived her Ph.D. from the University of Pennsylvania in the Department of
Computer and Information Science.
E-mail: klara@cs.uiuc.edu
 Mobile Networks and Applications 10, 219231, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

High Speed Networking Security:

Design and Implementation of Two New DDP-Based Ciphers
N. SKLAVOS
Electrical & Computer Engineering Department, University of Patras, Patras 26500, Greece

N.A. MOLDOVYAN
Specialized Center of Program Systems, SPECTR, Kantemirovskaya Str. 10, St. Petersburg 197342, Russia

O. KOUFOPAVLOU
Electrical & Computer Engineering Department, University of Patras, Patras 26500, Greece

Abstract. Using Data-Dependent (DD) Permutations (DDP) as main cryptographic primitive two new ciphers are presented: ten-round
Cobra-H64, and twelve-round Cobra-H128. The designed ciphers operate efficiently with different plaintext lengths, 64 and 128-bit, for
Cobra-H64 and Cobra-H128, respectively. Both of them use very simple key scheduling that defines high performance, especially in the case
of frequent key refreshing. A novel feature of Cobra-H64 and Cobra-H128 is the use of the Switchable Operations which prevent the weak
keys. The offered high-level security strength does not sacrifice the implementation performance, of both ciphers. Architecture, design
and hardware implementation of the two ciphers are presented. The synthesis results for both FPGA and ASIC implementations prove
that Cobra-H64 and Cobra-H128 are very flexible and powerful new ciphers, especially for high-speed networks. The achieved hardware
performance and the implementation area cost of Cobra-H64 and Cobra-H128 are compared with other ciphers, used in security layers of
wireless protocols (Bluetooth, WAP, OMA, UMTS and IEEE 802.11). From these comparisons it is proven that the two proposed are flexible
new ciphers with better performance in most of the cases, suitable for wireless communications networks of present and future.
Keywords: networking security, data-dependent permutations, Cobra-H64, Cobra-H128, encryption

1. Introduction
Security is a primary requirement of any wired and wireless
communication. Encryption algorithms are meant to provide
secure communications applications. However, if the sys-
tem is not designed property, it may fail. New encryption
algorithms have to perform efficiently in a variety of cur-
rent and future applications, doing different encryption tasks.
All hardware implementations have to be efficient, with the
minimum allocated number of logic gates. This means sim-
plicity in ciphers architectures with enough clever data
transformation components. The implementation of a com-
munication protocol, demands low power devices and fast
computation components which imply that the number and
complexity of the encryption operations should be kept as
simply as possible. A basic transformation in the operation of
todays ciphers is needed, including transformations in both
data and key blocks size. The ciphers of the near future have
to be key agile. Many applications need a small amount of
text to be encrypted with keys that are frequently changed.
Many well-known applications, like IPsec, use this way of ci-
phers operation. Although the most widely used mode of op-
eration is encryption with the same key for all the amount of
transport data, the previous mode is also very useful for future
applications. Ciphers that requiring subkeys precomputation
have a lower key agility due to the precomputation time, and
they also require extra RAM to hold the precomputed sub-
keys. This RAM requirement does not exist in the implemen-
tations of encryption algorithms, which compute their keys
during the encryption/decryption operation. Cellular phones
technology demands specific characteristics of the cryptog-
raphy science. Ciphers have to be compatible with wireless
devices restricted standards in hardware resources.
Data-Dependent (DD) Permutations (DDP) performed
with so called Controlled Permutation (CP) boxes [6,14,16]
appears to be very efficient cryptographic primitive for fast
hardware encryption. Security estimation of the DDP-based
ciphers CIKS-1 [11] and SPECTR-H64 [10] against linear
cryptanalysis has shown that DDP are efficient, provided they
are combined with other non-linear operations. The DDP-
based ciphers are proposed for hardware implementation with
low cost.
In this paper we present two new DDP-based ciphers
Cobra-H64 and Cobra-H128 and the results of their hard-
ware implementations. The design of the presented ciphers
takes into account some recommendations arising from the
linear and differential analysis of other DDP-based ciphers
[5,10,11]. Both proposed ciphers Cobra-H64 and Cobra-
H128 have been implemented in ASIC and FPGA hardware
modules. Two different VLSI architectures are examined for
each one of the proposed ciphers. The synthesis results of all
hardware integrations are presented in detail.
The paper is organized in the following way. In section 2
we consider construction of the controlled operational boxes
220
performing DDP. We present the design criteria and we de-
scribe the structure of the two new DDP-based block ciphers:
ten-round Cobra-H64 with 64-bit data input and twelve-round
Cobra-H128 with 128-bit data input. Section 3 describes in
details the encryption algorithm Cobra-H64. A feature of this
iterative cryptosystem is the use of two sellula-automaton-like
non-linear operations in its round transformation, as an addi-
tional primitive. Section 4 describes the cipher Cobra-H128
that has structure similar to the structure of Cobra-H64. In
section 5 we discuss the key scheduling and present results
on security estimation against differential analysis. Section 6
presents the hardware implementations cost and performance
(FPGA and ASIC). Comparisons of the proposed implemen-
tations, of both Cobra-H64 and Cobra-H128, with other block
ciphers are given. Finally conclusions and observations are
discussed in the last section.
2. Design of the controlled permutations
Controlled permutations can be easy performed with well
known interconnection networks (IN) [1,3] which were pro-
posed to construct key-dependent permutations [4,17]. How-
ever such use of IN do not effectively thwarts differential
cryptanalysis [20]. Regarding cryptographic applications it
is more attractive to use IN to perform DDP on data sub-
blocks [15] and subkeys [12]. An operational box Pn/m , per-
forming permutations on n-bit binary vectors depending on
some controlling m-bit vector V , is called Controlled Per-
mutation box (CPB). In the case that the controlling vector
depends on a data subblock, the CP box performs DDP. The
fast CP boxes can be constructed using elementary switching
elements P2/1 , figure 1(a), as elementary building blocks per-
forming controlled transposition of two one-bit inputs x1 and
x2 . In general case, each P2/1 -box is controlled with one bit
and forms two-bit output (y1 , y2 ), where y1 = x1+ and
y2 = x2 .
Taking into account that it is very desirable to mini-
mize the time delay, while performing CP-box permutations,
the layered topology of IN can be considered as the main one
since it permits to design very fast CPB. Layered CPB are
constructed as superposition of S = 2m/n active layers, sep-
arated with S 1 fixed permutations 1 , . . . , S1 that are
implemented in hardware as simple connections. Each active
layer, figure 1(b), in a CPB with n-bit input is represented by
the set of n/2 parallel elementary boxes P2/1 . The general
structure of the layered CPB is shown in figure 1(c). Its nota-
tion is presented in figure 1(d). In all figures of this paper, the
solid lines indicate data movement, while the dotted lines cor-
responding to CP-boxes indicate controlling bits. A CP-box
inverse of the box Pn/m is denoted as P1 n/m .
We assume that in a layered CP-box all elementary switch-
ing elements are consecutively numbered from left to right
from top to bottom and the j th bit of vector V controls the
j th switching element P2/1 . In accordance with the num-
ber of layers the vector V can be represented as concate-
nation of 2m/n vectors V1 , V2 , . . . , V2m/n {0, 1}n/2 , i.e.,
V = (V1 , V2 , . . . , V2m/n ).
SKLAVOS ET AL.
Figure 1. (a) P2/1 -box, (b) structure of one active layer, (c) general structure
of the layered CP boxes and (d) P1
n/m -box.
Controlled permutations performed with the box Pn/m ,
can be characterized using an ordered set of the modifi-
cations {0 , 1 , . . . , 2m 1 }, where each modification i ,
i = 0, 1, . . . , 2m 1, is a fixed permutation of some n-bit
sets. Permutations i are called CP-modifications.
Notation. Let {0, 1}n denote the set of all n-bit binary vec-
tors X = (x1 , . . . , xn ). Let X denote

also decimal value (or
simply value) of the vector X: X = ni=1 xi 2i1 .
Let X Y denote the bit-wise XOR operation performed
on X, Y {0, 1}n and XY denote bit-wise AND operation.
Let Y = Xk denote the cyclic rotation of the word X
by k bits (0
k < n), where Y = (y1 , . . . , yn ) is the output,
i {1, . . . , n k} we have yi = xi+k , and i {n k +
1, . . . , n} we have yi = xi+kn .
Let Xl = (x1 , . . . , xn/2 ) and Xl = (xn/2+1 , . . . , xn ) de-
note the least and the most significant bits of X {0, 1}n .
Definition. CP-boxes Pn/m and P1 n/m are mutual inverses, if
for all possible values of the vector V the corresponding CP-
modifications V and 1 V are mutual inverses.
One active layer can be considered as some single-layer
CP box Sn . It is evidently that P2/1 = P1 2/1 , therefore
1
Sn = Sn . A layered CP box Pn/m can be represented as
superposition Pn/m = S(V1 ) 1 S(V2 ) 2 s1
S(V2m/n ) . The respective box P1
n/m has the following struc-
ture P1 1 1
n/m = S 2m/n1 S(V2m/n1 ) 2m/n2
(V2m/n )
11 S(V1 ) . Thus, to construct inverse of the CP-box Pn/m it
is sufficient to number the boxes P2/1 from left to right from
1
bottom to top and to replace i by 2m/ni . We shall assume
that in the boxes P1
n/m switching elements P2/1 are consecu-
tively numbered from left to right from bottom to top. Note
that the vector Vj corresponding to the j th active layer in the
HIGH SPEED NETWORKING SECURITY
Figure 2. Structure of boxes: (a) P8/12 and (b) P1
8/12 .
Figure 3. Structure of the CP-boxes: (a) P32/96 and (b) P1
32/96 .
box Pn/m controls the (2m/n j + 1)th active layer in P1n/m
(see figure 2).
The cipher Cobra-H64 (Cobra-H128) uses the boxes P32/96
and P1 1
32/96 (P64/192 and P64/192). Each one of them is
constructed using four (eight) parallel boxes P8/12 and four
(eight) parallel boxes P1
8/12 which are shown in figure 3 (fig-
ure 4). The P8/12 boxes are connected with the P1
8/12 boxes in
accordance to the principal each to each.
While designing the single key encryption algorithms
Cobra-H64 and Cobra-H128 our strategy was oriented to the
extensive use of the controlled permutations that are very fast
and with low cost for hardware implementation. Our design
criteria were the following:
1. The encryption algorithm should be an iterated 64-bit or
128-bit block cipher.
2. The cipher should be fast, in the case of frequent key re-
freshing. Therefore the encryption algorithm should be
able to perform encryption and decryption with simple and
fast change of the used subkeys sequence.
3. Round transformations of data subblocks should be char-
acterized by high parallelism.
221
Figure 4. Structure of the CP-boxes: (a) P64/192 and (b) P1
64/192 .
4. Except DDP some additional non-linear operation should
be used in the round transformation.
The encryption/decryption schemes of Cobra-H64 and
Cobra-H128 are described by the following formulas: C =
T(e=0) (M, K) and M = T(e=1) (C, K), where M is the plain-
text, C is the ciphertext (M, C {0, 1}64 for Cobra-H64
and M, C {0, 1}128 for Cobra-H128). K is the secret key,
T is the transformation function, and e {0, 1} is a parame-
ter defining encryption (e = 0) or decryption (e = 1) mode.
The secret key is considered as concatenation of four subkeys:
K = (K1 , K2 , K3 , K4 ). For i = 1, 2, 3, 4 Ki {0, 1}32 for
Cobra-H64 and Ki {0, 1}64 for Cobra-H128. The ciphers
use no preprocessing to generate subkeys. The extended key
Q(e) is formed as simple sequence of subkeys Ki taken in
respective order.
The both ciphers use the same iterative structure which is
shown in the following figure 5. Encryption begins with the
Initial Transformation. Then r rounds of data transformation
are based on procedure Crypt(e) , followed by the Final Trans-
formation. First, the data input X is divided to subblocks L
and R. Then Initial Transformation is executed which per-
forms XOR-ing each between the data subblocks and two dif-
ferent subkeys: L0 = L O3 and R0 = R O4 .
The encryption procedure is performed in accordance with
the following pseudo-algorithm:
For j = 1 to r 1 do:
Execute transformation:
(e)
(Lj , Rj ) = Crypt(e) (Lj 1 , Rj 1 , Qj );
Swap the data subblocks:
Rj = T , Rj := Lj , Lj := T ;
End For loop;
Execute transformation:
(Lr , Rr ) = Crypt(e) (Lr1 , Rr1 , Q(e)
r ).
Procedure Crypt(e) represents round encryption function,
(e)
where Qj is the round key, used in the j th encryption round.
Encryption finishes with procedure of final transformation:
L = Lr O1 and R  = Rr O2 . The ciphertext block is
Y = (L , R  ). The ciphers use different procedures Crypt(e)
222 SKLAVOS ET AL.

Table 1
Specification of the subkeys A(i) in Cobra-H64.

j= 1 2 3 4 5 6 7 8 9 10
(1)
Aj = O1 O4 O3 O2 O1 O1 O2 O3 O4 O1
(2)
Aj = O2 O1 O4 O3 O4 O4 O3 O4 O1 O2
(3)
Aj = O3 O2 O1 O4 O3 O3 O4 O1 O2 O3
(4)
Aj = O4 O3 O2 O1 O2 O2 O1 O2 O3 O4

Figure 6. Swapping of subkeys.

Figure 5. Structure of Cobra-H64 and Cobra-H128.

and different numbers of rounds r = 10 for Cobra-H64 and

r = 12 for Cobra-H128.

3. The block cipher Cobra-H64

3.1. Formation of the round keys

(e)
Each of the round keys Qj consists of four e-dependent
(e)
round subkeys A(1) , A(2), A(3) , A(4) {0, 1}32, i.e., Qj =
(e)
(A(1), A(2) , A(3) , A(4))j .Table 1 specifies the rounds sub- Figure 7. Procedure Crypt(e) in Cobra-H64.
keys and their correspondence to the secret key.
For each one of the ten rounds each of these subkeys is 3.2. Data-dependent permutations
used while performing two operations G. Taking into ac-
count that the three inputs of the operation G are different, the To perform DDP Cobra-H64 uses CP boxes P(V )
32/96 and
role of each subkey changes from one round to another one. 
While data decryption the subkeys are generated as simple (P1
32/96 )
(V ) described in section 2 (figure 3). Controlling vec-

(e) tors for these CP boxes are formed using the same extension
swapping subkeys Qj with single-layer box P128/1 which is
box E implemented with simple connections. The input of
represented by two parallel boxes P(e)
64/1 (e = 0 for encryption the E-box is the current value of the left data subblock L. Let
(e)
and e = 1 for decryption). The box P64/1 is some single-layer the vector V = (V1 , V2 , V3 , V4 , V5 , V6 ) be the 96-bit output
CP box in which all elementary switching elements are con- of the E-box. The extension box provides the following rela-
trolled with the same bit e. The pairs (K1 , K3 ) and (K2 , K4 ) tions:
are inputs of the corresponding boxes P(e)
64/1 (figure 6). Four V1 = Ll , V2 = L6 , V3 = L12 ,
l l
(e)
32-bit outputs of two boxes P64/1
are the e-dependent sub- V4 = Lh , V5 = L6
h , V6 = L12
h .
keys Oi (i = 1, 2, 3, 4). Thus, we have Oi = Ki , if e = 0,
and O1 = K3 , O2 = K4 , O3 = K1 , O4 = K2 , if e = 1. Inverting a bit in L causes the inversion of three bits of V .
Correct change of the encryption mode for the decryption Thus, each bit of L influences three boxes P2/1 in the box
one is also defined by the respective change of the fixed per- P(V ) 1 (V  ) . While designing
32/96 and three P2/1 -boxes in (P32/96 )
mutation (e) in procedure Crypt(e) presented in figure 7. the box E we used the following criterion.
HIGH SPEED NETWORKING SECURITY
Figure 8. Structure of the switchable permutation.
Criterion. For arbitrary given vector L the permutation of
each input bit of the CP box must be defined by six different
bits of L.
Due to realization of this criterion each bit of L influences
exactly six bits of R. Such distribution of the controlling bits
provides that arbitrary input bit of the boxes P32/96 and P1
32/96
moves to each output position with the same probability pro-
vided L is a uniformly distributed random variable.
3.3. Switchable fixed permutation (e)
Switchable fixed permutation (e) performs permutation (0)
when enciphering, and (1) when deciphering. Change of the
permutation is performed as simple swapping outputs of the
fixed permutations (0) and (1) with single-layer box P(e)
64/1
(see figure 8). Permutations (0) and (1) contain two cy-
cles. The first cycle corresponds to identical permutation of
the most significant input bit x32 . The second cycle is de-
scribed by the following equations:
(0) (x1 , x2 , . . . , x31 ) = (x1 , x2 , . . . , x31 )5 ,
(1) (x1 , x2 , . . . , x31 ) = (x1 , x2 , . . . , x31 )26 .
The role of the fixed permutation, with such structure, is to
provide each input bit of the CP-box P(V )
32/96 influences each

output bit of the CP-box (P1
32/96 )
(V ) . Indeed, let consider
V .
(V )
the case that V = Each input bit of the P32/96-boxwith
the same probability moves to each input digit of the opera-
tion (e) . If it moves to the most significant digit, then it re-
turns to its initial digit at the output of the CP box (P1
32/96 )
If it moves to arbitrary other digit at the input of (e) , then it
can be moved in all output digits of the (P1 (V  ) -box ex-
32/96 )
cept its initial position.
Thus the permutation (e) improves the resultant DDP cor-
(V )
responding to performing sequential operations P32/96 and

(P1
32/96 )
(V ) . Indeed even in the case of V = V  the super-
position P(V )
32/96
(e) (P1 )(V  )
forms an effective CP-box
32/96
permutation, and all modifications are permutations having
the same cycle structure (all modifications contain one cycle
with length 1 and one cycle with length 31).
Actually, during encryption in general case we have
V = V  , since after execution of the operation P(V )
32/96 the
223
permutation I is performed on L. Investigating the role of
the fixed permutation between two mutually inverse CP-box
operations we have performed many statistic experiments.
These experiments have shown that the use of such permu-
tation, significantly improves the properties of the transfor-
mation performed with two mutually inverse CP-boxes. Such
fixed permutation defines some internal mechanism of the op-
timization of the distribution of the influence of the left data
subblock on the elementary switching elements of the both
CP-boxes.
3.4. Permutational involution I
Permutational involution I performed on the left data sub-
block is used to strengthen the avalanche effect. Let yi  and
yj  be the output bits corresponding to the input bits xi and
xj . To design involution I we have used the following two
criteria:
(i) i, j : |j i|
3 should be |j  i  |  4;
(ii) i should be |i i  |  6.
Integers 4 and 6 are selected in order to define each bit of
the left subblock influences as many as possible bits after the
outputs of the both operations G be XORed with R. These
criteria are satisfied by the involution
I = (1, 17)(2, 21)(3, 25)(4, 29)(5, 18)(6, 22)(7, 26)(8, 30)
(9, 19)(10, 23)(11, 27)(12, 31)(13, 20)(14, 24)(15, 28)
(16, 32).
For example, Out(1) = In(17) and Out(17) = In(1). The
involution I provides that changing one bit of the left data
subblock causes inversion from 2 to 8 bits of the right data
subblock after the outputs of both operations G are XORed
with R.
3.5. Non-linear operation G
Operation GA A (L) is described by the following expression:
W = L0 A0 L2 L3 L1 L2 L1 L3 L2 A1 A1 L3
A0 L1 L2 ,
(V  ) .
where binary vectors Lj , Aj , and Aj are expressed as fol-
lows:
L0 = L = (l1 , l2 , . . . , l32 ),
L1 = (1, l1 , l2 , . . . , l31 ),
L2 = (1, 1, l1 , l2 , . . . , l30 ),
L3 = (1, 1, 1, l1 , l2 , . . . , l29 ),
 
A0 = A = a1 , a2 , . . . , a32 
,

   

A1 = 1, a1 , a2 , . . . , a31 ,
 
A0 = A = a1 , a2 , . . . , a32 
,
 
A1 = 1, a1 , a2 , . . . , a31

,
 
A2 = 1, 1, a1 , a2 , . . . , a30 
.
224 SKLAVOS ET AL.

Figure 10. (a) Swapping subkeys and (b) vectors L and (L).

(e)
with two single-layer CP-boxes P128/1. Each of the CP-boxes
P(e) (e)
128/1 contains 64 parallel P2/1 -boxes controlled with the
(e)
same bit e. The left (right) inputs of the P2/1 -boxes corre-
spond to the left (right) 64-bit input of the CP-box P(e) 128/1 .
The pairs of subkeys (K1 , K3 ) and (K2 , K4 ) are inputs of the
(e)
corresponding boxes P128/1 (figure 10(a)). Four 64-bit out-
puts of two boxes P(e)
128/1 are denoted as Oi (i = 1, 2, 3, 4).
Thus, we have Oi = Ki , if e = 0, and O1 = K3 , O2 = K4 ,
O3 = K1 , O4 = K2 , if e = 1. The vectors L and (L) are
swapped with the third CP box P(e)
128/1 (figure 10(b)).

4.2. Data-dependent permutations

Figure 9. Round function Crypt(e) of Cobra-H128.
(V )
To perform DDP Cobra-H128 uses CP boxes P64/192 and
Table 2 
Specification of the subkeys A(i) in Cobra-H128. (P1
64/192)
(V ) described in section 2 (figure 4). In Cobra-

H128, the controlling vectors V and V  are formed using the

j= 1 2 3 4 5 6 7 8 9 10 11 12
same procedure. This procedure includes two XOR transfor-
(1)
Aj = O1 O4 O3 O2 O1 O3 O3 O1 O2 O3 O4 O1 mations and one fixed permutation that defines which bit of
(2) the vector V controls which elementary CP box P2/1.
Aj = O2 O3 O4 O1 O2 O4 O4 O2 O1 O4 O3 O2
(3)
Aj = O3 O2 O1 O4 O3 O1 O1 O3 O4 O1 O2 O3 Let consider the formation of the vector V . The cor-
(4)
Aj = O4 O1 O2 O3 O4 O2 O2 O4 O3 O2 O1 O4 respondence between bits of the controlling vector V =
(V1 , . . . , V6 ) and elementary switching boxes P2/1 of the
(V )
P64/192-box is given in table 3, where rows indicate ac-
4. The block cipher Cobra-H128 tive layers and numbers correspond to indices of the bits
of vectors L, L(1) , and L(4) . The rows corresponding
Procedure Crypt(e) of Cobra-H128 is presented in figure 9. to vectors V1 and V4 indicate bits of L. The rows cor-
responding to vectors V2 and V5 indicate bits of L(1) .
4.1. Formation of the round keys The rows corresponding to vectors V3 and V6 indicate bits
of L(4) . For example, accordingly to table 3 we have:
Each one of the twelve round keys Q(e) j consists of four
(4) (4) (4) (4) (4) (4) (4) (4)
V3 = (l13 , l14 , . . . , l32 , l1 , . . . , l12 , l10 , l11 , l9 ), V4 =
e-dependent round subkeys A , A , A(3), A(4) {0, 1}64,
(1) (2) (l33 , l34 , l35 , . . . , l64 ). The vector V  is formatted with simi-
i.e. Q(e)j = (A(1) , A(2), A(3) , A(4) )(e)
j . Each of the sub-
lar way.
(1) (2) (3) (4)
keys A , A , A , and A is specified in table 2 via The distribution of the bits of the controlling data subblock
key elements O1 , O2 , O3 , O4 as one of the subkeys Ki (i is a critical part in the design of the DDP-based ciphers. When
{1, . . . , 4}) depending on j {1, . . . , 12} and e {0, 1}. composing table 3, several iterations have been performed,
Each one of the subkeys Ki is used while performing two followed by differential analysis and modification of the dis-
operations (one CP-box operation and one operation G), for tribution table. Combining subkeys with controlling data sub-
each one of the twelve rounds. Note that the role of each sub- block makes the DDP in Cobra-H128 to be key-dependent. It
key changes from one round to another one. is easy to see that table 3 satisfies the criterion of section 3.2.
Change of the ciphering mode is performed using two This provides that (i) each bit of L influences exactly six
mechanisms: (i) changing key scheduling and (ii) swap- bits of R and (ii) arbitrary input bit of the boxes P64/192 and
ping vector L and output (L) of the permutation  be- P1
64/192 moves to each output position with the same proba-
fore fulfilling two operations G. To provide correct change bility, provided L is a uniformly distributed random variable.
(V )
of the transformation mode j {1, . . . , 12} the condi- In one round between the operations P64/192 and

tion (A(1), A(2) , A(3) , A(4))(1)
j = (A , A , A , A )13j
(3) (4) (1) (2) (0) (P1
64/192)
(V ) the permutational involution I is performed. It

must be hold. Fast change of the key scheduling is performed is described as follows: Y = (Y1 , . . . , Y8 ) = I(X1 , . . . , X8 ),
HIGH SPEED NETWORKING SECURITY 225

Table 3
(V )
Distribution of the controlling bits in the box P64/192 .

V1 31 32 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 1 2
V2 10 24 25 26 29 13 27 16 1 2 31 32 3 4 19 6 7 8 9 23 11 12 28 15 14 30 17 18 5 20 21 22
V3 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 1 2 3 4 5 6 7 8 12 10 11 9
V4 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
V5 55 56 57 58 59 60 61 62 63 64 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
V6 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 33 34 35 36 37 38 39 40 41 42 43 44

where 64-bit input (output) binary vector X(Y ) is represented
as concatenation of eight bytes Xi (Yi ), i {1, . . . , 8}, and
Y1 = X64 , Y2 = X54 , Y3 = X44 , Y4 = X34 ,
Y5 = X24 , Y6 = X14 , Y7 = X84 , Y8 = X74 .
Use of this permutational involution provides that in one
(V )
round each input bit of the box P64/192 influences all output

bits of the box (P1
64/192)
(V ) improving statistical properties
of the round function. Permutation I introduces no time delay
since it is implemented in hardware as simple connections of
wiring.
4.3. Non-linear operation G
and Cobra-H128 are fast in the case of key refreshing, since
they are free of external key scheduling.
The ciphers Cobra-H64 and Cobra-H128, in comparison
with the prototype SPECTR-H64, have the following fea-
tures:
1. They use the initial secret key in each round.
2. The round transformation includes fixed permutational in-
volution, performed on one of the data subblocks.
3. The round transformation includes one switchable opera-
tion that prevents the weak keys with the structure K =
(X, X, X, X).

The role of the fixed permutation  in procedure Crypt(e)
is to make each bit of the left data subblock influences more
different digits of R while combining outputs of two opera-
tions G with R. The permutation  contains four cycles of
the length 16 and is described as follows:
(1, 50, 9, 42, 17, 34, 25, 26, 33, 18, 41, 10, 49, 2, 57, 58)
(3, 64, 43, 24, 19, 48, 59, 8, 35, 32, 11, 56, 51, 16, 27, 40)
(4, 7, 28, 47, 52, 23, 12, 63, 36, 39, 60, 15, 20, 55, 44, 31)
(5, 14, 13, 6, 21, 62, 29, 54, 37, 46, 45, 38, 53, 30, 61, 22).
The transformation W = GA A (L) defining the operation
GA A is described as follows:
W = L0 A0 L1 A0 L2 L5 L6 A1 A2 A1
L3 L4 L1 L4 L6 L2 L6 A1 L1 L2 L4 A1 ,
where j {0, 1, . . . , 6} we have Lj = L64j and A0 =
A, A1 = (1, a1 , . . . , a63 ), A2 = (1, 1, a1 , . . . , a62 ), (A = A
or A = A ).
5. Discussion
5.1. Peculiarities of Cobra-H64 and Cobra-H128
In ciphers Cobra-H64 and Cobra-H128 extensive use of the
CP-box operations is applied. They are used in three differ-
ent ways: (i) as DDP that are one of two basic cryptographic
primitives, (ii) to swap subkeys when changing encryption
mode for decryption one, and (iii) to switch permutation (e)
when changing ciphering mode. In addition to DDP, these
ciphers use the non-linear operation G. The round transfor-
mation of both ciphers is characterized by high parallelism
that provides pre-requisites of their high performance. Due to
the use of very simple key scheduling the ciphers Cobra-H64
The ciphers Cobra-H64 and Cobra-H128 have very sim-
ilar structure of the procedure Crypt(e). The differences
are based on the use of (i) the operations having different
size of input (32-bit for Cobra-H64 and 64-bit for Cobra-
H128), (ii) different switchable operations: switchable per-
mutation (e) in Cobra-H64 and e-dependent swapping in
Cobra-H128, (iii) the use of the subkeys to calculate control-
ling vectors in Cobra-H128, while in Cobra-H64 the control-
ling vectors depend only on the current value of the left data
subblock.
5.2. Security estimations
We have considered different variants of the differential crypt-
analysis. We have obtained that the fewer active bits in the
difference the higher the probability of the differential charac-
teristic. This corresponds to the results of the analysis of other
DDP-based ciphers. Our best attack against Cobra-H64 and
Cobra-H128 corresponds to two-round difference with one
active bit. This difference passes two rounds with probability
p(2) = 1.16 219 for Cobra-H64 and p(2) = 1.13 229 for
Cobra-H128. Probability of the best two-round characteristic
of SPECTR-H64 is p(2) = 1.15 213 . Minimal number of
rounds required to thwart differential attacks is 8 for Cobra-
H64 and 10 for Cobra-H128 and SPECTR-H64.
Our preliminary linear analysis of Cobra-H64 and Cobra-
H128 has shown that they are secure against linear attacks for
number of rounds r  5. Accordingly to [10] SPECTR-H64
is secure against linear attack for values r  6. High degree of
the algebraic normal form and the complexity of the Boolean
function describing round transformation of the developed ci-
phers prevent the interpolation and high order differential at-
tacks. Our statistic experiments has shown that 4 round of
Cobra-H64 and 5 round of Cobra-H128 are sufficient to sat-
isfy test criteria proposed for the AES finalists [18].
226
Because of the very simple key scheduling used in the pro-
posed ciphers it appears to be important to study how a single
bit of key statistically influences ciphertext (keys propagation
property). For this purpose we have used the criteria of [18]
considering the secret key as input vector and fixing different
plaintexts. Such statistic testing has shown that five rounds
of Cobra-H64 and six rounds of Cobra-H128 are sufficient to
satisfy test criteria.
The used key scheduling is secure against basic related-
key attacks. In spite of the simplicity of the key schedule
the keys K  = (X, Y, X, Y ) or K  = (X, X, X, X), where
X, Y {0, 1}32 for Cobra-H64 or X, Y {0, 1}64 for Cobra-
H128, are not weak, since encryption and decryption require
changing the parameter e. It seems to be difficult to calculate
a semi-weak key-pair for presented ciphers, if it is possible
at all. Thus, the role of the switchable permutation (e) is
preventing the weak keys. For example, for SPECTR-H64
which uses no switchable operations for all X the 256-bit key
K = (X, X, X, X, X, X, X, X) is weak.
6. Hardware implementation
6.1. ASIC and FPGA devices
Hardware implementations of both proposed ciphers are de-
signed and coded in VHDL hardware description language.
Both Cobra-H64 and Cobra-H128 were implemented using
two complete different implementation hardware modules:
Application Specific Integrated Circuit (ASIC) and Field Pro-
grammable Gate Array (FPGA). The performance character-
istics of both ASICs and FPGAs are substantially different
compared with a general-purpose microprocessor. ASICs and
FPGAs have the advantage that can use all the resources for
Figure 11. (a) Full Rolling and (b) Pipeline architectures.
SKLAVOS ET AL.
pipelining data transformation, or parallel processing. On
the other hand, the internal structure of the microprocessors
functional units limits the parallel processing and pipelining
transformation. In addition, the instruction parallelism level
is a factor of great importance that must be taken under con-
sideration for microprocessor performance. Furthermore, the
hardware devices of these types can operate on arbitrary size
of words, in contrast with processors, that operate only on
fixed-sized words.
ASICs are in general far more expensive devices due to
time consuming and high cost fabrication procedure, which
is done by expertise industry departments. FPGAs can be
reached from anyone, since they are enough cheaper and can
be programmed or reconfigured by the designers/researchers.
FPGAs have the major advantage that can perform a com-
pletely different task/function after simple designers recon-
figuration. ASICs performance is tight and can not be mod-
ified after the chips fabrication. The offered reconfiguration
has a speed penalty in these devices. ASICs have higher speed
performance in comparison with FPGAs. This is due to the
fact that the reconfiguration in FPGAs causes delays, intro-
duced by the dedicated circuits parts needed to reconfigura-
tion. In general, any implemented system of digital logic in
an FPGA, is slower than ASIC implementation of the same
system.
6.2. Implementation architectures for Cobra-H64 and
Cobra-H128
Both Cobra-H64 and Cobra-H128 are examined in hardware
implementation by using two different architectures: Full
Rolling and Pipeline for both ASIC and FPGA devices. The
used Full Rolling architecture is shown in figure 11(a). It is
a typical architecture for secret key block cipher implementa-
HIGH SPEED NETWORKING SECURITY
tion. This architecture operates efficiently for both encryption
and decryption process. According to this architecture only
one block of plaintext/ciphertext is transformed at a time. The
necessary number of clock cycles to encrypt/decrypt a data
block is equal to the specified number of cipher rounds (10
for Cobra-H64 and 12 for Cobra-H128). The key expansion
unit produces the appropriate round keys which are stored and
loaded in the used RAM blocks. One round of the encryption
algorithm is performed by the Data Transformation Round
Core. This core is a flexible combinational logic circuit and it
is supported by a n-bit register and n-bit multiplexer (64-bit
for Cobra-H64 and 128-bit for Cobra-H128). In the first clock
cycle, the n-bit plaintext/ciphertext is forced into the Data
Transformation Round Core. Then in each clock cycle, one
round of the cipher is performed and the transformed data are
stored into the n-bit register. According to Full Rolling archi-
tecture a 64-bit data block is completely transformed every 10
clock cycles for Cobra-H64 (10 transformation rounds). The
operation of Cobra-H128 (12 transformation rounds) needs
12 clock cycles in order a 128-bit plaintext/ciphertext to be
generated
The second proposed architecture, figure 11(b), is a
N-stage pipeline architecture. The main characteristics of this
are: (i) the pipelining used technique, and (ii) the usage of a
RAM for the round keys storage and loading, which are pre-
computed. Pipelining is not possible to be applied in many
cryptographic applications. However, Cobra-H64 and Cobra-
H128 block ciphers structures provide the availability to be
implemented with pipelining technique. The pipelining ar-
chitecture offers the benefit of the high-speed performance.
The implementation can be applied in applications with hard
throughput needs. This goal is achieved by using a number
of operating blocks with a final cost to the covered area. The
proposed architecture uses 10 basic round blocks for Cobra-
H64 and 12 basic round blocks for Cobra-H128, which are
cascaded by using equal number of pipeline registers. Based
on this design approach, 10 and 12 different n-bit data blocks
can be processed at the same time, for Cobra-H64 and Cobra-
H128, respectively. Pipeline proposed architecture produces
a new plaintext/ciphertext block every clock cycle.
Implementation synthesis results.
Architecture\Hardware device FPGA technology
Covered area
CLBs FGs
Cobra-H64 615 1229
(Full Rolling)
Cobra-H64 3020 6040
(10-stage Pip.)
Cobra-H128 2364 4728
(Full Rolling)
Cobra-H128 22080 44160
(12-Stage Pip.)
227
6.3. VLSI implementations synthesis results
The synthesis results are shown in table 4, for Cobra-H64
and Cobra-H128, for both hardware implementations (ASIC
and FPGA), where: D Flip-Flops (DFFs), Configurable Logic
Blocks (CLBs), Function Generators (FGs).
The above synthesis results for both implementations
(ASIC and FPGA) prove that the Pipeline architectures of
Cobra-H64 and Cobra-H128 have very high speed perfor-
mance. Especially Cobra-H64 throughput is up to 5.5
and 7.1 Gbps for FPGA and ASIC implementation, respec-
tively. Cobra-H128 throughput reaches the values of 11 and
12.1 Mbps for the same implementation devices. On the other
hand, Full Rolling architectures for both proposed ciphers
allocate minimized area resources with good data rate. Of
course, for the Full Rolling architectures, of both Cobra-H64
and Cobra-H128, the main goal is the minimized allocated
area resources, with good achieved throughput. The opera-
tion frequency for both ciphers is very high, for both proposed
architectures and for all the examined hardware modules. Es-
pecially, Cobra-H64 frequency ranges between 82 MHz to
110 MHz. Cobra-H128 operates up to 90 MHz for FPGA
devices and up to 95 MHz for ASIC approaches. It is obvi-
ous that according to the applications major demands, area
or performance, the designer can use the Full Rolling or the
Pipeline architecture respectively, for each one of the two pro-
posed ciphers. The kind of application and the characteristics
of the application itself, will determine the use of the FPGA
or the ASIC as the integration device, for the implementa-
tion either Cobra-H64 or Cobra-H128. Both Cobra-H64 and
Cobra-H128 architecture simplicity in addition to the offered
high-level security strength make their hardware integration
very useful in wireless communications networks.
In order to evaluate the very good performance, in hard-
ware terms, of both Cobra-H64 and Cobra-H128 we com-
pare the proposed ciphers implementations with the encryp-
tion algorithms that are used in todays wireless protocols.
Especially in figures 12 and 13, the proposed ciphers perfor-
mance is compared with the best hardware implementations
of ciphers used in wireless communications protocols. In
IEEE 802.11 first versions (ad), the Wired Equivalent Pri-
vacy (WEP) is widely used to ensure privacy in the trans-
Table 4
ASIC technology
(Xilinx) (0.33um)
F Rate Area F Rate
(MHz) (Mbps) (sqmil) (MHz) (Mbps)
DFFs
204 82 525 2694 100 640
640 85 5500 14640 110 7.1 Gbps
399 86 917 6364 90 1 Gbps
1.5 Gbps 90 11500 48252 95 12.1 Gbps
228 SKLAVOS ET AL.

Figure 12. Proposed ciphers implementations FPGA comparisons.

mission channel. Especially WEP is based on RC4. The to achieve security due to the external attacks over the trans-
latest working group (802.11i) adopts AES as the block ci- mitted data. Finally, the Wireless Transport Layer Security
pher of this IEEE protocol. Bluetooth security is based on (WTLS) ensure encryption in both Wireless Application Pro-
SAFER+ cipher. UMTS uses both AES and SAFER+ in order tocol (WAP) and Open Mobile Alliance (OMA). DES, IDEA
HIGH SPEED NETWORKING SECURITY 229

Figure 13. Proposed ciphers implementations ASIC comparisons.

Figure 14. FPGA Area/Performance comparison.

230
and RC5 are the alternative ciphers that can be used for bulk
encryption in WTLS.
For the FPGA implementation approach as it is shown
in figure 12, Full Rolling (FR) architectures of both Cobra-
H64 and Cobra-H128 has minimized area resources com-
pared with FPGA approaches of AES [13], RC4 [7], SAFER+
[9], IDEA [2], and DES [8] (for the last ciphers we present the
best results corresponding to different implementation archi-
tectures). The pipeline architectures (P)s of Cobra-H64 and
Cobra-H128, as it was expected needs more resources than
then (FR)s. The main advantage of pipeline architectures, as
it is proven from figure 12, is the very high speed perfor-
mance, compared with the achieved throughput of the other
conventional ciphers [2,79,13].
In addition, the proposed architectures are compared with
other ASIC implementations [19,2123] in figure 13. For the
ASIC approaches, a covered area comparison is not efficient
because of the variety in units (mm2, sqmil, gates, transistors)
of the reported area of the conventional architectures [19,21
23]. Both Cobra-H64 and Cobra-H128 higher performance
compared with AES [21], IDEA [23], RC5 [19], and DES
[22], in all of the cases.
Furthermore, for the FPGA implementations, all the
proposed Cobra ciphers architectures are compared in the
Area/Performance model. This model can easily be calcu-
lated, by the equivalent coved area resources (in CLBs) and
the throughput (Mbps), by using the formula: Area/Perfor-
mance model = Area/Performance. The Area/Performance
model comparison is illustrated in figure 14. It is obvious that
the FRs architectures of both Cobra-H64 and Cobra-H128 are
much better compared with RC4 [7], SAFER+ [9], IDEA [2],
and DES [8], while they are too close to AES [13] value.
Cobra-H64 (P) architecture as well as Cobra-H128 (P) Area-
Delay model have higher value than FRs architectures. This
is the paid penalty of the very good performance (figure 12)
compared with the other conventional ciphers.
7. Conclusions
In this work, we propose two new fast ciphers: Cobra-H64
and Cobra-H128. These ciphers are based on DDP transfor-
mations. Security analysis has shown that the both ciphers
are secure against known attacks. Due to high parallelism of
computations in one round and the use of the switchable op-
erations one can use very simple key scheduling that makes
hardware implementation cheaper and faster in the case of
frequent change of keys. Both ciphers achieve high-speed
performance in FPGA devices and especially for ASIC ap-
proaches. The implementation cost and the performance of
the proposed ciphers are compared with the security layers of
the most widely used wireless protocols: IEEE 802.11, Blue-
tooth, WAP, OMA, and UMTS. These comparisons present
the advantages of the proposed ciphers in terms of area re-
sources, operating frequency, and throughput. These advan-
tages prove the suitability of the proposed ciphers in wireless
communications, which set hard specifications in security im-
plementations.
SKLAVOS ET AL.
References
[1] V.E. Benes, Mathematical Theory of Connecting Networks and Tele-
phone Traffic (Academic Press, New York, 1965).
[2] O.Y.H. Cheung, K.H. Tsoi, P.H.W. Leong and M.P. Leong, Tradeoffs in
parallel and serial implementations of the international data encryption
algorithm, in: Proceedings of CHES 2001, Lecture Notes in Computer
Science, Vol. 2162 (Springer, 2001) pp. 333337.
[3] C. Clos, A study of nonblocking switching networks, Bell System
Technical J. 32 (1953) 406424.
[4] A.P. Fournaris, N. Sklavos and O. Koufopavlou, VLSI architecture and
FPGA implementation of ICE encryption algorithm, in: Proceedings of
10th IEEE International Conference on Electronics, Circuits and Sys-
tems (ICECS03), United Arab Emirates, December 1417 (2003).
[5] N.D. Goots, B.V. Izotov, A.A. Moldovyan and N.A. Moldovyan, Mod-
ern Cryptography: Protect Your Data with Fast Block Ciphers (A-LIST
Publishing, Wayne, 2003) 400 p.,
http://www.alistpublishing.com
[6] N.D. Goots, A.A. Moldovyan and N.A. Moldovyan, Fast encryption al-
gorithm SPECTR-H64, in: Proceedings of the International Workshop,
Methods, Models, and Architectures for Network Security 2001, Lec-
ture Notes in Computer Science, Vol. 2052 (Springer, 2001) pp. 275
286.
[7] P. Hamalainen, M. Hannikainen, T. Hamalainen and J. Saarinen, Hard-
ware implementation of the improved WEP and RC4 encryption al-
gorithms for wireless terminals, in: Proceedings of the European Sig-
nal Processing Conference (EUSIPCO2000), September 58, Finland
(2000).
[8] J. Kaps and C. Paar, Fast DES implementations for FPGAs and its ap-
plication to a universal key-search machine, in: Proceedings of 5th An-
nual Workshop on Selected Areas in Cryptography (SAC 98), August
1718, Canada (1998).
[9] P. Kitsos, N. Sklavos, K. Papadomanolakis and O. Koufopavlou, Hard-
ware implementation of the Bluetooth security, IEEE Pervasive Com-
puting, Mobile and Ubiquitous Systems 2(1) (2003) 2129.
[10] Y. Ko, D. Hong, S. Hong, S. Lee and J. Lim, Linear cryptanalysis on
SPECTR-H64 with higher order differential property, in: Proceedings
of International Workshop, Methods, Models, and Architectures for
Network Security 2003, Lecture Notes in Computer Science, Vol. 2776
(Springer, Berlin, 2003) pp. 298307.
[11] C. Lee, D. Hong, S. Lee, S. Lee, H. Yang and J. Lim, A chosen plain-
text linear attack on block cipher CIKS-1, Lecture Notes in Computer
Science, Vol. 2513 (Springer, 2002) pp. 456468.
[12] V.M. Maslovsky, A.A. Moldovyan and N.A. Moldovyan, A method of
the block encryption of discrete data, Russian patent # 2140710, Bull.
No. 30 (1999).
[13] M. McLoone and J.V. McCanny, High performance single-chip FPGA
Rijndael algorithm implementation, in: Proceedings of CHES 2001,
Lecture Notes in Computer Science, Vol. 2162 (Springer, 2001) pp. 65
76.
[14] A.A. Moldovyan, Fast block ciphers based on controlled permutations,
Computer Science Journal of Moldova 8(3) (2000) 270283.
[15] A.A. Moldovyan and N.A. Moldovyan, A method of the cryptograph-
ical transformation of binary data blocks, Russian patent # 2141729,
Bull. No. 32 (1999).
[16] A.A. Moldovyan and N.A. Moldovyan, A cipher based on data-
dependent permutations, Journal of Cryptology 15(1) (2002) 6172.
[17] M. Portz, A generalized description of DES-based and benes-based
permutation generators, in: Advances in Criptology, Lecture Notes in
Computer Science, Vol. 718 (Springer, 1992) pp. 397409.
[18] B. Preneel, A. Bosselaers, V. Rijmen, B. Van Rompay, L. Granboulan,
J. Stern, S. Murphy, M. Dichtl, P. Serf, E. Biham, O. Dunkelman,
V. Furman, F. Koeune, G. Piret, J.-J. Quisquater, L. Knudsen and
H. Raddum, Comments by the NESSIE project on the AES finalists,
May 24 (2000), http://www.nist.gav/aes
[19] A. Schubert and W. Anheier, Efficient VLSI implementation of modern
symmetric block ciphers, in: Proceedings of ICECS99, Cyprus (1999).
[20] B. Van Rompay, L.R. Knudsen and V. Rijmen, Differential cryptanaly-
HIGH SPEED NETWORKING SECURITY
sis of the ICE encryption algorithm, in: Proceedings of the 6th Interna-
tional Workshop Fast Software Encryption FSE98, Lecture Notes in
Computer Science, Vol. 1372 (Springer, 1998) pp. 270283.
[21] B. Weeks, M. Bean, T. Rozylowicz and C. Ficke, Hardware perfor-
mance simulations of round 2 advanced encryption standard algo-
rithms, in: Proceedings of 3rd Advanced Encryption Standard (AES)
Candidate Conference, New York, USA, April 1314 (2000).
[22] D.C. Wilcox, L.G. Pierson, P.J. Roberston, E.L. Witzke and K. Gass,
A DES ASIC suitable for network encryption at 10 Gbps and bey-
oned, in: Proceedings of CHES99, Lecture Notes in Computer Sci-
ence, Vol. 1717 (Springer, 1999) pp. 3748.
[23] R. Zimmermann, A. Curiger, H. Bonnenberg, H. Kaeslin, N. Felber
and W. Fichtner, A 177 Mb/s VLSI implementation of the international
data encryption algorithm, IEEE Journal of Solid State Circuits 29(3)
(1994) 303307.
Nicolas Sklavos is a Ph.D. Researcher with the Elec-
trical and Computer Engineering Department, of the
University of Patras, Greece. His interests include
computer security, new encryption algorithms de-
sign, wireless communications, and reconfigurable
computing. He holds an award for his Ph.D. research
on VLSI Designs of Wireless Communications Se-
curity Systems from IFIP VLSI SOC 2003. He is
a referee of International Journals and Conferences.
He is a member of the IEEE, the Technical Chamber
of Greece, and the Greek Electrical Engineering Society. He has authored or
coauthored up to 45 scientific articles in the areas of his research.
E-mail: nsklavos@ee.upatras.gr
231
Nikolay A. Moldovyan is an honoured inventor
of Russian Federation (2002), a chief researcher
with the Specialized Center of Program Systems
SPECTR, and a Professor with the Saint Pe-
tersburg Electrical Engineering University. His
research interests include computer security,
cryptography, and currently developed concept of the
variable transformations as a new direction in ap-
plied cryptography. He received his Diploma and
Ph.D. in Academy of Sciences of Moldova (1981).
He is a member of the IACR.
E-mail: nmold@cobra.ru
Odysseas Koufopavlou received the Diploma of
Electrical Engineering in 1983 and the Ph.D. degree
in electrical engineering in 1990, both from Univer-
sity of Patras, Greece. From 1990 to 1994 he was at
the IBM Thomas J. Watson Research Center, York-
town Heights, NY, USA. He is currently an Asso-
ciate Professor with the Department of Electrical and
Computer Engineering, University of Patras. His
research interests include VLSI, low power design,
VLSI crypto systems, and high performance commu-
nication subsystems architecture and implementation. Dr. Koufopavlou has
published more than 100 technical papers and received patents and inventions
in these areas.
E-mail: odysseas@ee.upatras.gr
 Mobile Networks and Applications 10, 233249, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Media Synchronization and QoS Packet Scheduling Algorithms

for Wireless Systems
AZZEDINE BOUKERCHE and HAROLD OWENS II
School of Information Technology and Engineering (SITE), University of Ottawa, Canada

Abstract. Wireless multimedia synchronization is concerned with distributed multimedia packets such as video, audio, text and graphics
being played-out onto the mobile clients via a base station (BS) that services the mobile client with the multimedia packets. Our focus is on
improving the Quality of Service (QoS) of the mobile clients on-time-arrival of distributed multimedia packets through network multimedia
synchronization. We describe a media synchronization scheme for wireless networks, and we investigate the multimedia packet scheduling
algorithms at the base station to accomplish our goal. In this paper, we extend the media synchronization algorithm by investigating four
packet scheduling algorithms: First-In-First-Out (FIFO), Highest-Priority-First (PQ), Weighted Fair-Queuing (WFQ) and Round-Robin
(RR). We analyze the effect of the four packet scheduling algorithms in terms of multimedia packet delivery time and the delay between
concurrent multimedia data streams. We show that the play-out of multimedia units on the mobile clients by the base station plays an
important role in enhancing the mobile clients quality of service in terms of intra-stream synchronization and inter-stream synchronization.
Our results show that the Round-Robin (RR) packet scheduling algorithm is, by far, the best of the four packet scheduling algorithms in terms
of mobile client buffer usage. We analyze the four packet scheduling algorithms and make a correlation between play-out of multimedia
packets, by the base station, onto the mobile clients and wireless network multimedia synchronization. We clarify the meaning of buffer
usage, buffer overflow, buffer underflow, message complexity and multimedia packet delay in terms of synchronization between distributed
multimedia servers, base stations and mobile clients.
Keywords: distributed algorithms, media synchronization, mobile multimedia, wireless communications, packet scheduling algorithm,
quality of service (QoS)

1. Introduction dia synchronization is much easier when compared to a wire-

Wireless communication technological advancements have
created a new paradigm known as mobile distributed multi-
media systems. In a mobile distributed multimedia system,
diverse packets can be simultaneously manipulated. Packets
such as text, images, audio and video can be played out on a
mobile client via a base station. Some multimedia packets are
time dependent upon each other [25]. Because of the time de-
pendency that exists between some packets, we can classify
these packets as discrete packets or continuous packets. The
discrete packets include text and images, while the continuous
packets have both audio and video.
The basic abstraction, for a time constrained media ele-
ment, is a timed stream of media components (video frame
or audio sample). Typically, during play-out on the mobile
client, media components must be kept in temporal order.
This ordering process is known as media synchronization.
There are two kinds of timing aspects for constrained ele-
ments: (1) intra-media continuity is subject to a real-time-
constraint in handling media packets and (2) inter-media con-
tinuity is subject to temporal correlation during playback
of media packets [25]. The synchronization problem, in
wired communication systems, has been extensively stud-
ied [4,22,29]. In a wired communication system, multime-
This work was partially supported by Research Grants from NSERC, the
Canada Research Program, Canada Foundation for Innovation, Ontario
Distinguished Researcher Award (OIT/ODRA#201722).
less environment. This is mainly because of a plethora of
resources like memory, power and bandwidth. In contrast, a
wireless systems is a bit more complicated because of a dearth
of resources (memory, power and bandwidth) that needs to be
managed efficiently. Due to the abundance of resources in
the wired system and limited resource in the wireless system,
communication between wired and wireless systems causes
a major communication problem solved by performing syn-
chronization between the two types of systems.
In a wireless communication system, a base station (BS)
must be used in order to deliver the multimedia packets to the
mobile client [8,9]. The resources between the base station
and the mobile client are limited in terms of bandwidth, mem-
ory and power. If the multimedia packets are received from
the wired systems to the mobile client too quickly (flooding)
or too slowly (starvation), the limited resources could be com-
promised. The memory at the mobile client could be over-
flowed causing lost multimedia packets to be retransmitted
or underflowed causing un-smooth video and audio play-out.
This potential compromise of the mobile clients resources, is
why we investigate wireless multimedia synchronizationand
QoS packet scheduling.
This paper considers transmission of live audio stream and
the corresponding video stream from the distributed multi-
media servers to the base stations which service the mobile
clients. We investigate the effect of QoS packet scheduling
algorithms at the base station on the overall network syn-
234
chronization. We investigate four packet scheduling algo-
rithms: First-In-First-Out (FIFO), Priority-Queuing (PQ), and
Round-Robin (RR) and Weighted Fair-Queuing (WFQ) queu-
ing. In todays wireless and wired communication systems
FIFO is used because of the simplicity of the algorithm [14].
PQ is an priority-based scheduling algorithm useful for the
transfer of real-time traffic but usually produces unfairness
among traffic classes [14]. RR scheduling creates fairness by
given each packet a certain amount of play-out time. WFQ
scheduling, at different network layers, gives fair sharing of
bandwidth among various network traffic.
2. Related work
Media synchronization control is paramount for multimedia
applications such as video and audio stream over wired/wire-
less networks. Media synchronization control is necessary
for preserving the temporal relationships among plural media
streams by compensating for network delay jitters [13]. Me-
dia synchronization can be classified as intra-stream synchro-
nization and inter-stream synchronization [26]. The intra-
stream synchronization refers to the temporal constraints
within a single stream such as the time intervals between two
successive packets of a video or audio stream. Inter-stream
synchronization pertains to synchronization among multiple
streams of concurrent packets such as voice and audio stream
(lip sync). To ensure quality of service of multimedia services
and to minimize end-to-end multimedia packet delay for the
mobile client, intra-stream synchronization and inter-stream
synchronization must be guaranteed as illustrated in figure 1.
Research has been done in the area of QoS packet schedul-
ing. Most of the research and publications deal with network
level performance like packet overhead, throughput and de-
lay. For overall QoS of the mobile client, several areas such
as scheduling for synchronization, feedback techniques for
Figure 1. Intra-stream and inter-stream synchronization.
BOUKERCHE AND OWENS
synchronization [24], network-based schemes [28], buffering-
based schemes [4], and reactive control schemes have been
studied over the past few years. Network synchronization
techniques have been used to improve the quality of service,
in terms of multimedia synchronization for smooth play-out
of audio and video, at the mobile client. In our earlier work,
we have discussed MoSync, an algorithm [8] is the only algo-
rithm that deals with continuous media in high layers, partic-
ularly audio and video at the application layer. The primary
difference between packet-level performance and continuous-
media level performance is at the continuous-media level, the
temporal structure of the media is the focus [8].
To achieve the overall goal of having a unified network
(wired/wireless), extensive research has been done on the
effect of packet scheduling synchronization in the network.
Earlier network designs (wireless and wired) were meant for
non-delay affected applications such as text and graphics.
The new paradigm for, both wired and wireless networks,
are resource intensive. Applications like animated graphics,
on-demand video and voice-over IP require bounded delays
and guarantee throughput, but tolerate some errors [19]. The
growth of smart phones, PDAs and laptop computers have
increase the need for unification between wired and wire-
less networks. However, like all technological advancement,
new technology invents new problems, particularly synchro-
nization with multimedia packets, being played out on the
mobile client. The combination of wired and wireless net-
works differ greatly from the typical wired networks. First, all
packets must pass through a base station before reaching its
mobile terminal destination [8]. Second, a mobile clients re-
sources (battery, memory, screen, bandwidth, etc.) are very
limited. Due to resource limitations, conventional synchro-
nization strategies for delivery of multimedia packets to the
mobile client, cannot be applied in a mixed environment.
A single base station buffers multimedia packets for multiple
mobile clients and must service each mobile client in its cell
area. Packet buffering can cause congestion at the base station
and slow the processing of multimedia packets at the base sta-
tion. On the other hand, a burst of multimedia packets, from
the base station to a limited resource (battery, memory, screen,
bandwidth, etc.) mobile client, can cause buffer overflow at
the mobile client. If multimedia packets do not arrive on time
at the mobile client from the base station, buffer underflow at
the mobile client can occur. The base stations assume that the
packets will be available when multimedia packets requests
are made from the mobile client. The Internet only provides
best-effort and cannot guarantee on-time multimedia packet
delivery to the base station. The mobile client must inform
the multimedia server and base station about its play-out con-
ditions. The combination of wired and wireless network com-
munication brings about a new synchronization for multime-
dia applications methodology.
In earlier studies, the MoSync algorithm [8], upon which
we expounded, implemented a feedback-based synchroniza-
tion for multimedia applications. MoSync is a synchroniza-
tion scheme for wireless clients and distributed
multimedia systems that uses a Quasi-sink to control synchro-
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING
nization. The proposed solution copes with network jitters,
end-system jitters, clock drift and changing network condi-
tions [4,20]. MoSync can be employed for both intra-stream
during playback video and resynchronization streams in a
wireless network. Since MoSync uses delay time to allow
intra-synchronization, inter-synchronization is guaranteed by
the MoSync algorithm [8]. A set of experiments was done
earlier assuming FIFO play-out of the multimedia packets on
the mobile client from the base station. In this paper, we study
the hand-off problem and propose tow hand-off schemes to
enhance the performance of the media synchronization for
wireless systems, as well as investigate the multimedia packet
play-out onto the mobile client in RR, PQ, and WFQ from the
base station.
3. MoSync algorithm
The MoSync algorithm [8] assumes that the network can pro-
vide sufficient resources to deliver multimedia packets to the
mobile client. The multimedia data consists of many mul-
timedia packets that may be progressively transmitted over
the network. MoSync does not use a global synchronization
clock. In the MoSync algorithm, the servers time stamp each
multimedia packet, with current local time to allow the base
station to calculate round trip delay, jitters and inter-arrival
time. The multimedia packets carries the servers number and
packet sequence number but these additions are negligible so
multimedia synchronization is the primary focus of the this
paper.
The MoSync model consists of K scalable server nodes
and L mobile clients. A base station can communicate with
multiple mobile clients because the communication between
the base station and mobile client is wireless. Figure 2 illus-
trates the servers, base station and mobile client connectivity.
The system can contain multiple servers (data, voice, graph-
ics, video) depending upon the need of the mobile client. All
mobile clients receives the multimedia services from the same
servers. The communicate is done from the the servers to the
mobile client via the base station. The BS communicates with
many servers and mobile client at any given time. The BS in
the MoSync algorithm have special roles when servicing mo-
bile client in their area. The base station has three roles: as a
messenger it passes multimedia units to the mobile client, as
a filter it sends request to the servers per request of the mobile
Figure 2.
235
client and last, as a Quasi-receiver, it only receives the first
packets from the servers. When a mobile client request mul-
timedia packets from a server the multimedia packets are sent
via the base station. The servers can control the supply rate
of multimedia packets to mobile clients by using the mobile
clients feedback messages. MoSync focuses on three areas to
solve the combination of wired and wireless communication
synchronization problems:
1. No jitters and constant delay case (extending Biersack and
Geyer [4]).
2. Intra-stream and inter-stream synchronization problem
(using network load).
3. Resynchronization (using exponential smoothing forecast-
ing).
MoSync uses three simple protocols. In case one, to avoid
buffering of early arrival multimedia packets sub-stream the
start-up protocol is used. This allows evaluation of the round
trip delay for different sub-streams. In case two, end-system
jitters and network jitters is counteracted. Both end-system
and network jitters must be countered to avoid underflow or
overflow at the mobile client. For multimedia packets to play-
out smoothly the servers data transmission rate must match
the available service rate at the mobile client. The overall
goal when dealing with network and end-system jitters, is to
keep the average delay, buffer underflow and buffer overflow
at an acceptable minimal level. In case three, to solve the
average delay changes, clock drift and server dropouts expo-
nential smoothing forecasting is used [8]. When a mobile
client receives a multimedia packet, the mobile client calcu-
lates the expected arrive time of the next multimedia packet
and sends the calculation to the base station. The base station
then forwards the information to the servers. In both case two
and three the MoSync algorithm uses either the pessimistic or
optimistic synchronization protocol when requesting and re-
ceiving multimedia packets from the servers. With MoSyncs
pessimistic protocol, the base station updates the servers on
each multimedia packet request made by the mobile client.
With MoSyncs optimistic protocol, the server sends the mul-
timedia packets directly to the mobile client without contin-
uous updates being made by the base station once the first
multimedia packet from the server arrives to the mobile client.
The MoSync algorithm operates at the networks application
layer.
3.1. Basic concepts of the MoSync algorithm
The MoSync algorithm uses three types of nodes: server,
Quasi-receivers and receivers [8]. The algorithm contains
K servers, N base stations and L mobile clients, where K
servers are of type servers, M base stations are of type Quasi-
receiver and the L mobile clients are of type receivers. The
mobile client synchronizes the multimedia packets, received
from the multimedia server, by reporting buffer usage to
the base station and updating the base station with multime-
dia packet arrival time differences. After a base station has
236
information about the arrival time of each packet, it calcu-
lates the synchronization time for the next packet. A sched-
uler, in the server, manages the on-time transfer of the
subframes, as a part of the frames to the mobile client. The
mobile client requests multimedia service from the multi-
media servers, through the base station. When a base sta-
tion requests the first group of multimedia packets, it sends
the synchronization point information to all servers [8]. The
base station can receive messages from the mobile client and
from neighboring base stations. The mobile client messages
are for requesting multimedia packets from servers and done
messages to let base stations know all packets have been re-
ceived by the mobile client. The neighboring base station
messages are used to inform the base station if its neighbor
is on or off, to set-up hand-offs to neighboring base stations.
It is the mobile clients job to calculate the latest arrival time
and the differences between every multimedia packets arrival
time, the buffer usage and update its serving BS with the new
data.
3.2. MoSync algorithm and hand-off management schemes
A wireless network consist of a mobile client, a BS that ser-
vices the mobile client and MSC that performs the hand-off
between base stations. The BS is connected to a wired server
network that provides the mobile client with services, such as
data, graphics, audio and video, via the BS, through a wire-
less link. Each BS has a coverage area known as a cell. When
a mobile client, receiving services from a BS, moves out of
the coverage area of the servicing BS and into the cell of an-
other BS, the mobile clients services must be transfer to the
new BS to ensure continuous service. Hand-offs in a wire-
less network are classified as Soft Hand-offs and Hard Hand-
offs.
In a hard hand-off, the connection between the mobile
client and servicing BS, is broken before new services are es-
tablished with the new cells BS. The mobile client can only
be connected to one BS at any time, with a hard hand-off,
because frequencies between base stations must be different
to avoid interference between mobile client signals. Because
a mobile client can only be connected to one BS at a time,
during hand-off, delays between services are introduce into
the network. Hard hand-off is mostly used in FDMA and
TDMA [1].
In a soft hand-off, the connection between the servicing
BS and the mobile client is maintained while another con-
nection between the mobile client and the new cells BS is
established. The mobile client can be connected to one or
more base stations at any time, in a soft hand-off and con-
current connections between mobile clients and base stations
can exist because the frequency band at the base stations dif-
fers. Soft hand-off is mostly used in CDMA [1]. In a wireless
network, soft hand-off is the preferred method of hand-off be-
cause the mobile clients services are not interrupted when a
hand-off is performed. Because there is no interruption in ser-
vice, using soft-hand off, audio and video is played-out on the
mobile client smoothly.
BOUKERCHE AND OWENS
We have investigated several hand-off schemes for Mo-
Sync [10]. In our experiments, we settle with the two-hand
of scheme, which consists of two parts: setup hand-off and
end hand-off. In phase one, setup hand-off updates new ar-
rival base stations and maintains synchronization for newly
arrived mobile clients. If a mobile client can communicate
with another BS, the mobile client will send a new BS ar-
rived message to its primary BS. For the mobile client to
request service from the new BS, the new BS must have delay
time information about each multimedia server. If the new
base station does not have delay time information about the
multimedia servers, then the new BS sends a Request mes-
sage to all multimedia servers. When a mobile client receives
multimedia packets, the mobile client calculates the latest ar-
rival time and the differences between multimedia packets us-
ing the MoSync algorithm. In phase two, the end hand-off
deals with the ordering of multimedia packets and the flow
of the multimedia packets to the mobile client. Any BS can
serve as the newly selected primary BS. After informing the
mobile client, BS and multimedia servers, the end hand-off
phase selects the closest common node from the primary BS
and the newly selected BS. The common node must be within
the wireless network. If there is no common node, then the
MSC of the current primary BS will be the common node un-
til a common node can be found. The common node reroutes
the multimedia packets in time-stamped order using the newly
selected primary BS. Once the mobile client moves within
the coverage area of the selected BS, the end hand-off phase
is terminated. The MoSync algorithm works with both hard
hand-off and soft hand-off to synchronize multimedia packet
data flows [9].
4. Qos packet scheduling algorithms
In a wireless network, the mobile client requests multimedia
services through a base station. The base station forwards
the multimedia packet request to the distributed multimedia
server. The multimedia server sends the multimedia packet to
the base station where the packets are buffered and sched-
uled for play-out onto the mobile client that requested the
multimedia service. With todays wireless technology, the
mobile client can request video, audio, text and graphics ser-
vices from distributed multimedia servers via the base station.
The base station cannot play-out the multimedia packets onto
the mobile client until the latest packet has arrived. If an on-
demand-video movie is requested by the mobile client, the
base station must receive both audio and video packets for a
particular frame from the distributed multimedia servers be-
fore the frame can be played-out onto the mobile client. The
multi-data stream dependence between multimedia packets,
such as video and audio, is known as inter-stream synchro-
nization or lip-synchronization. Each audio and video mul-
timedia packet must be kept in temporal order for smooth
play-out audio and video at mobile client. The time de-
pendency of multimedia packets of the same data stream is
known as intra-stream synchronization. The mobile client
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING
calculates the arrive of the multimedia packets and informs
the base station of the arrival time difference. The updated
messages from the mobile client to the base station allows
the base station to synchronize its request to the multime-
dia server and synchronizes the multimedia packets play-out
onto the mobile client. In the MoSync algorithm the base sta-
tion schedules play-out of multimedia packets on the mobile
client in a FIFO order. The resources at the mobile client
are limited in terms of power, bandwidth and memory. The
object of multimedia synchronization is to keep the buffer
at the mobile client full so smooth play-out of multimedia
services can be presented. If the multimedia packets from
the servers are scheduled too early, buffer overflow occurs
at the mobile client. If multimedia packets from the servers
are scheduled after the deadline to maintain temporal order,
buffer underflow occurs at the mobile client. Mobile client
buffer underflow and overflow is a network problem that has
to be dealt with through synchronization of packet delivery,
to the mobile client from the servers, via the base station.
The scheduler in the base station effects network synchro-
nization in terms of buffer usage, overflow, underflow, mes-
sage complexity and multimedia packet delay at the mobile
client.
In this paper, we wish to investigate how packet schedul-
ing algorithms may affect the media synchronzation in wire-
less and mobile communication systems. We have investigate
several packet scheduling schemes, In this paper, we settle
down with the following QoS packet scheduling algorithms:
FIFO, PQ, WFQ and RR.
4.1. First-In-First-Out (FIFO)
In todays wired and wireless systems, First-In-First-Out
(FIFO) is the most common scheduling algorithm. There is
only one queue and all packets are treated equally. Packets
are stored at a buffering location and processed on a first come
first serve basis. FIFO does not recognize priority or classes
of traffic. If the buffering area does not have space to store
arriving packets, the buffer area discards the packets [14]. In
terms of data flow, since large flows may arrive first and cap-
ture large bandwidth, FIFO is a since of unfairness among
flows. On the average, a flow burst received is usually trans-
mitted in a similar burst using FIFO scheduling. FIFO is the
fastest queuing method. It is effective for large links that have
little delay and minimal congestion.
4.2. Priority Queuing (PQ)
Some multimedia requests, like on-demand-video, requires
large bandwidth and guaranteed processing order of multime-
dia packets. In this special case, Priority Queuing (PQ) can
be used to guarantee the bandwidth, memory and process-
ing time for a particular high priority data flow. PQ sets a
certain number of priority classes according to some parame-
ter. In our experiment, the packet with the lowest play-out
time receives higher priority. In a wired and wireless com-
munication system the priority can be an IP address, proto-
col or interface [14]. The packets are played out according
237
to the highest priority. PQ serves the highest priority pack-
ets first and then moves to the lowest after all higher packets
have been processed. Priority-based algorithms reduce delays
for high priority classes that are bandwidth sensitive, such as
video, audio and interactive traffic. PQ reduce delays for high
priority classes with out reducing the overall throughput of
the lower classes. PQ can lead to unfairness among classes
if not monitored closely because higher priority classes that
arrives on a continuous basis will always be processed before
the lower priority classes are processed. To reduce starvation
among lower priority classes, PQ monitors the traffic and will
update the lower priority classes priority to enable process-
ing of lower priority classes during continuous high priority
classes traffic flows.
4.3. Round Robin (RR)
Lately, there has been an increase in multimedia traffic such
as audio and video. The heavy demand for multimedia traffic
increases QoS needs to enhance the multimedia communica-
tion. Packet scheduling in the networks play an important part
in the QoS of each client in the network. These multimedia
services are bursty and bandwidth hungry. To ensure fairness
between processing of packets in the network, Round Robin
(RR) packet scheduling can be used. With RR packet schedul-
ing, arriving packets are store in the same buffer on a First-In-
First-Out (FIFO) strategy. The packets are removed from the
buffer in a FIFO basis. Each packet is guaranteed a certain
quantum of processing time. If the packet finish playing-out
during the allowable processing time, the packet is remove
from the queue. If the packet does not finish processing dur-
ing the allowable time, the packet is stamp with information
such as arrival time and burst time and is place back into the
buffer as a new arriving packet to be processed. The Round
Robin algorithm creates fairness among packets by guaran-
teeing play-out time and by isolating each traffic flow from
misbehaving traffic. Bursty packet like audio and video that
arrives for a particular flow are spread out and transmitted one
at a time according to the RR packet scheduling algorithm.
RR does not easily adjust to the network load due in part to
the quantum slice that is static.
4.4. Weighted fair queuing (WFQ)
Multiple traffic flows can be identical in priority and their net-
work resource usage. With the multimedia traffic of today,
assigning a priority or guaranteeing play-out time of multi-
media packets is not enough to ensure high QoS for a client.
The Weighted Fair Queuing (WFQ) algorithm combines the
PQ and RR into a dynamic fair queuing that divides the avail-
able bandwidth among traffic queues base on weight or pri-
ority. In todays networks, video packets are given a higher
priority than audio packets, audio packets are given a higher
priority than graphic packets and graphic packets are given a
higher priority than text during play-out of multimedia pack-
ets. WFQ ensures that each packet is treated fairly according
to its weight. Packets of the same priority is stored in the same
238
buffer. Multimedia packets of the same priority are played-
out in a round robin order. Hence, satisfactory response time
is a result of WFQ fairness. For delay sensitive applications
such as video-on-demand, animated graphics and interactive
applications, WFQ is the most appropriate scheduling algo-
rithm. Furthermore, WFQ dynamically adjusts to traffic flow.
Low volume and high volume traffic both get fair allocation
of available bandwidth. WFQ may not perform well in a wire-
less environment where the resources such as memory, power
and bandwidth are limited.
5. Simulation experiments
We have developed a discrete-event model to simulate a cellu-
lar wireless multimedia system, on a wireless and wired net-
work. Our assumption is that the communication between
servers and base stations are wired and connections between
mobile clients and base stations are wireless. In our model,
there are 300 channels available for 60 cells, with the system
load equally distributed over all cells. Each mobile client has
a buffer size thats at least 3 times that of the largest multi-
media packets and at most six time as large as the multimedia
packets received from the server. In this simulation, we em-
ploy a two-phase hand-off scheme using soft-hand hand-off
in CDMA [5].
We have performed several tests, using many jitters and re-
synchronization. We evaluate MoSyncs performance in two
environments: uniform and nonuniform jitters using FIFO,
PQ, WFQ and RR scheduling. Table 1 describes the parame-
ters we use in our simulation experiments. In the uniform
case, for the four packet-scheduling algorithms, we evalu-
ate each algorithm, assuming uniform multimedia unit de-
lay, under jitters. All cells have the same multimedia packets
demand and the requests from the mobile client have inter-
arrival time and average service time . We distribute the
delay evenly among all multimedia requests, maximum and
minimum delay time, assuming minimal delay time of 50 ms.
Due to the distribution of the delays, the jitters will cause dif-
ferent effects on the synchronization algorithm.
In the nonuniform case, we have assumed exponential dis-
tribution of network delay jitters. This distribution was used
in earlier studies of the MoSync algorithm [8]. For our exper-
iments, there are multimedia media delay ranges as follows:
(1) 0200 msec; (2) 0400 msec; (3) 0600 msec. The ranges
Table 1
Simulation parameters.
Packet scheduling algorithms FIFO, PQ, WFQ, RR
Number of cells 60
Number of multimedia servers 4 dia units that overflow a mobile client buffer.
Mobile client buffer size 30 times of a multimedia packets
Play-out time/multimedia packets 100 msec
Mean service time/session
Arrival rate in normal cell
Round-Trip-Time (RTT) to
request/deliver a multimedia packets 50 msec
Jitters (uniform) 020, 2040, 4060, 6080 msec
Jitters (non-uniform) 0200, 0400, 0600 msec
BOUKERCHE AND OWENS
are exponentially distributed, with mean values of 20, 40 and
60, respectively. The upper bound delay is set in the nonuni-
form distribution to prevent failure, caused by large delays,
in the stimulation. The mean communication session is set to
20 multimedia packets and the mean buffer size of a mobile
client is set to three times the size of the multimedia pack-
ets [8]. Our results will that each range of delay jitters cause
the network to behave differently.
5.1. Method of experiment
The BS plays-out the multimedia packets, sent by the mul-
timedia servers, onto the mobile client that requested the
packets. To avoid mobile client buffer overflow and buffer
underflow and to ensure intra-stream and inter-stream syn-
chronization of the multimedia packets, multimedia packets
request and delivery must be synchronized with the mobile
client that requests the multimedia packets.
The multimedia servers, which send the multimedia pack-
ets request, have a built-in scheduler that delivers the pack-
ets on time. The base station receives and buffers multime-
dia packets from the multimedia servers and schedules the
packets for play-out onto the mobile client that requested the
multimedia packets. In earlier studies, using the MoSync
algorithm, the base station plays-out the multimedia pack-
ets onto the mobile clients in a FIFO order. We will make
a correlation between the scheduling of multimedia pack-
ets by the BS, to be played-out onto the mobile client and
the effect the multimedia packets scheduling, by the BS, has
on intra-stream synchronization, inter-stream synchronization
and mobile clients buffer usage. We assume that the wired
network can provide efficient data flow from the multimedia
servers, to the mobile client, per request of multimedia pack-
ets.
5.2. Performance metrics
To assess the performance of each of the scheduling algo-
rithms, we choose the same performance metrics as in earlier
work [8,9]. Our areas of focus are as follows:
Message complexity: measures the overhead in terms of
the number of messages needed to satisfy the users mul-
timedia requests.
Buffer usage: measures the synchronization behavior for
each mobile client.
Overflow rate: measures the average number of multime-
Underflow rate: measures the average deficiency in the
clients buffer of the number of multimedia units currently
needed for smooth play-out.
Multimedia unit arrival rate: measures the reaction
of the network and the network condition by moni-
toring multimedia packets arrival time to client (intra-
synchronization).
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 239

Delay between multiple data streams: measures the delay

between the data streams that are serving the client con-
currently (inter-synchronization).

6. Simulation results

In this section, we analyze the results from the simulation ex-

periments we have obtained using FIFO, PQ, WFQ and RR
packet scheduling. We simulate MoSync with both intra-
synchronization and inter-synchronization. We use a control
mechanism to control buffer availability and reduce the num-
ber of messages between mobile client and BS. Because ar-
rival time is highly dependent on the network queue delay,
distributed clock drifts and violation of network bandwidth
guarantee, it is easier to delay arriving multimedia packets,
using this control mechanism, than to speed up the arrival
time of the packets.

6.1. Buffer usage

Figures 3 and 4 illustrate the buffer usage for uniform and

nonuniform cases. In figure 3, at 20 ms uniform delay jit-
ters, FIFO, PQ, WFQ and RR scheduling algorithms perform
equally. To play-out smoothly, all of the four algorithms re-
quire a buffer size sixteen times that of the multimedia pack-
ets. The four scheduling algorithms perform equally because
the network compensates for the small jitters by shifting the
delay to the upper layers, where the multimedia packets are
buffered and propagated with little delay between multime-
dia packets. At 40 ms uniform delay jitters, all algorithms
allow smooth play-out at fifty requests, but FIFO requires
about twice the buffer space to play-out smoothly. FIFO re-
quires twice the buffer space because the rate of multime-
dia packet arrival is almost proportional to the departure rate.
At 60 ms uniform delay jitters, all packet scheduling algo-
rithms synchronize equally requiring no more buffer space or
multimedia packets to play out smoothly. With 60 ms uni-
form jitters, the mobile clients and distributed multimedia
servers receive and send rates are matched. MoSync effi-
ciently buffers the data at the application layer and allows for
smooth play-out of multimedia packets by the BS onto the
client. At 6080 msec delay jitters in figure 4, FIFO needs
twice the number of multimedia packets to play out smoothly
compared to RR, PQ and WFQ. FIFO is really effected by the
large delays in the network whereas PQ, WFQ and RR are
affected because synchronization is easier and traffic is more
predictable, so RR, WFQ and PQ the MoSync algorithm can
synchronize accordingly. Notice up to 20 multimedia pack-
ets request with 6080 ms delay jitters all algorithm behaves
the same because for small numbers of multimedia packet re-
quest MoSync is able to buffer packets and play-out packets
smoothly onto the mobile client. After 20 multimedia pack-
ets request FIFO is affected by the non-predictability of the
arriving multimedia packets. With WFQ, PQ and RR algo-
rithm we know what multimedia packets will be serviced next Figure 3. Buffer usage for two-phase algorithm: uniform case.
by the BS. Figure 4 illustrates the nonuniform buffer usage.
240 BOUKERCHE AND OWENS

At 20 ms nonuniform delay jitters, all packet-scheduling al-

gorithms synchronize equally, requiring the same amount of
buffer space to play-out smoothly. With small amounts of
nonuniform network jitters, the network can easily counter
these delays by pushing delay to the upper most layers where
multimedia packets are buffered by the BS and played-out
onto the mobile client. At 40 ms nonuniform delay jitters the
packet scheduling algorithms all decrease and need more mul-
timedia packets to play-out smoothly. WFQ and PQ decrease
the slowest and are not affected by the large delays in the
network, while both RR and FIFO need about twice the num-
ber of multimedia packets to play-out smoothly. WFQ and
PQ reduce delay for a high-priority class that is both bursty
and low rate, relative to link bandwidth (video stream, inter-
active traffic, etc.), without reducing the throughput of lower
priority classes [12]. FIFO packet scheduling does not iso-
late flows from misbehaving or bandwidth heavy flows. Mis-
behaving or heavy bandwidth flows such as video can cap-
ture the limited wireless bandwidth increasing delays of other
lower bandwidth multimedia packets that need to be played
out. Finally, when we introduce 60 ms nonuniform delay jit-
ters, the scheduling algorithms all decrease quickly needing
more multimedia packets to play smoothly. PQ and WFQ
need 15 less multimedia packets than FIFO and RR to play-
out smoothly because under any delay, WFQ and PQ will
guarantee play-out of heavy bandwidth multimedia packets,
where RR and FIFO cannot because no priority is given to a
particular packet flow.

6.2. Overflow rate

Figures 5 and 6 illustrates the overflow rate for both the

nonuniform and uniform cases. At the networks application
layer multimedia packets are buffered and handle as continu-
ous data stream to allow for play-out of multimedia packets
without jitters. The object is not to overflow the mobile clients
buffer, hence, causing data in buffers to be rendered useless.
All packet scheduling algorithms perform well under the dif-
ferent ranges of network delay jitters. It is much easier to
control overflow then to control underflow. We can introduce
a delay great enough to eliminate all overflow at the mobile
client, to the maximum 150 multimedia packets request as
illustrated in figures 5 and 6. This works in theory, but in ap-
plication, will under-utilize the networks resources and not
allow for smooth play-out of multimedia packets. The Mo-
Sync algorithm is very efficient in preventing overflow. The
MoSync algorithms ability to adjust to network delay jitters
and prevent mobile clients buffer overflow is illustrated in
figures 5 and 6.

6.3. Underflow rate

Figures 7 and 8 illustrate the underflow rate for both the

nonuniform and uniform cases. In figure 7, at 20 ms nonuni-
form delay jitters, all algorithms perform well and produce no
Figure 4. Buffer usage for two-phase algorithm: nonuniform case. underflow because the network shifts the small delays to the
application layer where the multimedia packets are buffered
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 241

Figure 5. Overflow for two-phase algorithm: nonuniform case. Figure 6. Overflow for two-phase algorithm: uniform case.
242 BOUKERCHE AND OWENS

Figure 7. Underflow for two-phase algorithm: nonuniform case. Figure 8. Underflow for two-phase algorithm: uniform case.
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 243

at the BS and then played-out on the mobile client, without

jitters. In figure 7, with nonuniform delay jitters of 40 msec
distributions, WFQ and PQ produce a great deal of underflow
at higher number of requests, while FIFO and RR remain al-
most constant. At 60 msec distributed nonuniform delay jit-
ters, PQ and WFQ produce a lot of underflow at the mobile
client, while FIFO and RR remain almost constant because a
larger number of mobile clients are able to play-out their mul-
timedia packets. With WFQ and PQ, with a particular data
flow, only the mobile clients with higher priority requests are
able to play their multimedia packets out on-time. At greater
delays, the RR scheduling produces little or no, caused by
the ability of RR spread out a burst of multimedia packet re-
quest among flows, fairly, allowing for more mobile clients to
play-out the multimedia packets smoothly. In figure 8, uni-
form delays jitters of 20, 40, 60 and 80 msec are introduce
into the network. Figure 8 illustrates 80 ms delay jitters, RR
and FIFO produce little or no underflow while PQ and WFQ
produces a lot because with PQ we service the packets with
the smallest play-out time causing clients to wait on multi-
media packets until all higher priority multimedia packets are
played out. We can conclude that the RR packet scheduling is
the best among the four packet scheduling algorithms. WFQ
is used to ensure fairness among the various data streams but
does not easily adjust to traffic flow. We can notice little un-
derflow for the RR and FIFO packet scheduling for all uni-
form and nonuniform delay jitters. This implies that RR and
FIFO algorithms provide better intra-stream and inter-stream
synchronization. RR and FIFO based scheduling algorithms
isolate flows from one another. With RR, packets for a par-
ticular flow are spread out and serviced one at a time. With
FIFO, packets for a particular flow received are transmitted in
a similar burst. RR protects multimedia data flows from large
bandwidth or misbehaving data flows by guaranteeing each
flow a certain amount of time to play-out.

6.4. Message complexity

Figures 9 and 10 illustrate the message complexity of each

packet-scheduling algorithm. The message complexity is the
message overhead caused by the mobile client requesting
multimedia packets from the multimedia servers. The mes-
sage complexity of MoSync is O(I (K + L)) where L is re-
questing messages from mobile client, K is the number of
servers processing request and I is the number of rounds of
L + K messages. In the beginning round, the mobile clients
send a request message to the BS. The BS, upon receiving
the request message from the mobile client, forwards the re-
quest message to each K server. The servers reply with a K
dummy message to the forwarding BS. The BS, upon receiv-
ing the reply message from the K server, sends a ready mes-
sage to each mobile client. No multimedia packets are sent
during the beginning round. In later rounds, there are at most
L request messages from the mobile client, K messages from
the base station and L reply messages from the multimedia
servers. There are I rounds which imply that the total num- Figure 9. Message complexity for two-phase algorithm: nonuniform case.
ber of messages per request is I (2L + K) = O(I (K + L)). In
244 BOUKERCHE AND OWENS

the nonuniform case, at 20 ms, 40 ms and 60 ms delay jitters,

the message overhead per request is between 8 and 9 mes-
sages for small numbers of requests and large numbers of re-
quests. The four packet scheduling algorithms have no effect
on the number of messages per multimedia packet requests in
the nonuniform case. The message complexity, in this case,
implies that there is little or no buffer overflow at the mobile
clients. The number of messages per request increase when
buffer overflow at the mobile client occurs. All the packets in
the mobile clients buffer must be requested again, increasing
the rounds I, thus increasing the number of overhead mes-
sages. In the uniform case at 20 ms, 40 ms, 60 ms and 80 ms,
all of the four packet scheduling algorithms message com-
plexity is between 8 and 9 messages for small numbers of
requests and large numbers of requests. The message com-
plexity in this case implies little or no buffer over flow at the
mobile client. In cases where buffer overflow does not occur,
the mobile client need not request multimedia packets to re-
place lost multimedia packets caused by buffer overflow at the
mobile client. Figures 9 and 10 illustrate the MoSync algo-
rithms low message complexity rate and show that the play-
out of multimedia packets at the BS onto the mobile client
have little effect upon the MoSync algorithms message com-
plexity rate.

6.5. Multimedia unit arrival rate

Figures 11 and 12 illustrate the multimedia packets arrival

Figure 10. Message complexity for two-phase algorithm: uniform case.
rate (intra-stream synchronization) for both the nonuniform
and uniform cases. The multimedia packets arrival rate tells
us much about the condition of the network in terms of the
delay between multimedia packets delivered to the mobile
client. The object is to have a constant rate of multime-
dia packets, from the servers, that match the mobile clients
processing rate of multimedia packets. The MoSync algo-
rithm maintains intra-synchronization when the servers send
rate matches the mobile client play-out (processing rate) and
there is no buffer underflow or buffer overflow at the mobile
client. At 20 ms uniform delay jitters, the four packet schedul-
ing algorithms all ensure intra-stream synchronization. At
most the delay is 6 ms between successive multimedia pack-
ets. To play-out audio stream smoothly at the mobile client re-
quires intra-stream synchronization within 11 ms, for tightly
coupled audio [21]. At 40 ms uniform delay jitters, the over-
all delay increases in terms of arrive time because of the in-
creased delays introduced into the network. Intra-stream syn-
chronization remains tight in the 40 ms uniform case. For
large numbers of multimedia packet requests, the intra-stream
synchronization becomes erratic because no feedback mecha-
nism is used for large numbers of multimedia packet requests.
The multimedia servers send the multimedia packets directly
to the mobile client via the BS for large numbers of multime-
dia packet requests. At 60 and 80 ms uniform delay jitters,
intra-stream synchronization is guaranteed up to 100 multi-
media packet requests. In the nonuniform case, at 20 ms de-
lay jitters, the intra-stream synchronization is guaranteed up
to 100 multimedia requests. The four packet scheduling algo-
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 245

Figure 12. Multimedia units arrival for two-phase algorithm: nonuniform

Figure 11. Multimedia units arrival for two-phase algorithm: uniform case.
case.
246 BOUKERCHE AND OWENS

rithms guaranteed intra-stream synchronization at 60 ms and

80 ms uniform delay jitters. In the nonuniform case, a feed-
back mechanism is used to calculate the expected arrival time
of the next multimedia packet for smaller numbers of mul-
timedia packet requests. The next packet arrival calculation
allows the BS to schedule multimedia packet for smooth play-
out onto the mobile client. In the 40 ms and 60 ms nonuni-
form cases, we can see that the calculations of the next multi-
media packet arrival time allows for guaranteed intra-stream
synchronization up to 100 multimedia packet requests. The
on-time arrival of packets allow the BS to play-out the multi-
media packets onto the mobile client smoothly. At large num-
bers of requests, no feedback mechanism is used and mul-
timedia servers send the multimedia packets directly to the
mobile client, without continuous updates being made by the
mobile client to the servers via the base station. The MoSync
algorithm operates at the networks application layer. The ap-
plication layer absorbs network delay jitters by buffering data
and playing out multimedia packets as a stream of data onto
the mobile client. The MoSync algorithm maintains intra-
stream synchronization using WFQ, RR, PQ and FIFO queu-
ing during play-out of multimedia packets onto mobile client,
by the BS. We can conclude that the order in which the mul-
timedia packets are played-out onto the client has little effect
on intra-stream synchronization because all of the four packet
scheduling algorithms perform intra-stream synchronization
equally.

6.6. Delay between multiple data streams

Figures 13 and 14 illustrate the delay between multiple data

streams servicing the same mobile client concurrently (inter-
stream synchronization) for both the nonuniform and uni-
form case. In figure 13, under uniform delay jitters, the
MoSync feedback mechanism works very well. The four
packet scheduling algorithms average about 5 ms between
the data, text, audio and video data streams, servicing the
mobile clients. Notice, after 80, the number of requests for
the 4 algorithms increases in delay between data streams be-
cause no feedback mechanism is used for large numbers of
requests. For large numbers of multimedia packet requests,
the server sends the multimedia packets directly to the mo-
bile client to avoid feedback messages congestion at the BS.
The ability for the concurrent data streams to remain within
5 ms of packet delivery illustrates and proves that the Mo-
Sync algorithm guarantees inter-stream synchronization. In
figure 14, we use nonuniform delay jitters. At 20 ms delay jit-
ters, the four algorithms all performs extremely well because
small delays are compensated by shifting the small delay to
the higher layers of the network where they are buffered and
played-out smoothly. After 80 requests, the delay becomes
erratic because, for large numbers of multimedia packet re-
quests no feedback mechanism is used but still remains in the
bounds for smooth multimedia packet play-out onto the mo-
bile client. Under nonuniform 40 ms delay jitters, RR out
performs PQ, WFQ and FIFO. RR out performs PQ, WFQ Figure 13. Delay between data streams for two-phase algorithm: uniform
and FIFO inter-stream synchronization because RR has the case.
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING
Figure 14. Delay between data streams for two-phase algorithm: nonuniform
case.
247
ability to isolate flows by allowing all flows an equal amount
of time to play-out. Under large network delay conditions,
the network is unable to compensate for the large delay jitters
using FIFO, WFQ and PQ because only the flows with higher
priority are played-out resulting in very large delays for lower
priority flows. FIFO, WFQ, PQ does not isolate higher band-
width usage flows such as video from lower bandwidth usage
data flows such as text. When higher bandwidth usage flow
over utilize the bandwidth, lower concurrent bandwidth us-
age flows such as text, graphics and audio arrives to the mo-
bile client late, hence inter-synchronization is not maintained
and the quality of service at the mobile client is decreased in
terms of play-out of multimedia packets because mobile client
cannot play-out multimedia packets until the latest packet has
arrived.
RR scheduling allows the bursty flows to be spread out
among flows and not utilize the majority of the networks
bandwidth. RRs ability to utilize the bandwidth efficiently
allow each data flow to arrive to the mobile client on time.
When the multimedia packets of concurrent streams arrive
on time, the mobile client can give correct feedback to the
BS per multimedia packet request. When correct feedback
is given to the BS by the mobile client, the BS can syn-
chronize multimedia packets arrival times thus, improve the
QoS at the mobile client in terms of multimedia packet play-
out. Under 60 ms nonuniform delay jitters, RR out performs
PQ, WFQ and FIFO. With RR scheduling, the delay time be-
tween the concurrent data streams are half of the delay of
FIFO, WFQ and PQ scheduling algorithms. Because large
network delay jitters effect the inter-stream synchronization
different scheduling algorithms need to be used accordingly
when multimedia packets are played-out by the base station
onto the mobile client. We can conclude that RR clearly
out performs WFQ, PQ and FIFO scheduling under nonuni-
form delay network jitters in terms of inter-stream synchro-
nization. RR packet scheduling at the base station main-
tains inter-stream synchronization best among the 4 packet
scheduling algorithms. With RR scheduling, inter-stream
synchronization is maintain and on-time delivery of multime-
dia packets is guaranteed, enhancing the quality of service at
the mobile client in terms of smooth multimedia packet play-
out.
7. Conclusion and future work
In this paper we make a correlation between the play-out of
multimedia packets from the base station onto the mobile
client and wireless network multimedia synchronization. We
investigate four packet scheduling algorithms: First-Come-
First-Out (FIFO), Highest-Priority-First (PQ), Weighted-Fair
Queuing (WFQ) and Round-Robin (RR). We show the dif-
ferent algorithm under both uniform and nonuniform delay
jitters. We analyze their behavior, based on the QoS of the
mobile client in terms of buffer usage at the mobile client,
underflow at the mobile client, overflow at the mobile client,
the message complexity and intra-stream synchronization and
248
inter-stream synchronization during mobile client request and
play-out of multimedia packets. The different packet schedul-
ing algorithms affect the QoS of the mobile client, in terms
of buffer overflow, buffer underflows, buffer usage and intra-
stream synchronization and inter-stream synchronization. We
conclude that the RR packet scheduling is the most efficient in
the wireless multimedia in terms of synchronization between
multimedia servers, BS and mobile client. The RR packet
scheduling algorithm allows for 50 percent less delay between
concurrent data streams. The decrease in delay of concurrent
data streams enhances the QoS at the mobile client and al-
lowed for smooth multimedia packet play-out at the mobile
client. RR packet scheduling, with its fair packet schedul-
ing and flow isolation technique, is the best choice in terms
of buffer usage and inter-stream synchronization, at the ap-
plication layer, for multimedia packet play-out. Through our
research, we discover that the play-out of multimedia pack-
ets, by the BS onto the mobile client affects the overall net-
work synchronization and can be used to enhance the QoS at
the mobile client by allowing on-time delivery of multimedia
packets to the mobile client.
We see several directions for future research. We plan to
extend our MoSync scheme and make them power aware [6].
References
[1] S. Abhyankar, Handoff in mobile wireless network (November 2000).
[2] M. Adler, Y. Bartal, J.W. Byers, M. Luby and D. Raz, A modular analy-
sis of network transmission protocols, in: Proc. 5th Israeli Sympos. on
Theory of Computing and Systems (June 1997) pp. 1018.
[3] Z. Ali, M. Woo and A. Ghafoor, Distributed synchronization proto-
cols for multimedia services on Internet, IEEE/ACM Trans. Network-
ing 2(2) (1994) 2532.
[4] E. Biersack and W. Geyer, Synchronization delivery and play-out of
distributed stored multimedia streams, Multimedia Systems 7 (1999)
7090.
[5] U. Black, Second Generation Mobile and Wireless Networks (Prentice
Hall, NJ, 1999).
[6] A. Boukerche, Media synchronization protocols for mobile ad-hoc
wireless networks. In preparation.
[7] A. Boukerche and T. Huang, Real-time transport protocols of view mor-
phing over wireless communication.
[8] A. Boukerche, S. Hong and T. Jacob, An efficient synchronization
scheme of multimedia streams in wireless and mobile systems, IEEE
Trans. Parallel Distributed Syst. 13(9) (2002) 911923.
[9] A. Boukerche, S. Hong and T. Jacob, Synchronization and handoff
management schemes for wireless multimedia systems, Computer Net-
works 41 (2003) 347362.
[10] A. Boukerche, H. Sungbum and T. Jacob, A two-phase handoff man-
agement scheme for synchronizing multimedia units over wireless net-
works, in: IEEE Internat. Sympos. on Computer and Communications
(ISCC) (July 2003) pp. 10781084.
[11] H.-Y. Chen and J.-L. Wu, MultiSynch: A synchronization model for
multimedia systems, IEEE J. Selected Areas Comm. 14(1) (1996) 238
248.
[12] S. Floyed, B. Denny, V. Jacobson, S. Shenker and L. Zhang, Packet
scheduling research, in: DARTnet II Meeting (March 1995), Technical
Report (1995).
[13] Y. Ishibashi, S. Tasaka and H. Ogawa, A comparison of media syn-
chronization quality among reactive control scheme, Department of
Electrical and Computer Engineering, Nagoya Institute of Technology,
Nagoya 466-8555, Japan (2002).
BOUKERCHE AND OWENS
[14] K. Ito, S. Tasaka and Y. Ishibashi, Media synchronization quality of
packet scheduling algorithms, IEICE Trans. Comm. E85-B(1) (January
2002).
[15] J.P. Jarmasz and N.D. Georganas, Design a distributed multimedia
synchronization scheduler, in: Proc. IEEE Internat. Conf. Multimedia
Computing and Systems (June 1997) pp. 3442.
[16] T.V. Johnson and A. Zhang, Dynamic playput scheduling algorithms
for continuous multimedia streams, Multimedia Systems 7 (1999) 313
315.
[17] R. Krishnan and T.D.C. Little, Service aggregation through rate adapta-
tion using a single storage format, in: Proc. 7th Internat. Workshop on
Network and Operating System Support for Digital Audio and Video,
St. Louis, MO (1997) pp. 121232.
[18] L. Lamont, L. Li, R. Brimont and N.D. Georganas, Synchronization of
multimedia data for a multimedia news-on-demand application, IEEE
J. Selected Areas Comm. 14(1) (1996) 264278.
[19] K.S. Lee and M. El Zarki, Comparison of different scheduling al-
gorithms for packetized real-time traffic flows, in: Internat. Sympos.
on Wireless Personal Multimedia Communications, Denmark (2001)
pp. 2028.
[20] T.D.C. Little and F. Kao, An intermediate skew control system for mul-
timedia data presentation, in: Proceedings 3rd Internat. Workshop Net-
work and Operating System Support for Digital Audio and Video (No-
vember 1992) pp. 121132.
[21] D.M.H. Ly, User to network QoS parameter transformation in net-
worked multimedia systems, in: Workshop on Resource Allocation
Problems in Multimedia Systems (with RTSS96) (December 1996)
pp. 1119.
[22] D.L. Mills, Internet time synchronization: the network protocol, IEEE
Trans. Comm. 39(10) (1991) 14821493.
[23] M. Mowbray, G. Karlsson and K.T. Kohler, Capacity reservation for
multimedia traffic, Hewlett Packard Laboratories Report (2000).
[24] S. Pamanathan and V. Rangan, Adaptive feedback techniques
for synchronization multimedia retrieval over integrated networks,
IEEE/ACM Trans. Networking 1(2) (1993) 246260.
[25] N.U. Qazi, M. Woo and T.D.C. Little, A temporal reference framework
for multimedia synchronization, IEEE J. Selected Areas Comm. 14(1)
(1996).
[26] S. Ramanathan and P.V. Rangan, Adaptive feedback techniques for syn-
chronized multimedia retrieval over intergrated networks, IEEE/ACM
Trans. Networking 1(2) (1993) 246259.
[27] S. Shirani, F. Kossentini and R. Ward, Error concealment methods, a
comparative study, in: Proceedings of CCECE, Edmonton, Alberta,
Canada (May 1999) pp. 835840.
[28] T. Znati, R. Simon and B. Field, A network-based scheme for synchro-
nization of multimedia streams, in: Proc. Internat. Workshop Multime-
dia Synchronization, Tysons Corner, VA, May 19 (1995) pp. 1220.
[29] T. Znati, R. Simon and B. Field, A network-based scheme for synchro-
nization of multimedia streams, Technical Report TR95-02, Depart-
ment of Computer Science, University of Pittsburgh, PA (1995).
Azzedine Boukerche is a Canada Research Chair
and an Associate Professor of Computer Sciences at
the School of Information Technology and Engineer-
ing (SITE) at the University of Ottawa, Canada. He
is also the Founding and Director of PARADISE Re-
search Laboratory (PARAllel, Distributed and Inter-
active Simulation of LargE scale Systems and Wire-
less&Mobile Networking). Prior to this he was an
Assistant Professor of Computer Sciences at the Uni-
versity of North Texas, and Director of the Parallel
Simulations, Distributed and Mobile Systems Research Laboratory) at UNT.
He also worked as a Senior Scientist at the Simulation Sciences Division,
Metron Corporation located in San Diego. He was employed as a Faculty at
the School of Computer Science (McGill University, Canada) from 1993 to
1995. He spent the 19911992 academic year at the JPL-California Institute
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING
of Technology where he contributed to a project centered about the specifica-
tion and verification of the software used to control interplanetary spacecraft
operated by JPL/NASA Laboratory.
His current research interests include wireless and mobile networks, dis-
tributed and mobile computing, distributed systems, parallel simulation, dis-
tributed interactive simulation, and performance modeling. Dr. Boukerche
has published several research papers in these areas. He was the recip-
ient of the best research paper award at IEE/ACM PADS97, and the re-
cipient of the 3rd National Award for Telecommunication Software in 1999
for his work on a distributed security systems on mobile phone operations,
and has been nominated for the best paper award at IEEE/ACM PADS99
and at the ACM Modeling, Analysis and Simulation of mobile and wire-
less systems Conference 2001. He served the General co-Chair of the prin-
ciple Symposium on Modeling Analysis, and Simulation of Computer and
Telecommunication Systems (MASCOTS), in 1998, General Chair of the
Fourth IEEE International Conference on Distributed Interactive Simula-
tion and Real Time Applications (DS-RT2000), General Chair for the 3rd
ACM Conference on Modeling Analysis, and Simulation of Wireless and
Mobile Systems (MSWiM2000), as Program Co-Chair for the 5th IEEE In-
ternational Conference on Mobile and Wireless Computing and Communica-
tion (MWCN03), ACM/IFIP Europar 2003, IEEE Wireless Local Networks
(WLN03), the 35th SCS/IEEE Annual Simulation Symposium ANSS2002,
and the 10th IEEE/ACM Symposium on Modeling Analysis, and Simula-
tion of Computer and Telecommunication Systems (MASCOTS2002), the
3rd International Conference on Distributed Interactive Simulation and Real
Time Applications (DS-RT99), and ACM MSWiM2000, and as a Deputy
Vice Chair of Wireless and Mobilty Access Track for ACM WWW 2002, as
a Guest Editor for VLSI Design, the Journal of Parallel and Distributed Com-
249
puting (JPDC), ACM Wireless Networks (WINET), ACM Mobile Networks
and Applications (MONET), and the International Journal of Wireless and
Mobile Computing. He was the main organizer of a Special Session on Per-
formance Analysis of Mobile and Wireless Communication systems at the
7th IEEE HiPC Conference. He has been a member of the Program Commit-
tee of several international conferences such as ICC, VTC, ICPP, MASCOTS,
BioSP3, ICON, ICCI, MSWiM, PADS and WoWMoM, LWN, Networking
conferences.
Dr. A. Boukerche serves as a General Chair for the 1st International Con-
ference on Quality of Service on Heterogenous Wired/Wireless Networks
(QShine 2004), a Program Chair for the 5th IEEE Workshop on Wireless,
Mobile Ad-Hoc and Sensors Networks, as an Associate Editor for the In-
ternational Journal of Parallel and Distributed Computing (Mobile Comput-
ing Area), ACM/Kluwer Wireless Networks, Wiley International Journal on
Wireless Communications and Mobile Computing, and SCS Transactions on
Simulation, a Founding and a Steering Committee Chair of ACM MSWiM
Symposium, ACM Performance Evaluation of Wireless Ad hoc and Sensors
Networks (PE-WASUN), and IEEE DS-RT Symposium, and on the IEEE
Task Force on Cluster Computing (TFCC) Executive Committee. He is a
member of the IEEE and ACM.
E-mail: boukerch@site.uottawa.edu
Harold Owens II received his B.Sc. degree and M.Sc. degree from the De-
partment of Computer Science at the University of North Texas. His main
areas of research interests include mobile ad hoc networks, and wireless mul-
timedia systems.