Beruflich Dokumente
Kultur Dokumente
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.
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 level of the sensors. The coverage area (or simply coverage)
of the sensor network is the union of all such disks.
Ad-hoc sensor networks are emerging as a new sensing par- A sensor network is often used to detect intruders. An in-
adigm and have thus received massive research interest re- truder may start at a point S, follow an arbitrary trajectory
cently. Usually sensor nodes are semi-permanently deployed, (path) on the plane, and stop at some other point T on the
since the sensors themselves barely have any moving capac- plane. In some applications, a sensor network may need to
ity. However, the topology of such networks continuously keep track of the intruder at all times, as the intruder follows
changes over time due to a variety of reasons: For example, its trajectory; in some other applications, the networks func-
a sensor node may wear out due to its very limited battery tion may be simply to detect the presence of an intruder, in
power; a new sensor node may be inserted into the network; which case the network only needs to cover some part of the
or the layout of a sensor network may need to be changed trajectory. Thus, given two points S and T , two relevant types
in order to improve the quality of the network coverage in of trajectories on the plane are proposed [10]: the maximum
response to a change in the network requirements, which is breach path and the maximum support path. (In [10], these
accomplished by changing the placement of current (or in- paths are called maximal breach path and maximal support
serting, deleting) sensors in network. path, respectively.)
In this paper, we address the problem of how to dynami- The maximum breach path measures the vulnerability of a
cally maintain two important measures on the quality of the sensor network by, as the name suggests, completely avoid-
coverage of a sensor network: the best-case coverage distance ing the coverage area of the sensor network: It is a trajectory
and the worst-case coverage distance of the network. We also between the start point S and the stop point T that stays as
address a closely related problem, namely that of finding a far away from the sensors as possible. On the other hand,
shortest maximum support path. the maximum support path measures the efficiency of the net-
In a sensor network, each sensor bears the ability to detect work coverage: This path is a trajectory between S and T
objects around it. The coverage of a sensor is limited by its which stays as close to the sensors as possible. The distance
energy level. Assuming that a sensors detecting ability is of a point P to the sensor network is defined as the smallest
omnidirectional, we can model the coverage of a sensor as a Euclidean distance from P to one of the sensor nodes. A max-
disk (under 2-norm on the Euclidean plane1 ) centered at the imum breach path from S to T is a path from S to T such that
sensor. The radii of such disks are determined by the energy the minimum distance from a point P in the path to the sen-
This work was supported in part by NSF CAREER Award CCR-9985284. sor network is maximized: this distance is called the worst-
Corresponding author. case coverage distance of the network. Similarly, a maximum
1 A disk of radius r centered at (x, y) under 2-norm in R2 is the set of points
support path from S to T is a path such that the maximum
(p, q) such that (p x)2 + (q y)2 r. distance of a point P in the path to the sensor network is min-
10 HUANG ET AL.
imized: this distance is called the best-case coverage distance In order to dynamically maintain the best- and worst-case
of the network. coverage distance efficiently, we need to maintain some infor-
When the topology of a sensor network changes, the qual- mation on the current topology of the sensor network; when
ity of its coverage most probably will be affected. We would the network topology changes, we need to update this infor-
like to maintain an assessment on the quality of the network mation. We also perform queries for the current best-case
coverage which, as explained above, can be done by main- and worst-case coverage distances, based on the information
taining the worst-case and best-case coverage distances ef- maintained. Hence, the cost (or running time) of our algo-
ficiently at all times. This would give a clear indication on rithms are measured in terms of their respective update cost
how effective the network coverage is at any given point in i.e., the cost to update the topology information, which is
time, possibly calling for the insertion of new nodes in the charged per relevant topology change in the network and
network (e.g., when the coverage deteriorates due to node the query cost, which is the cost incurred when answering
failures) or to a network re-design phase. Whenever neces- a query for the current best-case or worst-case coverage dis-
sary, the actual paths which give the best-case and worst-case tance.
coverage distances can be retrieved. As we will see later, in In sections 4 and 5, we formally define a relevant topol-
sections 4 and 5, our algorithms for maintaining the worst- ogy change which will henceforth be called an event for
case and best-case coverage distances have poly-logarithmic the problems of maintaining the best-case and worst-case cov-
update and query costs, as defined later. To the best of our erage distances, respectively.
knowledge, this is the first work which formalizes and ad- The remainder of the paper is organized as follows. Sec-
dresses this problem in a dynamic scenario. tion 1.1 states our results. In section 2, we present some re-
For a moment, let us assume that all sensors have the same lated work in the literature. Section 3 covers some prelim-
energy power and thus all disks have the same radius r. We inaries and sketches the basic framework of our solutions.
call such a sensor network a uniform sensor network with cov- We present the low constant approximation algorithms for the
erage radius r. In a uniform sensor network, all of the paths best- and worst-case coverage distance in sections 4 and 5 re-
whose minimum distance of a point in the path to a sensor spectively. In section 6 we address the closely related problem
is larger than the coverage radius are equivalent, in the sense of efficiently finding a shortest maximum support path. Sec-
that the sensors in the network will not be able to detect an in- tion 7 concludes the paper with some possible lines for future
truder using any such path. Similarly, all of the paths whose work.
maximum distance of a point in the path to a sensor is smaller
than the coverage radius are equivalent, in the sense that any 1.1. Our results
such path is entirely contained in the coverage area of the net-
work. The worst coverage radius (see [10]) is defined to be In this section, we summarize the main results of this paper.
the maximum coverage radius r such that there exists a trajec- One of the main contributions of this work is to take into ac-
tory P between given points S and T which does not intersect count the dynamic nature of sensor networks, and to propose
the interior region of the area covered by the uniform sensor a framework which can be used to continuously monitor the
network (i.e., P may touch the coverage area, intersecting quality of the network coverage. Let n denote the current
it at a discrete number of points only). We can think of the number of sensors in the network.
worst-coverage radius as being the maximum energy that can In the following sections, we present two algorithms to
be assigned to the sensor nodes which still would not prevent maintain low constant approximations on the best-case and
an intruder from escaping from S to T without being detected worst-case coverage distances. Both algorithms have low
(for simplicity, we assume that a sensor will not be able to update and query costs. Namely, our algorithms achieve a
detect an intruder who only touches its coverage area). Cor- (1 + )-approximation
on the best-case coverage distance,
respondingly, the best coverage radius (see [10]) is defined and a ( 2 + )-approximation on the worst-case cover-
to be the minimum coverage radius r such that there exists age distance, for any fixed > 0. The amortized up-
a trajectory between S and T that is totally covered by the date cost per event of the best-case coverage distance algo-
uniform sensor network. rithm is O(log3 n), and the respective query cost is worst-case
We introduce uniform sensor networks as a merely concep- O(log n). For the worst-case coverage algorithm, the update
tual tool in order to facilitate the presentation of our approxi- cost per event is worst-case O(log2 n) and the query cost is
mation algorithms and their analyses, following a similar ap- worst-case O(1). A formal definition of an event for each of
proach as Li et al. [9]. (The actual sensor network in consid- the problems considered follows in sections 4 and 5, respec-
eration has nodes with arbitrary energy levels and therefore tively.
is not assumed to be uniform.) In fact, if we think of a uni- As a byproduct of our algorithm for maintaining the worst-
form sensor network built on top of the placement of the sen- case coverage distance, we extend the concatenable queue
sor nodes currently deployed in the general sensor network in data structure to also serve as a priority queue. All the opera-
consideration, the worst-coverage radius of the uniform net- tions on this extended data structure have worst-case O(log n)
work is indeed equal to the worst-case coverage distance of running time.
the general sensor network, and the best-coverage radius is We also present an O(n log n) algorithm for computing
indeed equal to the best-case coverage distance. an exact shortest maximum support path between two given
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS 11
points S and T , improving on the best-known previous results port path with O(n2 log n) worst-case communication com-
by Li et al. [9]. A shortest maximum support path from S plexity, and an algorithm that computes a 2.5-approximation
to T is a maximum support path from S to T such that the of a shortest maximum support path (i.e. the total length of
Euclidean length of the trajectory followed in this path is min- the obtained path is at most 2.5 times the length of a short-
imum. In [9], two algorithms are presented for computing the est maximum support path) with O(n log n) communication
maximum support path: One algorithm computes an exact complexity.
shortest maximum support path in O(n2 log n) time; the other Meguerdichian et al. [11] proposed an exposure-based for-
algorithm provides a 2.5-approximation on the shortest max- mulation for analyzing the coverage of paths taken by polyg-
imum support path in O(n log n) time. One should note that onal objects: they define a path-dependent integral, which
the algorithms presented by Li et al. can be implemented in consists of the trajectories of all the points of a polygonal ob-
a distributed fashion (we use the communication complexity ject (the polygonal object is able to rotate), and not only of
as the time bound for the sequential versions of their algo- the trajectory of the objects center point.
rithms), whereas the algorithms presented in this paper are all Recently, Zhang and Hou [13] proved that if the commu-
centralized. nication rage of a sensor is at least twice its sensing range,
The update costs of our algorithms for approximately a complete coverage of a convex area implies connectivity
maintaining the best- and worst-case coverage distances are among the working set of nodes and derive optimality con-
much cheaper than maintaining the best- or worst-case cov- ditions under which a subset of working sensor nodes can be
erage distances using the best-known algorithms in the liter- chosen for full coverage. Wang et al. [12] designed a Cov-
ature prior to this work. In fact, the best previously known erage Configuration Protocol (CCP) that can provide differ-
algorithm for maintaining the best-case (resp., worst-case) ent degrees of connected coverage and present a geometric
coverage distance maintains the exact distance by repeat- analysis of the relationship between coverage and connectiv-
edly re-computing the maximum support path (resp., max- ity. Huang and Tseng [8] present an algorithm with runtime
imum breach path) using the O(n log n) algorithm by Li of O(n2 log n) that decides whether every point in a given ser-
et al. [9] (resp., the O(n2 log n) algorithm by Meguerdichian vice area is covered by at least one sensor.
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) 3. Preliminaries
these two distances.
Before heading into the technical details of our algorithms,
we introduce some basic concepts which will be used in both
2. Related work sections 4 and 5. The first concept we introduce is that of
growing disks, which will help us translate our problems into
Meguerdichian et al. [10] considered the problems of finding graph connectivity problems.
the maximum breach path and the maximum support path on The growing disks concept was previously proposed in [9].
a sensor network. They [10] present an O(n2 log ) runtime We restate it in terms of the coverage radius of a uniform sen-
algorithm for the maximum breach path problem, where n is sor network as defined in section 1. (In section 1, we saw how
the number of sensors in the sensor network, and is the dif- the coverage radius of a virtual uniform overlay sensor net-
ference between the highest and the lowest weight of an edge work directly relates to the worst-case and best-case coverage
in the Voronoi Diagram of the sensor network. Their algo- distances of the actual network.) Assume we have a uniform
rithm for computing the maximum support path has the same sensor network with coverage disks centered at the sensors.
running time as their maximum breach path algorithm. The Define U (r) to be the region on the plane composed of the
O(log ) factor can be easily converted into O(log n) in the union of all of the coverage disks when the coverage radius
algorithm that solves the maximum breach path problem if we is r. Let U (r) be the complement of the region U (r). At the
perform a binary search over a sorted list of the radii of sen- very beginning, we set the coverage radius to be equal to 0.
sors instead of using a linear search as in [10]. The algorithms Then U (r) is the union of discrete singletons. As the cov-
presented in [10] heavily rely on geometric structures such as erage radius grows, the disks centered at the sensors become
the Voronoi Diagram and Delaunay triangulation of the net- larger and might get connected into larger regions. Therefore,
work, which cannot be constructed efficiently in a distributed U (r) might get disconnected into separate regions. For any
manner. two given points S and T , the best coverage radius is the min-
Li et al. [9] prove the correctness of the algorithms given imum r such that S and T are in the same connected region of
in [10]. They also show how to find a maximum support U (r), while the worst coverage radius is the minimum r such
path in O(n log n) time using a centralized algorithm, or with that S and T belong to two disconnected regions in U (r).
O(n log n) communication complexity bits in a distributed Hence, the best and worst coverage radius problems translate
fashion. In addition, Li et al. [9] present two algorithms for to connectivity problems on U (r) and U (r), respectively. Fig-
computing a shortest (with respect to the Euclidean length of ure 1 illustrates these ideas. We will further translate the best
the trajectory followed in this path) maximum support path: and worst coverage radius problems into graph connectivity
an algorithm that computes an exact shortest maximum sup- problems.
12 HUANG ET AL.
changes and an update on the connectivity information is trig- worst-case coverage distance. We first present a (1 + )-
gered. Hence the update cost is the cost to update the connec- approximation algorithm (for any > 0) for a simplified sen-
tivity information of G(U (r)) per event. When a certificate sor network model, where the coverage disks are considered
fails and an event occurs, it takes constant time for the kinetic under infinity-norm. Since there is only a 2 gap between
data structure to process the failure (due to the setup of at most infinity-norm and 2-norm, a (1 + /2)-approximation fac-
a constant number of new certificates). We do not explicitly tor for infinity-norm dilates into a ( 2 + )-approximation
take this cost into account when computing the update cost of factor when applied to the 2-norm scenario, for any fixed
the maintenance of the connectivity information of G(U (r)), > 0. The infinity-norm of a vector v = (x1 , . . . , xd ) in a
since it would not change the asymptotic bound on the update d-dimensional space is defined as v = max(|x1 |,
cost. . . . , |xd |). Under infinity-norm, the distance between two
We adapt the main theorem in [5, theorem 5.4], to better points on the plane is the maximum of the difference of
serve our purposes. The uniform disk graph G(U (r)) corre- their x coordinates and the difference of their y coordinates.
sponds to a unit-disk graph if we rescale one unit to be equal Hence the coverage region of a sensor is square shaped and
to 2r. its boundary is composed of four line segments. As we will
see later, this simple boundary shape allows for an efficient
Lemma 1 (Adapted from [5, theorem 5.4]). In [5], an algo- maintenance scheme.
rithm to dynamically maintain the connectivity of G(U (r)) is Recall the solution framework presented in section 3. The
presented. The update cost is amortized O(log2 n) per event. core of our algorithm is to check, for any two given points S
The query cost is worst-case O(log n/ log log n). and T , whether the region in U (r) containing S is connected
to that containing T . If we can maintain some information
We still need to show how to determine which disks con- such that each query on connectivity of regions takes only
tain the given points S and T , at any given time. We sort all poly-logarithmic time, the cost of update against mobility is
sensors according to their distances to the fixed point S. We also poly-logarithmic.
maintain a binary heap on this ordering. Once the ordering In our algorithm, regions in U (r) are represented by their
changes, we update the heap in O(log n) time. This intro- boundaries. Only one region in U (r) may be infinite in area.
duces a new type of event namely, when a sensor changes We call such an unbounded region the outer face. All of the
its location and needs to be re-inserted in this ordering be- other (bounded) regions are called inner faces. Since we con-
sides the other two events defined earlier. The update cost sider the infinity-norm, each disk is represented by a square
for this event is O(log n). To check which disk contains S, on the plane. Thus the boundary of any inner face is a sim-
we find the closest sensor p to S. We check if the distance ple cycle composed of a sequence of line segments, while the
from p to S is larger than the coverage radius. If so, then S is boundary of the outer face comprises several simple cycles.
not contained in any disk. Otherwise, we know that the disk To differentiate these cycles, we call a cycle that is the bound-
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 (a)
G(U (r)).
5. Dynamic worst-case coverage distance 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,
In this section, we present our ( 2 + )-approximation C , and E and E are relocated. Edges (C, C ) and (E, E ) are removed, and
algorithm, for any > 0, to dynamically maintain the edges (C, E ) and (E, C ) are inserted.
14 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.
Figure 3. Outer- and inner-cycles. When the outer cycle in (a) breaks into
5.1. Dynamics of cycles
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). In this section, we first formally define the representation we
use for the connectivity graph G(U (r)). Second, we ad-
ary of an inner face an inner cycle, and a cycle on the bound- dress the dynamics of the connectivity graph. And finally,
ary of the outer face an outer cycle. Figure 3 illustrates some we present an algorithm for maintaining the connectivity in-
of these concepts. The shaded areas in the figure define U (r), formation on the regions of U (r).
and the unshaded areas define U (r). In (b), U (r) is divided The boundary of a standalone square is the simplest cycle
into two regions, the unbounded region is the outer face, the in G(U (r)). We represent a square by eight vertices and eight
bounded region is the inner face. The boundary of the inner edges as shown in figure 2. For every corner X of a square,
face is an inner cycle and that of the outer face is an outer we introduce two vertices X and X . Hence we have O(n)
cycle. In (c), the boundary of the outer face consists of two vertices and edges in G(U (r)), where n always denotes the
disjoint outer cycles. current number of sensors in the network. The extra vertices
Below we describe a method which translates the connec- help us to efficiently maintain the graph when squares start to
tivity of regions in U (r) into a graph connectivity problem. move and overlap on the plane (including when sensors are
The first step is to represent outer cycles and inner cycles by added or removed from the network). In the following, we
a graph. There are only vertical line segments and horizon- will show that the dynamics of sensors will not change the
tal line segments in both outer and inner cycles, and those O(n) bound on the number of vertices and edges.
line segments only meet at their endpoints. Hence we can When two squares meet, at most two pairs of line seg-
draw a graph such that the vertices are the endpoints and the ments of their boundaries intersect. Without loss of gener-
edges are the line segments. We call this graph the connectiv- ality, suppose a vertical edge B C intersects with a horizon-
ity graph G(U (r)). (For convenience, the connectivity graph tal edge E F at a point Z, and the new boundary comprises
will actually be implemented in a slightly different way, as we edges B Z and ZF . Then we simply relocate vertices C and
explain in section 5.1.) E to Z, insert an edge CE and remove edges CC and EE
Every outer or inner cycle is a cycle in the graph and any from G(U (r)). Figure 2 illustrates this operation. Note that
two of them are disjoint, i.e., disconnected in the graph. This we do not introduce any new vertex or remove any old ver-
coincides with the fact that any two distinct inner faces are tex. In fact, since G(U (r)) contains no information of the
disconnected, and that any inner face is disconnected from vertexs location, we do not need to perform any relocation
the outer face. of a vertex when we operate on G(U (r)). The cases of a verti-
The connectivity of G(U (r)) is thus analogous to that of cal edge intersecting with a vertical edge, and of a horizontal
U (r): Two regions are connected in U (r) if and only if their edge intersecting with a horizontal edge are analogous, and
boundary cycles are connected in the graph, or they are both can thus be also handled by at most two edge insertions and
part of the outer face boundary. Thus we could apply the algo- at most two edge deletions. Since we never change the num-
rithm proposed by Holm et al. [7], which dynamically main- ber of vertices in the graph, and since each vertex has degree
tains graph connectivity, to maintain the connectivity of the at most 2, the O(n) upper bound on number of vertices and
regions in U (r). The update cost per edge insertion or dele- edges in G(U (r)) always hold. The following fact follows:
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS 15
Fact 1. When two squares meet or separate, up to four edge belong to the same cycle. If so, we may also need to update
insertions and deletions are needed to update the connectivity the topmost edge of the cycle. When event (III) occurs, we
graph G(U (r)). update the orderings with respect to distances to S and T .
We use the kinetic data structure as defined in [3] as our
When the topology of the network changes, cycles in event monitor (unlike G(U (r)), G(U (r)) is not a unit-disk
G(U (r)) may also undergo changes. A cycle may break into graph and therefore the results in [5] do not apply). Each
two smaller cycles; or two cycles may merge into a longer event can be detected and processed in constant time.
cycle. Both these operations impose changes on the connec- In the following, we present our update scheme. We will
tivity of G(U (r)). Cycles break or merge only when two sen- also show that the update cost per event is O(log n). We store
sors coverage disks meet or separate. Hence we need to de- a cycle as a sequence of consecutive edges. In section 5.2
tect the time when those happen in order to trigger an update. we introduce a data structure which supports the following
When a cycle breaks, it could break into an outer cycle and operations on sequences of edges:
an inner cycle (as shown in figure 3). We need to differentiate
outer cycles from inner cycles since all outer cycles define INSERT insert an edge into a sequence
the same region, namely the outer face. In order to determine DELETE delete an edge from a sequence
whether a cycle is an outer cycle, one only needs to identify CONCATENATE concatenate a sequence to the end of an-
the topmost edge of the cycle: If the topmost edge of the cycle other sequence
is the top boundary of a square, then the cycle is an outer SPLIT split a sequence into two sequences
cycle; otherwise, the topmost edge of a cycle is the bottom SWAP swap the y position of two edges
boundary of a square, and the cycle is an inner cycle. Hence MAX return the topmost edge of a sequence
we need to maintain the topmost edge of each cycle as sensors MEMBER return the representative edge of a sequence
move. The topmost edge of a cycle may change only when
two horizontal line segments swap their y position. Therefore Each of these operations can be executed in worst-case
we also need to monitor these line segment swaps. running time O(log n), as stated in lemma 4.
Recall that the original problem we aim to solve is to check The update per type (I) or (II) event is as follows. When
whether the region containing a given point S is connected to squares move and the shape of a cycle changes, up to a con-
that containing T . We need to determine which region con- stant number of INSERT and DELETE operations are needed
tains a given point and also to update this information as sen- to update the cycle per event. When two edges in a cycle ex-
sors move. As described in section 4, we sort all sensors ac- changes their y position, we execute SWAP to update the y
cording to the distance from the fixed point S and maintain a position per event. We can execute MAX to know whether a
binary heap on this ordering, with update cost O(log n) on the cycle is an outer cycle or not. Recall that a cycle is an outer
heap. In order to check which region S belongs to, we need cycle if and only if the topmost edge of the cycle is the top
to find the cycle representing the region. Again we find the boundary line segment of a square. Cycle merges or breaks
closest sensor p to S and check if the distance is smaller than can be carried out by a constant number of CONCATENATE
the radius of the coverage disk of p. If so, then the point S and SPLIT operations. Since only a constant number of IN-
does not belong to any region of U (r). Otherwise, we check SERT, DELETE, CONCATENATE, SPLIT, SWAP and MAX
the eight vertices of the square representing the closest sensor operations are executed per event, the update cost per event is
to S, find the closest one of these vertices to S, and the cycle worst-case O(log n). As we have explained earlier, the update
containing this closest vertex represents the region contain- cost per type (III) event is also O(log n).
ing S. This query takes constant time. We maintain a similar A data structure that supports the operations above can also
data structure for T . Thus we also need to monitor and de- be used for efficiently performing a connectivity check. As-
tect the time when two sensors swap their relative position in sume that S and T are not covered by any sensor in U (r)
these orderings. and therefore both belong to U (r). We can find the closest
We summarize all of the above in the following three vertices u and v to points S and T , respectively, in constant
types of events, which we need to monitor in order to trig- time. Then we check if u and v belong to the same cycle by
ger mandatory updates, as sensors move on the plane: performing two MEMBER operations. If so, then S and T
belong to the same region in U (r). Otherwise, we need to
(I) Two vertical line segments swap their x position,
check whether the closest cycles to S and T are both outer
(II) Two horizontal line segments swap their y position, and cycles by two executions of the MAX operation. If both of
them are outer cycles, then both S and T belong to the outer
(III) Two sensor swap their position in the orderings of the
face, and hence are in the same region. Otherwise, S and T
sensors distance to the given points S and T .
belong to two disconnected regions. This procedure can be
When events (I) or (II) occurs, we can check in constant implemented in O(log n) time.
time whether two coverage disks meet or separate. If they do, We summarize all of the above in lemma 2.
we check whether the event leads to a cycle break or merge,
and update the data structure accordingly. When event (II) oc- Lemma 2. For any two given points S and T , we maintain a
curs, we can check whether the two horizontal line segments data structure with O(log n) update cost per event such that
16 HUANG ET AL.
the query to check whether the region in U (r) containing S is INSERT or DELETE, it takes an additional O(log n) time to
connected to that containing T takes O(log n) time. update the y coordinate of all internal nodes due to the edge
insertion or deletion. Both CONCATENATE and SPLIT are
Combining lemma 2 with the algorithmic framework pre- implemented by up to O(log n) joins or breaks of trees at the
sented in section 3, we have our (1 + )-approximation al- root node. Since updating the y coordinate at the root node
gorithm, for any > 0, for the worst-case coverage distance takes constant time (by comparing all the children of the root),
under infinity-norm, as stated in the lemma below. If every we incur at most an additional O(log n) time per CONCATE-
time we perform an update operation, we keep track of the NATE or SPLIT. Thus the asymptotic running time of IN-
smallest ri such that S and T are disconnected in G(U (ri )), SERT, DELETE, CONCATENATE, and SPLIT remains un-
then each query operation can be performed in O(1) time. changed. The running time of MEMBER is not affected by
SWAP or MAX operations.
Lemma 3. Under infinity-norm, our algorithm dynamically
maintains a (1 + )-approximation of the worst-case cover- Lemma 4. The extension of the concatenable queue data
age distance for any > 0. The update cost is worst-case structure supports the operations of INSERT, DELETE,
O(log n log1+ R) per event, and the query cost is worst- CONCATENATE, SPLIT, SWAP, MAX and MEMBER.
case O(1). Each operation has worst-case running time of O(log n).
Hence the ( 2+)-approximation algorithm for the worst-
6. Exact shortest maximum support path
case coverage distance under 2-norm follows:
We consider the problem of finding a maximum support path
Theorem 2. Our algorithm dynamically maintains a ( 2+) between S and T such that the Euclidean length of the tra-
-approximation of the worst-case coverage distance, for any jectory followed by this path is minimum. Below we present
> 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). known previous results by Li et al. [9]. One should note that
the algorithms presented in [9] can be implemented in a dis-
Corollary 2. If > 0 is fixed, then our algorithm has worst- tributed fashion, whereas the algorithm we present in this sec-
case O(log2 n) update cost per event, and worst-case O(1) tion is intrinsically centralized.
query cost. 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)
5.2. Extended concatenable queue time. Next, we obtain a collection of uniform disks by setting
the radius of each sensor to be rbest . Let U denote the union
In this subsection we introduce a data structure that sup- of all these uniform disks. Define the complement region of
ports the operations INSERT, DELETE, CONCATENATE, the union C = R2 \ U .
SPLIT, SWAP, MAX and MEMBER efficiently. The data The problem of finding a shortest maximum support path
structure is an extension of the concatenable queue data struc- is equivalent to the problem of finding a shortest S, T -path
ture [1]. In [1], a concatenable queue is implemented by a in R2 avoiding C, since we are seeking for a maximum sup-
23 tree (a RedBlack tree would also work, for example), port path and rbest is the best coverage radius. (Since rbest is
and all the data is stored at the leaf nodes. A concaten- the best coverage radius, any maximum support path is con-
able queue supports the operations INSERT, DELETE, CON- tained in U ; in fact any path from S to T in U is a maximum
CATENATE, SPLIT and MEMBER, and each operation takes support path.) A shortest maximum support path can only
time O(log n) in the worst case. In the following paragraphs, contain straight line segments as edges, otherwise the path
we will show how to also implement the SWAP and MAX would not be shortest. Therefore, we can replace each arc in
operations on a concatenable queue in O(log n) time. C by a straight line segment. In such fashion we obtain a new
We associate each edges y coordinate to the correspond- set of obstacles C as a collection of polygonal objects with
ing leaf node in the 23 tree. To each internal node t, we asso- possible holes that have a total O(n) number of vertices.
ciate the maximum y coordinate of a leaf node in the subtree We can remove these holes and obtain slightly larger num-
rooted at t. This is done by comparing all the y coordinates ber of disjoint polygonal objects by cutting the existing ob-
associated to ts children in the tree, taking constant time per jects with segments that connect the vertices of the holes and
internal node. When the y coordinate of an edge changes, and the external boundary in an arbitrary fashion. Note that the
a SWAP operation is invoked, it takes at most O(log n) time to number of total vertices of the disjoint polygonal objects has
climb up the tree and update all the internal nodes on the way not changed, i.e., it is still O(n). Thus, our problem translates
up. Starting from any given edge on a cycle, it takes O(log n) to that of finding a shortest path in R2 that avoids the polyg-
time to reach the root of the 23 tree where we can find the onal obstacles in C . The idea now is to use an algorithm by
topmost edge of the cycle. Hence the O(log n) running time Hershberger and Suri [6], which finds a shortest path between
of MAX follows. S and T on the Euclidean plane that avoids polygonal obsta-
We need also to justify that the above modification does cles in O(n log n) time. Hence the total running time of our
not increase the running time of all other operations. Per each algorithm is O(n log n).
DYNAMIC COVERAGE IN AD-HOC SENSOR NETWORKS 17
As described in section 2, two algorithms are presented [12] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless and C.D. Gill, Integrated
for computing the maximum support path by Li et al. in [9]: coverage and connectivity configuration in wireless sensor networks,
One algorithm computes an exact shortest maximum sup- in: Proc. of the 1st ACM Conf. on Embedded Networked Sensor Systems
(2003).
port path in O(n2 log n) time; the other algorithm provides a
[13] H. Zhang and J.C. Hou, Maintaining sensing coverage and connectivity
2.5-approximation on the shortest maximum support path in in large sensor networks, Technical Report UIUCDCS-R-2003-2351,
O(n log n) time. Our algorithm improves on the running time UIUC (2003).
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 Hai Huang is a Ph.D. student in the Department of
the sequential versions of their algorithms), whereas our al- Computer Science and Engineering at Arizona State
gorithm is centralized in nature. 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-
7. Future work
sity, P. R. China, in 1996 and 1999, respectively. His
current research work focus on clustering and rout-
In this paper, we present poly-logarithmic dynamic algo-
ing problems in mobile ad-hoc networks, and on the
rithms to maintain approximations of two relevant measures chordal graph completion problem with applications
namely, the best- and worst-case coverage distances of the to scientific computing.
quality of the network coverage in wireless sensor networks. E-mail: hai@asu.edu
An interesting open question is whether we can maintain ex-
act best-case and worst-case coverage distances for Euclidean Andra W. Richa joined the Department of Com-
metric with poly-logarithmic update time. 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
Acknowledgement from the School of Computer Science at Carnegie
Mellon University, in 1995 and 1998, respectively.
We express our thanks to Micha Sharir for his comments. 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
References Federal University of Rio de Janeiro, Brazil, in 1992
and 1990, respectively. Prof. Richas main area of research is in network
[1] A.V. Aho, J.E. Hopcroft and J.D. Ullman, The Design and Analysis of algorithms. Some of the topics Dr. Richa has worked on include packet
Computer Algorithms (Addison-Wesley, Reading, MA, 1974). scheduling, distributed load balancing, packet routing, mobile network clus-
[2] J. Basch, Kinetic data structures, Ph.D. dissertation, Stanford Univer- tering and routing protocols, and distributed data tracking. Prof. Richas data
sity (1999). tracking (or lookup) algorithm has been widely recognized as the first bench-
[3] J. Basch, L.J. Guibas and J. Hershberger, Data structures for mobile mark algorithm for the development of distributed databases in peer-to-peer
data, in: Proc. of 8th ACMSIAM Symposium on Discrete Algorithms networking, having received over 55 academic journal or conference publi-
(1997) pp. 747756. cations, and being implemented as part of two of the current leading pojects
[4] B.N. Clark and C.J. Colbourn, Unit disk graphs, Discrete Math. 86 in peer-to-peer networking. Dr. Richas was the recipient of an NSF CA-
(1990) 165177. REER Award in 1999. For a selected list of her publications, CV, and current
[5] L.J. Guibas, J. Hershberger, S. Suri and L. Zhang, Kinetic connectivity research projects, please visit http://www.public.asu.edu/aricha.
for unit disks, Discrete Comput. Geom. 25 (2001) 591610. E-mail: aricha@asu.edu
[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- Michael Segal was born at October 12, 1972 in
ministic fully-dynamic graph algorithms I: Connectivity and minimum USSR. In 1991 he immigrated to Israel and started
spanning tree, Technical Report DIKU-TR-97/17, Department of Com- to study computer science in Ben-Gurion University
puter Science, University of Copenhagen (1997). of the Negev. He finished his B.Sc., M.Sc. and Ph.D.
[8] C.-F. Huang and Y.-C. Tseng, The coverage problem in a wireless sen- degrees in 1994, 1997, and 1999, respectively. Dur-
sor networks, in: Proc. of the 2nd ACM Internat. Conf. on Wireless ing a period of 19992000 Dr. Michael Segal held a
Sensor Networks and Applications (2003) pp. 115121. MITACS National Centre of Excellence Postdoctoral
[9] X.-Y. Li, P.-J. Wan and O. Frieder, Coverage in wireless ad-hoc sensor Fellow position in University of British Columbia,
networks, IEEE Trans. Comput. 52 (2003) 111. Canada. Dr. Segal joined the Department of Com-
[10] S. Meguerdichian, F. Koushanfar, M. Potkonjak and M.B. Srivastava, munication Systems Engineering, Ben-Gurion Uni-
Coverage problems in wireless ad-hoc sensor networks, in: Proc. of the versity, Israel in 2002 where he holds now a position of senior lecturer and
20th IEEE INFOCOM (2001) pp. 13801387. serves as departments Deputy Chairman. His primary research is algorithms
[11] S. Meguerdichian, F. Koushanfar, G. Qu and M. Potkonjak, Exposure in (sequential and distributed), data structures with applications to optimization
wireless ad-hoc sensor networks, in: Proc. of the 7th ACM MOBICOM problems, mobile wireless networks, communications and security.
(2001) pp. 139150. E-mail: segal@cse.bgu.ac.il
Mobile Networks and Applications 10, 1934, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
1.2. Formulation of topology control problems Definition 2.1. A property P of the (directed or undirected)
graph associated with an ad hoc network is monotone if the
Topology control problems have been studied under two property continues to hold even when the powers assigned to
graph models. The discussion above corresponds to the di- some nodes are increased while the powers assigned to the
rected graph model studied in [23]. The undirected graph other nodes remain unchanged.
model proposed in [16] represents the ad hoc network as an
undirected graph in the following manner. First, the directed Example. For any k 1, the property k-N ODE C ONNECTED
graph model for the network is constructed. Then, for any for undirected graphs is monotone since increasing the pow-
pair of nodes u and v, whenever both the directed edges (u, v) ers of some nodes while keeping the powers of other nodes
and (v, u) are present, this pair of directed edges is replaced unchanged may only add edges to the graph. However, prop-
by a single undirected edge {u, v}. All of the remaining di- erties such as ACYCLIC or B IPARTITE are not monotone.
rected edges are deleted. Under this model, the goal of a
topology control problem is to assign transmission powers to Some of the topology control problems considered in this
nodes such that the resulting undirected graph has a specified paper are NP-complete. For such problems, we study approx-
property and a specified function of the powers assigned to imation algorithms. In this context, an approximation algo-
nodes is minimized. Note that the directed graph model al- rithm provides a performance guarantee of if for every in-
lows two-way communication between some pairs of nodes stance of the problem, the solution produced by the approxi-
and one-way communication between other pairs of nodes. mation algorithm is within the multiplicative factor of of the
In contrast, every edge in the undirected graph model corre- optimal solution. A polynomial time approximation scheme
sponds to a two-way communication. (PTAS) is an approximation algorithm that, given a problem
In general, a topology control problem can be specified instance and an accuracy requirement , produces a solution
by a triple of the form M, P, O. In such a specification, that is within a factor 1 + of the optimal solution.
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- 3. Previous work and summary of results
per O {M AX P, T OTAL P} (abbreviations of Max Power and 3.1. Previous work
Total Power). For example, consider the D IR , S TRONGLY
C ONNECTED, M AX P problem. Here, powers must be as- The form of topology control problems considered in this
signed to transceivers so that the resulting directed graph is paper was proposed by Ramanathan and Rosales-Hain [23].
strongly connected and the maximum power assigned to a They presented efficient algorithms for two topology control
transceiver is minimized. Similarly, the U NDIR , 2-N ODE problems, namely U NDIR , 1-N ODE C ONNECTED, M AX P
C ONNECTED, T OTAL P problem seeks to assign powers to and U NDIR , 2-N ODE C ONNECTED, M AX P. After deter-
the transceivers so that the resulting undirected graph has a mining the minimum value for the objective, their algorithms
node connectivity (see below for definition) of (at least) 2 and also reduce the power assigned to each transceiver such that
the sum of the powers assigned to all transceivers is mini- each power level is minimal while maintaining the desired
mized. graph property. In addition, they presented efficient distrib-
uted heuristics for these problems.
Several groups of researchers have studied the D IR ,
2. Additional definitions S TRONGLY C ONNECTED, T OTAL P problem [5,7,8,16].
However, it is not difficult to see that their NP-hardness re-
This section collects together the definitions of some graph sults as well as approximation algorithms also hold for the
theoretic and algorithmic terms used throughout this paper. U NDIR , 1-N ODE C ONNECTED, T OTAL P problem. The pa-
Given an undirected graph G(V , E), an edge subgraph per [5] proves that the problem is NP-hard and presents
G (V , E ) of G has all of the nodes of G and the edge set E an approximation algorithm with a performance guarantee
is a subset of E. Further, if G is an edge weighted graph, then of 2. The other references consider a geometric version of
the weight of each edge in G is the same as it is in G. the problem along with a symmetry assumption concerning
The node (edge) connectivity of an undirected graph is the transmission power thresholds. More precisely, these refer-
smallest number of nodes (edges) that must be deleted from ences assume the following: (a) Each transceiver is located
the graph so that the resulting graph is disconnected. For at some point of d-dimensional Euclidean space. (b) For
example, a tree has node and edge connectivities equal to 1 any pair of transceivers u and v, p(u, v) = p(v, u) =
while a simple cycle has node and edge connectivities equal the Euclidean distance between the locations of u and v.
to 2. When the node (edge) connectivity of a graph is greater For a justification of this model, see [16]. They show
than or equal to k, the graph is said to be k-node connected that the D IR , S TRONGLY C ONNECTED, T OTAL P problem
(k-edge connected). Given an undirected graph, polynomial is NP-hard when transceivers are located in 3-dimensional
algorithms are known for finding its node and edge connec- space. They also present an approximation algorithm with
tivities [30]. a performance guarantee of 2 for the problem in any met-
The main results of this paper use the following definition. 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, general approach. As an illustration of our general ap-
T OTAL P problem where there is an additional constraint on proach, we present a constant factor approximation algo-
the diameter of the resulting undirected graph. Clementi rithm for the U NDIR , 2-N ODE C ONNECTED, T OTAL P
et al. [7] show that the 2-dimensional version of the problem. No approximation algorithm was previously
D IR , S TRONGLY C ONNECTED , T OTAL P problem remains known for this problem. In analyzing this approxi-
NP-hard. They also show that the 2-dimensional version mation algorithm, we use some properties of critically
with a diameter constraint can be efficiently approximated to 2-node connected graphs [10,21,32]. By a minor modi-
within some constant factor and that the 3-dimensional ver- fication to this approximation algorithm, we also obtain
sion does not have a polynomial time approximation scheme. a constant factor approximation algorithm for producing
Under a slightly different model, where there is an explicit re- 2-edge-connected graphs. As in the case of minimizing
lationship between the transmission power and distance, ref- maximum power, our general heuristic for approximating
erences [2,3] study topology control problems for connectiv- total power is also applicable to graph properties specified
ity properties. The complexity of several problems under this by proper functions.
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 an implementation of the above approximation algorithm
ratio is improved to 5/3 in a journal submission based on [3]. and compare its performance with an algorithm discussed
Additional related work may be found in [13,1820, in [23].
22,31].
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 LLOYD ET AL.
Theorem 4.1. For any monotone and polynomial time tes- isfies the given proper function. Moreover, it is easy to see
table graph property P, the problems U NDIR , P, M AX P and that any graph property specified using a proper function is
D IR , P, M AX P can be solved in polynomial time. a monotone property. Thus, our results apply to this class of
network design problems as well.
Proof. We will present the proof for D IR , P, M AX P. (The We now present a polynomial time approximation scheme
proof for U NDIR , P, M AX P is virtually identical.) for U NDIR , P, M AX P and D IR , P, M AX P problems. As a
Consider an instance of D IR , P, M AX P. By lemma 4.1, compensation for the slight deviation from the optimal value,
there is an optimal solution in which every transceiver is as- this approach has the potential to reduce the running time sub-
signed the same power value. We can estimate the number of stantially.
candidate optimal power values as follows. Let T denote the
set of all transceivers in the system and let |T | = n. Consider Theorem 4.2. Let P be a monotone graph property that can
any transceiver u T . The number of different power values be tested for an n-node graph in time FP (n). For any fixed
that need to be considered for u is at most n 1, since at most > 0, the problems U NDIR , P, M AX P and D IR , P, M AX P
one new power value is needed for each transceiver in T {u}. can be approximated to within the factor 1 + in O((n2 +
Therefore, for all of the n transceivers, the total number of FP (n)) log log(max / min)) time, where max and min are re-
candidate power values to be considered is n(n 1) = O(n2 ). spectively the maximum and minimum power threshold val-
For each candidate power value, the corresponding di- ues in the given problem instance.
rected graph can be constructed in O(n2 ) time. Let FP (n)
denote the time needed to test whether property P holds for Proof. We will present the proof for D IR , P, M AX P.
a directed graph with n nodes. Thus, the time needed to Since the number of power threshold values is O(n2 ), the val-
test whether property P holds for each candidate solution ues of min and max can be found in O(n2 ) time. Note that
value is O(n2 + FP (n)). An optimal solution can be ob- for any candidate power value (which is assigned to all the
tained by sorting the O(n2 ) candidate solution values and us- nodes), testing whether P holds for the induced graph can be
ing binary search to determine the smallest value for which done in O(n2 + P(n)) time.
property P holds. Since the number of candidate solu- Let k be the smallest integer such that (1 +)k min max.
tion values is O(n2 ), the time taken by the sorting step is Thus, k = O(log(max / min)). Consider the following set of
O(n2 log n). The binary search would try O(log n) candidate k + 1 power values: {min, (1 + ) min, (1 + )2 min, . . . ,
solution values and the time spent for testing each candidate (1 + )(k1) min, max}. By doing a binary search on this
is O(n2 + FP (n)). Thus, the total running time of this algo- set, we can determine the smallest integer j such that
rithm is O((n2 + FP (n)) log n). Since FP (n) is a polynomial, the power value (1 + )j min causes the induced graph to
the algorithm runs in polynomial time. have the property P. The binary search uses O(log k) =
O(log log(max / min)) calls to the algorithm for testing P.
As an illustration of the above theorem, let P denote the Thus, the running time of the algorithm is O((n2 + FP (n)) log
property 2-N ODE C ONNECTED for undirected graphs. It is log(max / min)).
known that this property can be tested in O(n2 ) time for a Further, since j is the smallest value for which the power
graph with n nodes [30]. For this property, the general al- value (1 + )j min causes the induced graph to have the prop-
gorithm outlined in the proof of theorem 4.1 yields an algo- erty P, the optimal value must be at least (1 + )(j 1) min.
rithm with a running time of O(n2 log n). This running time Thus, the solution found by the algorithm is within a factor
matches the time of the algorithm given in [23]. However, it (1 + ) of the optimal value.
should be noted that the algorithm in [23] not only finds an op-
timal solution but also reduces the power of each transceiver When the ratio max / min is substantially smaller than 2n ,
so that the power levels are minimal. There is no increase in the above approximation scheme reduces the number of calls
their asymptotic running time. to the property testing algorithm to a value that is asymptoti-
Instead of requiring the entire graph to be connected, one cally smaller than O(log n).
may require connectivity only for a specified subset of the
nodes. Such a requirement arises in the context of multicas- 4.2. Difficulty of generalizing to nonmonotone properties
ting (see, for example, [27]), where the subset of nodes in-
cludes the sender and all the intended receivers. Connect- We now show that there is a natural nonmonotone graph
edness of a specified subset of nodes can be seen to be a property for which the problem of minimizing the maximum
monotone property. Thus, the general approach presented power is NP-hard. As mentioned earlier, this result points out
above leads to a polynomial time algorithm for this prop- that if the monotonicity requirement is omitted, then an effi-
erty as well. In fact, the result extends to large class of cient algorithm for minimizing maximum power may not be
network design problems that can be specified using proper possible.
functions [1,12]. As noted in [12], the Steiner tree problem The property that we use for this purpose is G IS A
and the Steiner forest problem can be specified using this for- T REE. Surprisingly, we show that this property makes the
malism. Given a network and a proper function specifica- topology control problem NP-complete even without any
tion, it is easy to test in polynomial time if the network sat- minimization objective. The proof of lemma 4.2 utilizes a
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 23
reduction from Exact Cover by 3-Sets (X3C), which is known outline for such an approximation algorithm for topology con-
to be NP-complete [11]. trol problems of the form U NDIR , P, T OTAL P. We observe
that this general outline encompasses the approximation al-
Lemma 4.2. To determine whether there is a power assign- gorithm for U NDIR , 1-N ODE C ONNECTED, T OTAL P pre-
ment such that the resulting undirected graph G is a tree is sented in [16]. Based on the general outline, we also develop
NP-complete. an approximation algorithm with a constant performance
guarantee for U NDIR , 2-N ODE C ONNECTED , T OTAL P.
Proof. See appendix A. A slight modification of this approximation algorithm yields
an approximation algorithm for the problem of obtaining a
Theorem 4.3. There is a nonmonotone property P for which 2-edge-connected graph while minimizing total power.
U NDIR , P, M AX P is NP-hard. In presenting our general scheme, we assume (as done in
section 4.1) that the property P to be satisfied by the graph is
Proof. Let P denote the property G IS A T REE. monotone and that it can be tested in polynomial time. We
The NP-hardness of U NDIR , P, M AX P follows from also assume symmetric power thresholds as in [7,8,16]; that
lemma 4.2. is, for any pair of transceivers u and v, the power thresholds
p(u, v) and p(v, u) are equal.
4.3. Difficulty of minimizing the number of nodes of An outline for our general approximation algorithm (called
maximum power 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-
An extension of U NDIR , P, M AX P for monotone graph nomial time. The time complexity of step 2 depends cru-
properties is explored in this section. While such problems cially on the property P. For some properties such as 1-
can be solved efficiently, our algorithm in section 4.1 assigns N ODE C ONNECTED, step 2 can be done in polynomial time.
the maximum power value to all of the nodes. From a prac- For other properties such as 2-N ODE C ONNECTED, step 2
tical point of view, it is important to reduce the number of cannot be done in polynomial time, unless P = NP [11].
nodes with maximum power without affecting the required In such cases, an efficient algorithm that produces an ap-
property. In this section, we show that this additional require- proximately minimum solution can be used in step 2. The
ment renders the problem NP-hard even for certain monotone following theorem proves the correctness of the general ap-
graph properties. A formal statement of the decision version proach and establishes its performance guarantee as a func-
of the problem is as follows. tion of some parameters that depend on the property P and
the approximation algorithm used in step 2 of the general out-
Max-power users. line.
Instance: A positive integer M, a positive number P (max-
imum allowable power value), a node set V , a power Theorem 5.1. Let I be an instance of U NDIR , P, T OTAL P
threshold value p(u, v) for each pair (u, v) of transceivers where P is a monotone property. Let OPT(I ) and GTP(I )
and a graph property P. denote respectively the total power assigned to the nodes in
an optimal solution and in a solution produced by Heuristic
Question: Is there a power assignment where the power
G EN -T OTAL -P OWER for the instance I .
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 (i) The graph G resulting from the power assignment pro-
the resulting undirected graph G satisfies P? duced by the heuristic (i.e. step 3) satisfies property P.
(ii) Consider the complete graph Gc (V , Ec ) constructed in
Theorem 4.4. There is a monotone and polynomial time
step 1 of the heuristic. Let H (V , EH ) be an edge sub-
testable property P for which the problem Max-power users
graph of Gc with minimum total edge weight satisfying
is NP-complete.
property P and let W (H ) denote the total edge weight
of H . Let step 2 of the heuristic produce an edge sub-
Proof. See appendix B. graph G (V , E ) of G with total edge weight W (G ).
Suppose there are quantities > 0 and > 0 such that
5. A general approach for minimizing total power (a) W (H ) OPT(I ) and (b) W (G ) W (H ). Then,
GTP(I ) 2OPT(I ). That is, Heuristic G EN -T OTAL -
5.1. Approximating minimum total power P OWER provides a performance guarantee of 2.
Topology control problems in which the minimization objec- Before proceeding to the proof of this result, we illus-
tive is total power tend to be computationally intractable. For trate its use by discussing how the 2-approximation algorithm
example, the problem is NP-hard even for the (simple) prop- presented in [16] for the U NDIR , 1-N ODE C ONNECTED,
erty 1-N ODE -C ONNECTED [16]. A common way of coping T OTAL P problem can be derived from the above general out-
with such problems is to develop polynomial time approxima- line. In step 2 they use an efficient algorithm for constructing
tion algorithms for them. In this section, we present a general 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 5.2.1. An approximation algorithm for U NDIR , 2-N ODE
of a minimum spanning tree of Gc . Thus, using the notation C ONNECTED, T OTAL P
of theorem 5.1, = = 1 for their approximation algo- This section presents an approximation algorithm for the
rithm. Since 1-N ODE -C ONNECTED is a monotone property, U NDIR , 2-N ODE C ONNECTED, T OTAL P problem. The
it follows from theorem 5.1 that the performance guarantee of NP-hardness of this problem is established in [4]. Our
their algorithm is 2. 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
Proof of theorem 5.1. Part (i). The edge subgraph G (V , E ) C ONNECTED, T OTAL P with n transceivers. For each trans-
constructed in step 2 of the heuristic satisfies property P. We ceiver u, (u) denotes the power assigned to u in an optimal
show that every edge in E is also in the subgraph G induced solution. Further, OPT(I ) denotes the sum of the powers as-
by the power assignment produced in step 3. Then, even if G signed to the nodes in an optimal solution.
has other edges, the monotonicity of P allows us to conclude We obtain an approximation algorithm for the U NDIR ,
that G satisfies P. 2-N ODE C ONNECTED, T OTAL P problem from the outline
Consider an edge {u, v} with weight p(u, v) in E . Recall of figure 1 by using an approximation algorithm from [14] for
that p(u, v) is the minimum power threshold for the existence the minimum weight 2-N ODE -C ONNECTED subgraph prob-
of edge {u, v} and that the power thresholds are symmetric. lem in step 2 of the outline. This approximation algorithm
Since step 3 assigns to each node the maximum of the weights provides a performance guarantee of (2 + 1/n). Using the
of edges incident on that node, we have (u) p(u, v) and notation of theorem 5.1, we have (2 + 1/n).
(v) p(u, v). Therefore, the graph G induced by the We also show (see lemma 5.1 below) that for the com-
power assignment also contains the edge {u, v} and this com- plete edge weighted graph Gc (V , Ec ) constructed from the
pletes the proof of part (i). instance I in step 1 of the outline, there is an edge sub-
Part (ii). By conditions (a) and (b) in the statement of the graph G1 (V , E1 ) such that G1 is 2-N ODE -C ONNECTED and
theorem, we have W (G ) OPT(I ). We observe that the total weight W (G1 ) of the edges in G1 is at most (2
GTP(I ) 2 W (G ). This is because in step 3 of the heuris- 2/n)OPT(I ). Using the notation of theorem 5.1, this result
tic, the weight of any edge is assigned to at most two nodes implies that (2 2/n).
(namely, the endpoints of the edge). Combining the two in- Thus, once we establish lemma 5.1, it would follow
equalities, we get GTP(I ) 2OPT(I ), and this completes from theorem 5.1 that the performance guarantee of the re-
the proof of theorem 5.1. sulting approximation algorithm for the U NDIR , 2-N ODE
C ONNECTED, T OTAL P problem is 2(2 2/n) (2 + 1/n),
which approaches 8 asymptotically from below. The remain-
5.2. New approximation algorithms der of this section is devoted to the formal statement and proof
of lemma 5.1.
This section presents two new approximation algorithms de-
rived from the general approach outlined in figure 1. These Lemma 5.1. Let I denote an instance of the U NDIR ,
algorithms are for the two monotone properties 2-N ODE 2-N ODE C ONNECTED, T OTAL P problem with n transceiv-
C ONNECTED and 2-E DGE C ONNECTED respectively. The ers. Let OPT(I ) denote the total power assigned to the trans-
corresponding problems are denoted by U NDIR , 2-N ODE ceivers in an optimal solution to I . Let Gc (V , Ec ) denote
C ONNECTED, T OTAL P and U NDIR , 2-E DGE C ONNECTED , the complete graph constructed in step 1 of Heuristic G EN -
T OTAL P. 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.
The following lemma bounds the total weight W1 (G1 ) of is at most W1 (G1 ). By lemma 5.3, W1 (G1 ) is bounded
all the edges in G1 under the edge weight function w1 chosen by (2 2/n)OPT(I ). Therefore, W (G1 ) is also bounded
above. 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
Lemma 5.3. W1 (G1 ) (2 2/n)OPT(I ). all its edges is at most (2 2/n)OPT(I ). This completes the
proof of lemma 5.1.
Proof. As mentioned above, each edge of G1 is either a tree
edge or a back edge. Consider the tree edges first. For each The following is a direct consequence of the above discus-
tree edge {u, v}, where v is the parent of u, w1 (u, v) = (u). sion.
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 Theorem 5.3. There is a polynomial time approximation al-
in T ). The power value of the root r in the optimal solution, gorithm with a performance guarantee of 2(22/n)(2+1/n)
namely p , is not assigned to any tree edge (since the root (which approaches 8 asymptotically from below) for the
has no parent). Thus, the total weight of all of the tree edges U NDIR , 2-N ODE C ONNECTED, T OTAL P problem.
under the weight function w1 is bounded by OPT(I ) p .
Now consider the back edges. For each back edge {u, v}, 5.2.2. An approximation algorithm for U NDIR , 2-E DGE
where v is an ancestor of u, w1 (u, v) = (u). Since G1 C ONNECTED, T OTAL P
is critically 2-N ODE -C ONNECTED, by lemma 5.2, each node A result analogous to theorem 5.3 can also be obtained for
has at most one back edge to an ancestor. Thus, the weight U NDIR , 2-E DGE C ONNECTED, T OTAL P where the goal is
(u) is assigned to at most one back edge. Again, the to induce a graph that has the monotone property 2-E DGE
power value p of the root r in the optimal solution is not C ONNECTED. This problem has also been shown to be NP-
assigned to any back edge. Thus, the total weight of all of the complete in [4]. To obtain an approximation algorithm for
back edges under the weight function w1 is also bounded by this problem from the general framework, we use an ap-
OPT(I ) p . proximation algorithm of Khuller and Vishkin [15]. Their
Therefore, the total weight W1 (G1 ) of all of the edges approximation algorithm produces a 2-edge-connected sub-
in G1 under the edge weight function w1 is at most 2OPT(I ) graph whose cost is at most twice that of a minimum 2-edge
2p . Since p is the largest power value assigned to a node connected subgraph. In the notation of theorem 5.1, we have
in the optimal solution, p is at least OPT(I )/n. Hence, 2. Again using the notation of theorem 5.1, it is possible
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
identical to that for the 2-Node-Connected case, except that
The following lemma relates the weight w1 (u, v) of an we need an analog of theorem 5.2. Before stating this ana-
edge {u, v} to the power threshold p(u, v) needed for the ex- log, we have the following definition (which is analogous to
istence of the edge. definition 5.2).
Lemma 5.4. For any edge {u, v} in G1 , p(u, v) w1 (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
Proof. Consider any edge {u, v} in G1 . Since G1 is an edge conditions. (i) G is 2-E DGE -C ONNECTED. (ii) For every
subgraph of G (the graph induced by the chosen optimal edge e E, the subgraph of G obtained by deleting the edge e
power assignment), {u, v} is also an edge in G . Also, re- is not 2-E DGE -C ONNECTED.
call that the minimum power threshold values are symmetric.
Therefore, (u) p(u, v) and (v) p(u, v). Hence We can now state and prove the analog of theorem 5.2 for
min{ (u), (v)} p(u, v). The weight assigned to the critically 2-edge connected graphs.
edge {u, v} by the edge weight function w1 is either (u) or
(v). Therefore, w1 (u, v) min{ (u), (v)}. It follows Lemma 5.5. If a graph G is critically 2-E DGE -C ONNECTED
that w1 (u, v) p(u, v). then no cycle of G has a chord.
We are now ready to complete the proof of lemma 5.1. Proof. The proof is by contradiction. Suppose G is crit-
ically 2-E DGE -C ONNECTED but there is a cycle C=
Proof of lemma 5.1. Starting from an optimal power assign- v1 , v2 , . . . , vr , with r 4, with a chord {vi , vj }. Con-
ment to the instance I , construct the graph G1 (V , E1 ) as de- sider the graph G obtained from G by deleting the chord
scribed above. Since the graph Gc constructed in step 1 of {vi , vj }. We will show that G is 2-E DGE -C ONNECTED, thus
the heuristic (Figure 1) is a complete graph, every edge in contradicting the assumption that G is critically 2-E DGE -
G1 is also in Gc . Consider the edge subgraph G1 (V , E1 ) of C ONNECTED.
Gc where E1 = E1 . Since G1 is 2-N ODE -C ONNECTED, To show that G is 2-E DGE -C ONNECTED, it suffices to
so is G1 . By lemma 5.4, for each edge {u, v} in E1 , show that G cannot be disconnected by deleting any single
p(u, v) w1 (u, v). Therefore, the total weight W (G1 ) edge. Consider any edge {x, y} of G , and let G denote the
of all of the edges in G1 under the edge weight function p graph created by deleting {x, y} from G . Since we deleted
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 27
only one edge from G , all the nodes of the cycle C are in there are no existing approximation algorithms3 specifically
the same connected component of G . Thus, if we create the for U NDIR , 2-N ODE C ONNECTED, T OTAL P, in the exper-
graph G1 by adding the chord {vi , vj } to G , the two graphs iments described here we compare the performance of our
G1 and G have the same number of connected components. algorithm with Ramanathan and Rosales-Hains algorithm
However, G1 is also the graph obtained by deleting the edge in [23]. Recall that their algorithm finds an optimum so-
{x, y} from G. Since G is 2-E DGE -C ONNECTED, G1 is con- lution for U NDIR , 2-N ODE C ONNECTED, M AX P in which
nected. Thus, G is also connected. We therefore conclude the power level of each node is minimal. Our experiments
that G is 2-E DGE -C ONNECTED, and this contradiction com- were conducted using a customized implementation on both
pletes the proof of lemma 5.5. randomly generated networks and on networks derived from
realistic data generated by the TRANSIMS project [29].
The remainder of the proof to show that (2 1/n)
is identical to that for the 2-Node-Connected case. With 6.1. Randomly generated networks
(2 1/n) and 2, the following theorem is a di-
rect consequence of theorem 5.1. 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
Theorem 5.4. There is a polynomial time approximation the Log-distance Path Loss Model:
algorithm with a performance guarantee of 8(1 1/n)
(which approaches 8 asymptotically from below) for the Gt Gr 2 d
P L(d) = 10 log10 + 10 log ,
U NDIR , 2-E DGE C ONNECTED, T OTAL P problem. (4)2 d02
10
d0
where is the path loss exponent, d0 is the close-in reference
Our performance guarantee results are somewhat pes-
distance, is the radio wavelength, Gt is the transmitter an-
simistic since they are derived from a general framework.
tenna gain, Gr is the receiver antenna gain, and d is the sepa-
Using a different method of analysis, Calinescu and Wan [4]
ration distance between transmitter and receiver (see [24] for
have shown recently that both of our heuristics provide a per-
detailed descriptions of these parameters). All of the parame-
formance guarantee of 4.
ters are chosen to emulate a 2.4 GHz wireless radio, and if d
As in the case of minimizing maximum power, our general
is less than a certain threshold, the transmission power is set
framework for minimizing total power can also be used to to the minimum transmission power of 1 dBm.
obtain polynomial time approximation algorithms for topol- The experiments are conducted by varying the density of
ogy control problems wherein the connectivity requirements the network and the spatial distribution of the nodes. In total
are specified using proper functions. To obtain this result, there are 38 sets of experiments, and 10 trials are run on each
we use the general method outlined in [1,12] as the algo- set. Each of the results we cite is the average over the 10 trials.
rithm in step 2 of our general heuristic. The method of [1,12] The node density varies from 0.625 node/sq mile to
gives a 2-approximation algorithm for network design prob- 6.25 nodes/sq mile (10 nodes to 100 nodes) in a 4 mile by
lems specified using proper functions. Using the notation of 4 mile area. The experiments are conducted using two node
theorem 5.1, = 2. It is also straightforward to show that the distributions: one uniform and one skewed. Specifically, in
complete graph constructed in step 1 of our heuristic has a re- the uniformly distributed networks, all nodes are placed us-
quired subgraph of weight at most the optimal solution value. ing a random uniform distribution. In the networks with a
In other words, 1. Thus, we obtain a 4-approximation skewed distribution, the network area is equally divided into
algorithm for the general class of problems defined in [1,12]. a 2 by 2 grid, with 80% of the nodes uniformly distributed
An important example of a problem in this class is the Steiner in two diagonal squares, and the other 20% of the nodes uni-
variant of connectivity, where the goal is to assign power lev- formly distributed in the other two diagonal squares.
els so as to connect only a specified subset of nodes of a graph In each experiment, after generating a placement of the
rather than all the nodes. An approximation
algorithm with a nodes, both our approximation algorithm (M IN T OTAL) and
performance guarantee of (1+ln 3 ) is known for the Steiner the algorithm of [23] (M IN M AX) are run on the network
tree problem in graphs [25]. Thus, using this approximation consisting of those nodes. Each algorithm assigns powers
algorithm, our approach yields a (2 + ln 3)-approximation for to nodes such that the resulting network is 2-N ODE C ON -
the Steiner variant. 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.
6. Experimental results Prior to discussing the results, we first provide figure 3 that
shows the actual topologies for one simulated network with
In the preceding section, we showed that our algorithm for 60 nodes. Figures 3(a) and (b) are respectively the topologies
U NDIR , 2-N ODE C ONNECTED, T OTAL P provides a con- 3 While this paper was under review, an algorithm that provides a con-
stant factor approximation. In this section, we report on stant factor approximation for the geometric version of U NDIR, 2-N ODE
the experimental performance of this algorithm. Since C ONNECTED, T OTAL P was presented in [4].
28 LLOYD ET AL.
resulting from our approximation algorithm (M IN T OTAL) about 40%76% of the maximum power using M IN M AX al-
and Ramanathan and Rosales-Hains algorithm (M IN M AX). gorithm of [23], and about 25%64% using our algorithm. In
other words, for a given average node density, the maximum
6.1.2. Experimental results and discussion power in a skewed network is higher than that in a uniformly
In reporting our experimental results, we plot four different distributed network, while the average power in the skewed
quantities: (i) average power assigned to a node, (ii) maxi- network is lower. The reason is that in a skewed network the
mum power assigned to any node, (iii) average degree of a node density varies significantly from region to region. With
node and (iv) maximum degree of all of the nodes. The ex- a larger number of nodes in a smaller area, the average dis-
perimental results on power and node degree are shown in tance between two nodes is less, hence the required power
figure 4. levels are, on the average, smaller.
In figures 4(a) and (c), Min Max AVG (Min Max As a general rule, smaller is better in regard to node de-
MAX) and Min Total AVG (Min Total MAX) are grees in the network induced by the power assignments (e.g.,
the average (maximum) power using the M IN M AX and increases spatial spectrum reuse). In that context, in the case
M IN T OTAL algorithms, respectively. where nodes are uniformly distributed, the average (maxi-
mum) degree of the network with power assigned by our
In figures 4(b) and (d), Min Max AVG (Min Max M IN T OTAL algorithm is consistently smaller than the aver-
MAX) and Min Total AVG (Min Total MAX) are the age (maximum) degree of the network with power assigned
average (maximum) degrees using the M IN M AX and the by the M IN M AX algorithm in [23]. When using either of
M IN T OTAL algorithms, respectively. the algorithms, the average degree does not vary much as the
Figure 4 illustrates the results on power and node de- network density changes. Specifically, the average degree is
gree. In the cases where nodes are uniformly distrib- around 2.73 using our algorithm, which is very close to the
uted, our M IN T OTAL algorithm consistently outperforms the smallest possible degree, since in a 2-node-connected graph,
M IN M AX algorithm in [23] in regard to average power by the degree of each node must be at least 2.
5%19%. This improvement increases as the density of the The results in regard to node degrees under the skewed
network increases. In contrast, the maximum power assigned node distribution are similar to those for the uniform case.
by our algorithm is 14%37% larger than that of [23]. The
average power is about 60%83% of the maximum power 6.2. The TRANSIMS networks
using the M IN M AX algorithm, and about 39%70% using
our algorithm. Those numbers decrease as the density of the In addition to the randomly generated networks, we also con-
network increases, which implies that the average power de- ducted an experimental study on a more realistic network
creases faster than the maximum power, and a smaller per- obtained from the TRANSIMS Portland Study by the Los
centage of nodes have the maximum power as the network Alamos National Laboratory [29]. This data set contains
density increases. locations of 1716 nodes over a 3 km by 3 km area. The
In skewed placements of nodes, our M IN T OTAL algorithm locations were generated by carrying out a detailed simula-
outperforms the M IN M AX algorithm with respect to aver- tion of the traffic in the Portland, OR, metropolitan area us-
age power by 6%25%. We observe that the difference be- ing the TRANSIMS simulation tool. Since the running time
tween average power and maximum power is larger in skewed of our algorithm would be prohibitively high if run on all
placements than in uniform placements. The average power is 1716 nodes, we selected two characteristic areas of this net-
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 29
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
almost on a straight line, are connected to each other, and 7. Directions for future research
the edges between them overlap in the figure.
Our experiments with TRANSIMS data show that topol- Our work provides several directions for future research.
ogy control can significantly reduce the average transmis- First, it will be of interest to investigate whether approxima-
sion power. In figure 5(a), where each node has a transmis- tion algorithms with performance guarantees better than 4 can
sion range of 75 meters, the induced graphs in areas 1 and 2 be developed for inducing 2-node connected and 2-edge con-
are not even connected. After the application of M IN T O - nected graphs. Second, it will be useful to consider topology
TAL algorithm, the induced graphs in both areas are 2-node- control problems for other graph properties. In that direc-
connected and the average range for the two areas is reduced tion, some complexity and approximation results for proper-
to 55.07 meters and 51.95 meters, respectively. ties such as bounded diameter and lower bounds on node de-
In area 1, the average range assigned by our M IN T OTAL grees under the objective of minimizing total power are pre-
algorithm is 18.7% lower than that assigned by the M IN M AX sented in [17]. A third direction is to investigate the behav-
algorithm, while the maximum transmission range of our al- ior of topology control problems under the asymmetric power
gorithm is 22.2% higher than the M IN M AX algorithm. The threshold model. Some results in that direction are also pre-
induced maximum and average degrees are always smaller sented in [17]. Finally, it will be of interest to develop distrib-
using the M IN T OTAL algorithm than using the M IN M AX uted versions of algorithms for topology control problems.
algorithm. For area 2, our M IN T OTAL algorithm assigns av- References [4,18,23] present some results along that direc-
erage range 29.4% lower, but 45% higher maximum range. tion.
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 Acknowledgements
constantly outperforms the M IN M AX algorithm on average
power (transmission range), and the margin is larger when We thank the reviewers for a careful reading of the manuscript
the network is more skewed. and for providing valuable suggestions.
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 31
Appendix A. Proof of lemma 4.2 (1 j m). It can be seen that the graph G resulting from
this power assignment contains only the following edges:
We first restate the lemma.
(a) The edge {R, vj }, for each j , 1 j m.
Lemma 4.2. To find a power assignment such that the result- (b) For each node vj whose corresponding set Cj is in C ,
ing undirected graph G is a tree is NP-complete. there are edges from vj to the three nodes corresponding
to the elements in Cj .
By abuse of terminology, we use U NDIR , T REE, to de-
note this problem. The NP-hardness of this problem is es- By choosing R as the root and using the fact that C is an
tablished using a reduction from the X3C problem defined exact cover, it can be verified that G is a tree: the root node R
below. is adjacent to each of the set nodes; and, each element node
appears as one of the three children of a set node correspond-
Exact cover by 3-sets (X3C). ing to a subset in the collection C .
Only if. Now, suppose the U NDIR , T REE, instance has
Instance: A set S = {x1 , x2 , . . . , xn } of elements, where a solution. Let p (x) denote the power assigned to node x and
n = 3r for some integer r; a collection C = {C1 , C2 , . . . , let G denote the graph induced by the power assignment.
Cm } of subsets of S such that |Cj | = 3, 1 j m. We first observe that p (R) 1; otherwise, R would be
Question: Does C contain an exact cover for S, that is, is an isolated node and thus G cannot be a tree. Similarly,
there a subcollection C of C such that the sets in C are p (vj ) 1 for every set node vj and p (ui ) 2 for every
pairwise disjoint and their union is equal to S? 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
Note that whenever there is a solution to an instance of power assigned is at least 2. Therefore, there are two cases to
X3C, the number of sets in the solution is exactly r (i.e. n/3). consider:
Case 1. The maximum power assigned is 2. Let X =
Proof of lemma 4.2. In the U NDIR , T REE, problem, we {vjk : p (vjk ) = 2}. We claim that the collection C =
are given a collection of nodes, and a (symmetric) power {Cjk : vjk X} is an exact cover for S. We prove this by first
threshold p(u, v) for each pair of nodes. The question is showing that each element xi appears in some subset of C .
whether there exists a power assignment such that the graph To see this, we note that the graph G is connected (since it is
induced by the power assignment is a tree. a tree). Thus, there is at least one edge from the element node
It is easy to see that U NDIR , T REE, is in NP since one ui (corresponding to element xi ) to some other node of G.
can guess a power assignment and verify in polynomial time Since the maximum power assigned to any node is 2 and the
that the resulting graph is a tree. We prove the NP-hardness power threshold for the element node ui to have an edge to R
of the problem by a reduction from X3C (defined above). or an edge to any other element node is 3, ui must be adjacent
Given an instance I of X3C consisting of a set S with n to a set node vj . Further, because the threshold values are
elements and a collection C of m subsets, we construct an symmetric, p (vj ) = 2. Thus, vj X and the correspond-
instance I of the U NDIR , T REE, problem as follows. The ing subset Cj is in C . Hence, each element appears in some
node set V of I contains a total of n + m + 1 nodes: There is subset in the collection C .
one node (called an element node) ui corresponding to each We now show that the subsets in the collection C are pair-
element xi of S (thus, there are totally 3r element nodes), wise disjoint. Suppose some pair of subsets Ca and Cb in C
one node (called a set node) vj corresponding to each set Cj have a common element xi . By our choice of C , the power
of C (thus, there are totally m set nodes), and a special node values assigned to the corresponding set nodes va and vb are
(called the root node) denoted by R. The power thresholds both 2. Further, the power assigned to node ui is also 2. Thus,
are chosen as follows. (The reader should bear in mind that
in the graph G, ui is adjacent to both va and vb . As observed
the power thresholds are symmetric; that is, for any pair of
earlier, the root node R is adjacent to both va and vb . Now,
nodes u and v, p(u, v) = p(v, u).)
the four edges {R, va }, {va , ui }, {ui , vb } and {vb , R} create a
p(R, vj ) = 1, 1 j m, 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
p(ui , vj ) = 2, if xi Cj , 1 i n, 1 j m.
exact cover for S.
For all other pairs of nodes, the power thresholds are set Case 2. The maximum power assigned is 3. First, note
to 3. This completes the construction of the instance I of that at most two nodes can have power 3, since if three nodes
U NDIR , T REE, . It is easy to verify that the construction have power 3, then they are mutually adjacent, and thus G is
can be carried out in polynomial time. We now argue that not a tree.
there is a solution to the U NDIR , T REE, instance if and Second, if the power assignment is as in the following
only if there is a solution to the X3C instance. cases, we argue that there is an equivalent assignment in
If. Suppose the X3C instance has a solution C . We choose which the maximum power is 2. These cases are: only one
the following power assignment: p (R) = 1, p (ui ) = 2 (1 node has power 3; R and one set node vi have power 3; and,
i n), p (vj ) = 2 if Cj is in C and p (vj ) = 1 otherwise one element node ui and one set node vj have power 3 where
32 LLOYD ET AL.
xi Cj . In any of these cases, the resulting graph G has Question: Does there exist a subcollection C C, such
no edge with power threshold 3, so an assignment with max- that |C | K and the union of the sets in C is equal to S?
imum power 2 can be obtained by reducing the power level
Let P be the property THE DIAMETER OF G IS LESS
of the nodes with power 3 while keeping the assignments to
THAN OR EQUAL TO 6. This property implies that in G,
all of the other nodes unchanged. The induced graph does not
each node is at most 6 hops away from any other node. Ob-
change. Thus, the new assignment is a solution with maxi-
viously, P is monotone, and can be tested in O(n3 ) time by
mum power 2 to the instance of U NDIR , T REE, . Follow-
using the FloydWarshall algorithm, where n is the number
ing the argument in case 1, a solution to X3C can be con-
of nodes in the graph [9]. Thus, Max-power users is in NP. To
structed.
prove the NP-hardness, we provide a reduction from S C.
Finally, we claim that there are no such valid power as-
signments in the remaining cases (i.e. R and ui have power 3; Given an instance I of S C, an instance I of Max-power
vi and vj have power 3; ui and uj have power 3; or, ui and vj users is constructed as follows: For each element xi of S,
have power 3 where xi / Cj ). The reasons are the following: 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 ,
1. If two set nodes vi and vj have power 3, then the edges s2 , s3 . The power threshold function p is defined as follows.
{R, vi }, {R, vj } and {vi , vj } form a cycle. (It should be noted that the power thresholds are symmetric.)
2. If the root node R and one element node ui have po- p(ui , vj ) = 1, if xi Cj ,
wer 3, the edge {R, ui } is in G. Therefore, edge{ui , vj },
p(w, vj ) = P , 1 j m,
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 , p(w, s1 ) = p(s1 , s2 ) = p(s2 , s3 ) = 1.
therefore each vj with power 2 is adjacent to exactly 3 el- For all other pairs of nodes x and y, p(x, y) = P + 1.
ement nodes. No two set nodes can be adjacent to the The value of M is set to K + 1. This completes the con-
same element node, otherwise those three nodes and R struction of an instance I of Max-power users. It is clear that
form a cycle. Hence, totally 3k (where k is the number the construction can be done in polynomial time. Now, we
of set nodes with power 2) element nodes are adjacent to show that there is a solution to the Max-power users instance
some set node. Further, no two element nodes can be ad-
if and only if there is a solution to S C.
jacent to each other since the power thresholds between
If. Suppose C is a solution to the instance of S C. We
such nodes are 3. Thus, there are 3k + 1 element nodes.
construct a power assignment p as follows.
This is a contradiction since we know in this instance of
U NDIR , T REE, , the number of element nodes is a mul- p (w) = P ,
tiple of 3. p (vi ) = P , if Ci C
3. If two element nodes ui and uj have power 3, the edge (Note: there are at most K such nodes),
{ui , uj } is in G. Recall that all set nodes must be adjacent p (x) = 1 for any other node x.
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 We now argue that p is a solution to the instance of Max-
element nodes are adjacent to some set node. So, together power users. Obviously, the maximum power assigned is P
with uj , there are 3k + 1 element nodes a contradiction. 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
4. If one element node ui and one set node vj have power 3,
diameter at most 6), we show that w is within 3 hops of every
where xi / Cj , then ui is adjacent to vj . Therefore, there
other node. This follows from the following observations.
are 4 nodes adjacent to vj , which are ui and three element
nodes whose corresponding elements are in set Cj . Hence, 1. Nodes s1 , s2 , and s3 are respectively 1, 2, and 3 hops away
there are totally 3k+1 element nodes a contradiction. from w.
This completes the proof of the case 2 as well as that of 2. For any Ci C, if Ci C , then the edge {vi , w} E.
lemma 4.2. Hence, node vi is only one hop away from w.
3. For any xi S, node ui is 2 hops away from w, since ui
Appendix B. Proof of theorem 4.4 is adjacent to some node vi that has an edge to w. (Other-
wise, C does not cover the element xi .)
Proof. We use a reduction from S ET COVERING (S C), a 4. For any Ci C, if Ci / C , then vi is 3 hops away from w,
well-known NP-complete problem [11]. since vi is adjacent to some uj .
Only if. Suppose we have a power assignment p that
Set covering (Sc).
is a solution to the instance of Max-lpower users, and that
Instance: A set S = {x1 , x2 , . . . , xn }, a collection C = G(V , E) is the resulting graph. We construct a solution to the
{C1 , C2 , . . . , Cm }, where Ci is a subset of S (1 i m), S C instance as follows. If there is an edge between w and
and a positive integer K m. vi in E (there are at most M 1 such edges), then include
ALGORITHMIC ASPECTS OF TOPOLOGY CONTROL PROBLEMS 33
set Ci in C . We claim that C is a solution to the instance of ruary 1997) Lecture Notes in Computer Science, Vol. 1200 (Springer,
S C. Since |C | M 1 = K, we need only show that C New York, 1997) pp. 363374.
covers S. Since the diameter of G(V , E) 6, s3 is at most [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-
6 hops away from any other node. It follows that w must be ric power thresholds, in: Proc. of the Internat. Conf. on Ad Hoc and
within 3 hops of every other node. For each vi , if edge {vi , w} Wireless Networks (ADHOC-NOW03), eds. S. Pierre, M. Barbeau and
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 puter Science, Vol. 2865 (Springer, New York, 2000) pp. 187198.
an element xi S that is not in any set of C . Then, ui is not [18] M. Kubisch, H. Karl, A. Wolisz, L. Zhong and J. Rabaey, Distributed
algorithms for transmission power control in wireless sensor networks,
adjacent to any node vj that is one hop away from w. Thus, in: Proc. of the IEEE Wireless Communications and Networking Con-
ui must be adjacent to some node vj that is at least 3 hops ference (WCNC 2003), New Orleans, LA (March 2003) pp. 558563.
away from w. Thus, node ui is at least 4 hops away from w [19] L. Li and J.Y. Halpern, Minimum energy mobile wireless networks
a contradiction. This completes the proof of theorem 4.4. 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
References cone-based distributed topology control algorithm for wireless multi-
hop networks, in: Proc. of the ACM Principles of Distributed Comput-
[1] A. Agrawal, P. Klein and R. Ravi, When trees collide: An approxima- ing Conference (PODC01) (August 2001) pp. 264273.
tion algorithm for the generalized Steiner problem on networks, SIAM [21] M.D. Plummer, On minimal blocks, Trans. AMS 134 (October
J. Computing 24(3) (1995) 440456. December 1968) 8594.
[2] D.M. Blough, M. Leoncini, G. Resta and P. Santi, On the symmetric [22] V. Radoplu and T.H. Meng, Minimum energy mobile wireless net-
range assignment problem in wireless ad hoc networks, in: Proc. of the works, IEEE J. Selected Areas Commun. 17(8) (1999) 13331344.
2nd IFIP Internat. Conf. on Theoretical Computer Science, Montreal [23] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
(August 2002) pp. 7182. wireless networks using transmit power adjustment, in: Proc. of the
[3] G. Calinescu, I. Mandoiu and A. Zelikovsky, Symmetric connectivity IEEE INFOCOM 2000, Tel Aviv, Israel (March 2000) pp. 404413.
with minimum power consumption in radio networks, in: Proc. of the [24] T.S. Rappaport, Wireless Communications: Principles and Practice
2nd IFIP Internat. Conf. on Theoretical Computer Science, Montreal (Prentice-Hall, Englewood Cliffs, NJ, 1996).
(August 2002) pp. 119130. [25] G. Robins and A. Zelikovsky, Improved Steiner tree approximation in
[4] G. Calinescu and P.-J. Wan, Symmetric high connectivity with min- graphs, in: Proc. of the 11th Ann. ACMSIAM Symposium on Discrete
imum total power consumption in multihop packet radio networks, Algorithms (SODA 2000), San Francisco, CA (2000) pp. 770779.
in: Proc. of the Internat. Conf. on Ad Hoc and Wireless Networks [26] E.M. Royer, P. Melliar-Smith and L. Moser, An analysis of the opti-
(ADHOC-NOW03), eds. S. Pierre, M. Barbeau and E. Kranakis, Mon- mum node density for ad hoc mobile networks, in: Proc. of the IEEE
treal, Canada (October 2003) Lecture Notes in Computer Science, Internat. Conf. on Communication (ICC01), Helsinki, Finland (June
Vol. 2865 (Springer, New York, 2003) pp. 235246. 2001) pp. 857861.
[5] W. Chen and N. Huang, The strongly connecting problem on multihop [27] E.M. Royer and C. Perkins, Transmission range effects on AODV mul-
packet radio networks, IEEE Trans. Commun. 37(3) (1989) 293295. ticast communication, ACM Mobile Networks Appl. (Special Issue on
[6] X. Cheng, B. Narahari, R. Simha, M. Cheng and D. Liu, Strong min- Multipoint Communication in Wireless Networks) 7(6) (2002) 455
imum energy topology in wireless sensor networks: NP-completeness 470.
and heuristics, IEEE Trans. Mobile Computing 2(3) (2003) 248256. [28] H. Takagi and L. Kleinrock, Optimal transmission ranges for randomly
[7] A.E.F. Clementi, P. Penna and R. Silvestri, Hardness results for the distributed packet radio terminals, IEEE Trans. Commun. 32(3) (1984)
power range assignment problem in packet radio networks, in: Proc. 246257; also see in: Multiple Access Communications, Foundations
of the 3rd Internat. Workshop on Randomization and Approximation for Emerging Technologies, ed. N. Abramson (IEEE Press, New York,
in Computer Science (APPROX 1999) (July 1999), Lecture Notes in 1992) pp. 342353.
Computer Science, Vol. 1671 (Springer, New York, 1999) pp. 195208. [29] TRANSIMS, http://transims.tsasa.lanl.gov/.
[8] A.E.F. Clementi, P. Penna and R. Silvestri, The power range assign- [30] J. van Leeuwen, Graph algorithms, in: Handbook of Theoretical Com-
ment problem in packet radio networks in the plane, in: Proc. of the puter Science, Vol. A, ed. J. van Leeuwen (MIT Press/Elsevier, Cam-
17th Annual Symposium on Theoretical Aspects of Computer Science bridge, MA, 1990) chapter 10.
(STACS 2000) (February 2000) pp. 651660. [31] R. Wattenhofer, L. Li, P. Bahl and Y. Wang, Distributed topology con-
[9] T. Cormen, C. Leiserson, R. Rivest and C. Stein, Introduction to Algo- trol for power efficient operation in multihop wireless ad hoc networks,
rithms, 2nd ed. (MIT Press/McGraw-Hill, Cambridge, MA, 2001). in: Proc. of the IEEE INFOCOM 2001, Anchorage, Alaska (April
[10] G.A. Dirac, Minimally 2-connected graphs, J. Reine Angewandte 2001) pp. 13881397.
Math. 228 (1967) 204216. [32] D.B. West, Introduction to Graph Theory (Prentice-Hall, Englewood
[11] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide Cliffs, NJ, 1996).
to the Theory of NP-Completeness (Freeman, San Francisco, CA,
1979).
[12] M. Goemans and D.P. Williamson, A general approximation technique Errol L. Lloyd is a Professor of Computer and Infor-
for constrained forest problems, SIAM J. Computing 24(2) (1995) 296 mation Sciences at the University of Delaware. Pre-
317. viously he served as a faculty member at the Univer-
[13] L. Hu, Topology control for multi-hop packet radio networks, IEEE sity of Pittsburgh and as Program Director for Com-
Trans. Commun. 41(10) (1993) 14741481. puter and Computation Theory at the National Sci-
[14] S. Khuller and B. Raghavachari, Improved approximation algorithms ence Foundation. From 1994 to 1999 he was Chair
for uniform connectivity problems, J. Algorithms 21 (1996) 434450. of the Department of Computer and Information Sci-
[15] S. Khuller and U. Vishkin, Biconnectivity approximations and graph ences at the University of Delaware. Concurrently,
carvings, J. ACM 41 (1994) 214235. from 1997 to 1999 he was Interim Director of the
[16] L.M. Kirousis, E. Kranakis, D. Krizanc and A. Pelc, Power consump- University of Delaware Center for Applied Science
tion in packet radio networks, in: Proc. of the 14th Annual Sympo- and Engineering in Rehabilitation. Professor Lloyd received undergraduate
sium on Theoretical Aspects of Computer Science (STACS 97) (Feb- degrees in both computer science and mathematics from Penn State Uni-
34 LLOYD ET AL.
versity, and a Ph.D. in computer science from the Massachusetts Institute Ram Ramanathan is a Division Scientist at BBN
of Technology. His research expertise is in the design and analysis of al- Technologies. His research interests are in the area
gorithms, with a particular concentration on approximation algorithms for of wireless and ad hoc networks, in particular, rout-
computationally difficult problems. He has published over thirty journal pa- ing, medium access control and directional anten-
pers and numerous conference papers. In 1989 Professor Lloyd received an nas. He is currently the principal investigator for a
NSF Outstanding Performance Award, and in 1994 he received the Univer- project on architecture and protocols for opportunis-
sity of Delaware Faculty Excellence in Teaching Award. tic access of spectrum using cognitive radios. Re-
E-mail: elloyd@cis.udel.edu cently, he was one of one of two principal inves-
tigators for the DARPA project UDAAN (Utilizing
Directional Antennas for Ad hoc Networking), and
Rui Liu received the BS degree in mathematics from the co-investigator on NASAs Distributed Spacecraft Network project. Ram
Peking University, Beijing, China, in 1998; the MS is actively involved in the evolution of mobile ad hoc networking, and has
degree in applied mathematics from University of recently served on the program committees of the ACM MobiHoc sympo-
Delaware in 2000. He is a doctoral candidate in sium, and ACM Mobicom. He is on the editorial board of Ad Hoc Networks
computer and information sciences at the University journal. He has won three best paper awards at international conferences
of Delaware. His research interests include design at ACM SIGCOMM 92, at IEEE INFOCOM 96, and IEEE MILCOM 02.
and analysis of algorithms for combinatorial opti- Dr. Ramanathan holds a B.Tech from the Indian Institute of Technology,
mization problems, parallel and distributed comput- Madras, and an M.S. and a Ph.D. from the University of Delaware. He is
ing and computer networks. a senior member of the IEEE.
E-mail: ruliu@cis.udel.edu E-mail: ramanath@bbn.com
Madhav V. Marathe received his B. Tech in com- S.S. Ravi received his Ph.D. in computer science
puter science from IIT Madras and his Ph.D. in from the University of Pittsburgh in 1984. Since that
computer science from the University at Albany time, he has been with the Department of Computer
SUNY in 1994. Since that time, he has been with Science, University at Albany State University of
Los Alamos National Laboratory, where he currently New York, where he is currently a Professor. His re-
leads the Mathematics and Computer Science team search interests include design and analysis of algo-
in the Basic and Applied Simulation Science Group. rithms, mobile computing, fault-tolerance and VLSI
His research interests include mobile computing, design and testing.
simulation of large socio-technical systems such as E-mail: ravi@cs.albany.edu
transportation, telecommunication and markets and
design and analysis of algorithms.
E-mail: marathe@lanl.gov
Mobile Networks and Applications 10, 3545, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
the physical topology of the infrastructured network. For in- bility of establishing wireless connections with mobile termi-
stance, in nowadays cellular systems like GSM, the physical nals. A distinguished source node s V provides high speed
graph G is a tree, stations correspond to its leaves and the ad- services to the users moving along the network. We observe
jacency graph is an hexagonal grid (see, for instance, [22]). that, according to the wireless nature of the system, during
Standard ATM layouts must be constructed in order to estab- the handover phase mobile terminals do not necessarily have
lish a different VC for each station, but their performance is to move along the network G, but they can switch directly
evaluated by means of a further parameter, the virtual channel from one station to another, provided that they are adjacent in
distance, that measures the time needed to reconstruct virtual the physical space. It is thus possible to define a (connected)
channels during handover phases, that is when mobile termi- adjacency graph A = (U, F ), whose edges in F represent
nals switch between adjacent stations. More precisely, the adjacencies between stations.
distance between the virtual channels of two adjacent nodes A layout for G = (V , E) with source s V is a col-
is equal to the number of VPs that must be deleted and added lection of simple paths in G, termed virtual paths (VPs for
to one VC in order to obtain the other one. In order to make short), and a mapping that defines, for each station u U ,
the rerouting phase imperceptible to users and thus to obtain a unique virtual channel VC(u) connecting s to u, i.e., a sim-
a sufficient quality of service, the maximum distance between ple path from s to u in the virtual topology defined by the VPs
two virtual channels must be maintained as low as possible. of . In other words, VC(u) is a collection of VPs whose
Therefore, a natural combinatorial problem arises in which concatenation forms a path in G from s to u.
suitable trade-offs must be determined between the different
performance measures. Definition 2.1 [15]. The hop count h(u) of a node u U in
In [12] it has been shown that the layout construction prob- a layout is the number of VPs contained in VC(u), that
lem is intractable, that is NP-hard. Moreover, optimal lay- is |VC(u)|. The maximal hop count of is Hmax ()
out constructions are given when the physical and adjacency maxuU {h(u)}.
graphs are coincident and correspond to basic interconnection
networks, such as chains and rings. Such results hold under Definition 2.2 [15]. The load l(e) of an edge e E in a
the assumption that all the VCs induce shortest paths in the layout is the number of VPs that include e. The
underlying network. maximal load Lmax () of is maxeE {l(e)}.
In this paper we consider the determination of WATM lay-
outs for chains in the non-shortest path case in which the As already observed, when passing from a station u U
lengths of the paths induced by the VPs is not constrained. to an adjacent one v U , the virtual channel VC(v) must
We first show that deciding the existence of a layout with be reconstructed from VC(u) changing only a limited number
maximum hop count h, load l = 1 and distance d = 1 is of VPs. Once fixed VC(u) and VC(v), denoted as VC(u, v)
NP-complete even when the adjacency graph is a chain of the set of VPs in the subchannel corresponding to the longest
base stations with the source coinciding with one of its end- common prefix of VC(u) and VC(v), this requires the deletion
points. Moreover, such a hardness result is extended to every of all the VPs of VC(u) that occur after VC(u, v), plus the
fixed value of d. We then consider the case in which the phys- addition of all the VPs of VC(v) after VC(u, v). The number
ical and adjacency graph coincide with chains and provide of removed and added VPs, denoted as D(VC(u), VC(v)), is
optimal layout constructions for d 2. Finally, optimal lay- called the distance of VC(u) and VC(v) and naturally defines
out constructions are obtained also for any d within the class a channel distance measure d between pairs of adjacent nodes
of the so-called canonic layouts, that so far have been always in A.
shown to be the optimal ones.
The paper is organized as follows. In the next section we Definition 2.3 [12]. The channel distance of two nodes u
introduce the model, the notation and the necessary defini- and v, such that, {u, v} F (i.e., adjacent in A) is d(u, v) =
tions. In section 3 we provide the above mentioned hardness D(VC(u), VC(v)) = h(u)+h(v)2|VC(u, v)|. The maximal
results for the layout construction problem. In section 4 we distance of is Dmax () max{u,v}F {d(u, v)}.
provide the optimal layouts for chains when d = 2 and in
section 5 the optimal canonic ones for any d. Finally, in sec- It is now possible to give the following definition concern-
tion 6, we give some conclusive remarks and discuss some ing layouts for WATM networks.
open questions.
Definition 2.4. A layout with Hmax () h, Lmax ()
l and Dmax () d is a h, l, d-layout for G, s and A.
2. The WATM model
In the following we will always assume that all the VPs
We model the network as an undirected graph G = (V , E), of are contained in at least one VC. In fact, if such prop-
where nodes in V represent switches and edges in E are point- erty does not hold, the not used VPs can be simply removed
to-point communication links. In G there exists a subset of without increasing the performance measures h, l and d.
nodes U V constituted by base stations, i.e., switches Before concluding the section, let us remark that for prac-
adapted to support mobility and having the additional capa- tical purposes and quality of services guarantees, it makes
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS 37
sense to consider the case where d h. In fact, while a Starting from an instance of DP, we construct a network
little communication delay proportional to the hop count in G = (V , E ), a source s V and a chain adjacency graph
general can be tolerated, connections gaps due to rerouting of A = (U, F ) that admit a h, 1, 1-layout with h = 2k 1 if
virtual channels must not be appreciated by mobile users. On and only if there exist the requested k edge-disjoint paths in
the other hand, when d 2h, our model coincides with the the instance of DP.
classical one presented in [15] for standard ATM networks, Let G = (V , E ) be such that, given k 1 nodes
since the difference between any two virtual channels is al- w1 , . . . , wk1 not contained in the initial graph G, V = V
ways at most equal to 2h. {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,
3. Hardness and approximation results the source s = s1 .
In this section we show that constructing optimal dynamic Assume first that there is a 2k 1, 1, 1-layout for
layouts is in general an NP-hard problem, even when l = 1, G = (V , E ), s and A = (U, F ). By the considerations
d = 1 and the adjacency graph is a chain of stations with the at the beginning of this section, for each e F , a VP in
source being one of its endpoints. must exist connecting the two endpoints of e. We can assume
Before proving our results, let us briefly outline the basic that for each i, 1 i < k, the VP connecting ti to si+1 is
characteristics of a layout with maximum delay d = 1. ti , wi , si+1 , i.e., it is constituted by the new added path in
Given any two stations u1 , u2 U adjacent in A = G that goes from ti to si+1 through the new node wi . In
(U, F ), during an handover from u1 to u2 if d = 1 by de- fact, if this does not hold, it is possible to add to the new
finition only one VP can be modified. This means that either VP ti , wi , si+1 , deleting the old one and then, in order to
VC(u2 ) is a prefix of VC(u2 ) and thus VC(u2 ) is obtained keep l = 1, if there is another VP stepping through wi , it is
from VC(u1 ) adding a new VP from u1 to u2 , or vice versa. modified in such a way that its subpath between ti and si+1
In any case, a VP between u1 and u2 must be contained in the coincides with the old deleted VP.
layout. As a direct consequence, the virtual topology defined Therefore, since l = 1 and for all i, 1 i k, the VP
by the VPs of coincides with the adjacency graph A. More- between si and ti does not step through any of the nodes
over, A must be acyclic. In fact, moving from a station in one w1 , . . . , wk1 , there must exist k edge-disjoint paths in G
direction along a cycle it is not possible to rebuild the virtual connecting the pairs (s1 , t1 ), . . . , (sk , tk ).
channel of the station itself when it is reached twice. Finally, Vice versa, if there are k edge-disjoint paths in G connect-
if the source coincides with a base station, the maximum hop ing the pairs (s1 , t1 ), . . . , (sk , tk ), a 2k 1, 1, 1-layout for
count of is the eccentricity of s in A, that is, the maximum G = (V , E ), s and A = (U, F ) can be constructed as fol-
distance in A between s and the other stations. lows. For each i, 1 i k, the VP between si and ti is given
We are now ready to prove our first hardness result. 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 .
Theorem 3.1. Given a network G = (V , E), a source s V , Since s = s1 and the eccentricity in A = (U, F ) of the sta-
a chain adjacency graph A = (U, F ) and a positive inte- tion s1 is 2k 1, the layout thus constructed gives directly
ger h, deciding the existence of a h, 1, 1-layout for G with a h, 1, 1-layout with h = 2k 1.
source s is an NP-complete problem.
Notice that in the above construction the source s corre-
Proof. First of all, observe that, for any h, l, d, the prob- sponds to an endpoint of the chain A = (U, F ), so as already
lem of deciding the existence of a h, l, d-layout is in NP, as remarked the NP-completeness holds also under this restric-
given G = (V , E), s V , A = (U, F ) and a layout , it is tion.
possible to check in polynomial time whether Hmax () h, The above result can generalized to any fixed d > 0 as
Lmax () l and Dmax () d. follows.
We prove the claim by providing a polynomial time re-
duction from Disjoint Paths problem (DP), known to be Theorem 3.2. For any fixed integer d > 0, given a network
NP-complete [20]. An instance of this problem is consti- G = (V , E), a source s V , a chain adjacency graph
tuted by a graph G = (V , E) and a collection of node pairs A = (U, F ) and a positive integer h, deciding the existence
{(s1 , t1 ), . . . , (sk , tk )}. We want to determine whether there of a h, 1, d-layout for G with source s is an NP-complete
exist k edge-disjoint paths in G, each connecting a different problem.
pair (s1 , t1 ), 1 i k.
Without loss of generality, it is possible to assume that Proof. Given G, a source s and an adjacency graph A =
all the pairs (si , ti ), 1 i k, are disjoint, i.e., all (U, F ), it is sufficient to construct in polynomial time G , s
nodes s1 , . . . , sk , t1 , . . . , tk are different. In fact, any instance and A = (U , F ), such that, G, s, A admit a h, 1, 1-layout
not satisfying this property can be trivially modified into an if and only if G , s , A admit a h , 1, d-layout for a suitable
equivalent one in which every node v occurring in k k h > 0.
pairs is connected in G to k new nodes v1 , . . . , vk and the By theorem 3.1, it is possible to assume that A = (U, F )
k pairs contain in the order v1 , . . . , vk instead of v. is a chain of the nodes u1 , . . . , uk and s = u1 .
38 FLAMMINI ET AL.
G = (V , E ) is obtained from G = (V , E) by adding {s, t} F . The paths returned by ADP plus an eventual path
for each ui , 1 i < k, d 1 other stations ui,1 , . . . , ui,d1 connecting s to the node with minimum eccentricity in A if
connected by edges {ui , ui,1 } and {ui,j , ui,j +1 }, 1 j < s is not a base station form the VPs of an O(r)-approximate
d 1, in such a way that ui , ui,1 , . . . , ui,d1 form a chain of solution for the layout problem.
d nodes. s = u1 and the new adjacency graph A = (U , F ) To the best of our knowledge, the best general
algorithm
is, such that, U = U {ui,1 , . . . , ui,d1 | 1 i k} and for DP has an approximation ratio r = O( |E| log |V |),
F = {{ui , ui,1 | 1 i d 1} {{ui,j , ui,j +1 | 1 i < k, while r = O(polylog|V |) [28]. Therefore, an O( |E| log |V |)-
1 j < d 1} {{ui,d1 , ui+1 | 1 i < k}. Hence, approximation algorithm exists for the maximum load min-
A = (U, F ) is a chain. imization in layout with d = 1, while any algorithm with
Since l = 1, in any layout for G , s , A , all edges edges an asymptotic better approximation ratio would improve
{ui , ui,1 } and {ui,j , ui,j +1 }, 1 i < k and 1 j < d 1, upon [28].
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 4. Optimal chain layouts for d 2
{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- Starting from the hardness results shown in the previous sec-
mon prefix of the virtual channels VC(ui,d1 ) and VC(ui+1 )) tion, we now focus on specific topologies and provide optimal
to ui+1 . layouts for chain networks when the maximum channel dis-
Since s = u1 and u1 has eccentricity h = max{h, k 2 + tance d is at most 2. More precisely, we consider the case
d 2} in A = (U , F ), then G, s, A admit a h, 1, 1-layout in which the physical graph is a chain Cn of n nodes, that is
if and only if G , s , A admit a h , 1, d-layout, hence the V = {1, 2, . . . , n}, E = {{v, v + 1} | 1 v n 1} and
theorem holds. the adjacency graph A coincides with Cn . Moreover, without
loss of generality, we take the leftmost node of the chain as
Again, the NP-completeness still holds if the source s is an the source, i.e., s = 1, as otherwise we can split the layout
endpoint of the chain adjacency graph. construction problem into two equivalent independent sub-
Before concluding the section, let us, finally, show that for problems for the left- and the right-hand sides of the source,
d = 1 a stronger hardness result holds. To this aim observe respectively.
first that as remarked at the beginning of this section, the vir- Given fixed h,l,d and a h, l, d-layout for a chain Cn ,
tual topology induced by any h, l, 1-layout coincides with we say that is optimal if no h, l, d-layout exists for any
the adjacency graph. Moreover, if the source coincides with a chain Cm with m > n.
base station, h is equal to the eccentricity in A of s, otherwise By the considerations of the previous section for d = 1, the
connecting s by a VP to a node of minimum eccentricity in A virtual topology induced by the VPs of any h, l, 1-layout
it is possible to obtain a layout with a maximum hop count coincides with the adjacency graph A and thus with Cn . As a
equal to its eccentricity increasing the load at most of one. consequence, the largest chain admitting a h, l, 1-layout is,
Therefore, as far as approximation results are concerned, the such that, n = h + 1. Therefore, in the remaining part of this
interesting parameter to be approximated remains the maxi- section we focus on the case d = 2.
mum load. In the following we denote by u, v the unique VP cor-
The problem of minimizing the maximum load is equiv- responding to the simple path from u to v in Cn and by
alent from an approximation point of view to the optimiza- s, v1 v1 , v2 . . . vk , v or simply s, v1 , v2 , . . . , vk , v the
tion version of the decision problem DP in which we want to virtual channel VC(v) of v given by the concatenation of the
determine k paths connecting the k sourcedestination pairs VPs s, v1 , v1 , v2 , . . . , vk , v.
(s1 , t1 ), . . . , (sk , tk ) in such a way as to minimize the maxi- The following lemma establishes that, when moving in one
mum number of paths sharing a same edge. direction along a chain, some VPs are accumulated, that is
In fact, any r-approximation algorithm A for the layout they cannot be removed from the VCs of the successive nodes
problem yields directly a O(r)-approximation algorithm ADP encountered along the same direction.
for DP. Informally, ADP simply consists in running A on the
instance of the layout problem obtained by adding a new Lemma 4.1. Given a h, l, 2-layout for a chain network
source s, connecting s to each si , 1 i k, and letting and a node v, if VC(v) = s, v1 , v2 , . . . , vk , v and in VC(v)
A = (U, F ) be, such that, U = {s, s1 , . . . , sk , t1 , . . . , tk } and there exist two consecutive VPs vi1 , vi , vi , vi+1 with
F = {{s, si } | 1 i k} {{si , ti } | 1 i k}. The vi1 < v and vi < v (resp. vi1 > v and vi > v), then
k VPs connecting each si to ti , 1 i k, in the layout for every u v (resp. u v), s, v1 , v2 , . . . , vi1 , vi is a
returned by A correspond to an O(r)-approximate solution prefix of VC(u).
for DP. A reverse reduction can be determined by observing
that an r-approximation algorithm ADP for DP yields directly Proof. Assume first that vi1 < v and vi < v and let
an O(r)-approximation algorithm A for the layout construc- u > v be the first node, such that, vi , vi+1
/ VC(u). Since
tion problem that consists in running ADP on the instance ob- d(u 1, u) 2, to reach u we can only add vi , u. This
tained by associating a sourcedestination (s, t) to each edge process can be iterated to every node w with w > u, hence the
WIRELESS ATM LAYOUTS FOR CHAIN NETWORKS 39
claim holds. An analogous argument applies when vi1 > v Assume then that the VP vk , v 1 is not contained in
and vi > v. VC(u) for every u < v 1. In this case also vk , v is not
contained in any VC(u) with u < v 1, because otherwise,
Another useful property of h, l, 2-layouts is that the pre- with no matter of the sense in which vk , v is traversed in
fixes of a VC are the VCs of their final nodes. VC(u), VC(v) could not contain vk , v 1, as by hypotheses
vk , v 1 does not belong to VC(u).
Lemma 4.2. There exists an optimal h, l, 2-layout for If vk , v is contained in VC(v + 1), then by lemma 4.1
a chain network, such that, for every node v with VC(v) = VC(v) is a prefix of all the VCs VC(u) with u v. Therefore,
s, v1 , v2 , . . . , vk1 , vk , v, VC(vi ) = s, v1 , v2 , . . . , vi for the layout obtained by deleting the VPs vk , v1 and vk , v,
every i k. 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 , . . . ,
Proof. Let be any optimal h, l, 2-layout. We now prove v 1, v, . . . , u does not increase the hop count of any node,
that, if for a given node v each prefix of every virtual channel the load of any edge and the channel distance of the ad-
VC(u) with u < v is the virtual channel of the corresponding jacent nodes. Therefore, since in the new layout the vir-
final node, then can be modified in such a way that such tual channels of the nodes before v are not modified and
a property is satisfied also by VC(v). This clearly proves the VC(v) = s, v1 , v2 , . . . , v 1, v = s, v1 , v2 , . . . , vk1 , v,
lemma. s, v1 , v2 , . . . , vi is a prefix of VC(v 1) for every vi with
Trivially the property is satisfied by the virtual channel of 1 i < k and thus VC(vi ) = s, v1 , v2 , . . . , vi .
source s, since it is empty. Therefore, let v 2 be any node, If vk , v is not contained in VC(v + 1), then by lemma 4.1
such that, the property is true for all the VCs of the previous s, v1 , v2 , . . . , v 1, vk is a prefix of all the VCs VC(u) with
nodes and let VC(v) = s, v1 , v2 , . . . , vk1 , vk , v. Recalling u v. Moreover, all the VPs starting at vk contained in
that d(v 1, v) 2, it is possible to distinguish among the the VC of some node u > v are not contained in any VC
following cases: VC(w) with w v 1, as otherwise by hypothesis VC(u)
would not contain vk , v 1. Notice also that the other
1. VC(v 1) = s, v1 , v2 , . . . , vk , v 1, endpoint of each such VP is greater than v, as otherwise by
the maximum channel distance it cannot be used in the VCs
2. VC(v 1) = s, v1 , v2 , . . . , vk , v, v 1, of the nodes after v. It is thus possible to modify the lay-
3. VC(v 1) = s, v1 , v2 , . . . , vk , that is vk = v 1, 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
4. VC(v 1) = s, v1 , v2 , . . . , vk , vk+1 , v, and some node u > v are substituted with v, w and, finally, each
5. VC(v 1) = s, v1 , v2 , . . . , vk1 , that is vk1 = v 1. 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
Since by the hypothesis the claim is true for VC(v 1) the above considerations, the new layout does not increase the
and s, v1 , v2 , . . . , vk is a prefix of VC(v 1), VC(vi ) = hop count of any node, the load of any edge and the channel
s, v1 , v2 , . . . , vi for each vi with 1 i k. Therefore, distance of the adjacent nodes. Moreover, it does not modify
every prefix of VC(v) is a VC. the virtual channels of the nodes before v. Therefore, again
If VC(v 1) = s, v1 , v2 , . . . , vk1 , that is vk1 = v 1, VC(v) = s, v1 , v2 , . . . , v 1, v = s, v1 , v2 , . . . , vk1 , v,
we further distinguish the following two subcases. s, v1 , v2 , . . . , vi is a prefix of VC(v 1) for every vi with
(I) vk > v. In this case the VP v, vk must be 1 i < k and thus VC(vi ) = s, v1 , v2 , . . . , vi .
deleted in the VC of a node u with v < u vk . If
u = vk VC(vk ) = s, v1 , v2 , . . . , vk , otherwise, since Motivated by the previous lemma, even if not explicitly
d(u 1, u) 2, VC(u) = s, v1 , v2 , . . . , vk , u and iterating stated, in the remaining part of this section, we restrict our
the same argument to the VP u, vk we finally have that again attention to layouts in which all the prefixes of each VC are
VC(vk ) = s, v1 , v2 , . . . , vk . Therefore, since for each vi the VCs of the corresponding final nodes. In fact, this does
with i < k s, v1 , v2 , . . . , vi is a prefix of VC(v1), we have not affect the correctness of our results, since the optimality
VC(vi ) = s, v1 , v2 , . . . , vi for every vi with 1 i k. is preserved under such assumption.
(II) vk < v 1. If the VP vk , v 1 is contained in the The following corollary is a direct consequence of the pre-
VC of a node u < v 1, then starting from the source s, in vious lemma.
VC(u) vk , v 1 is not traversed from vk to v 1, otherwise
Corollary 4.3. Every VP of a h, l, 2-layout for a chain
by hypothesis vk , v 1 would be contained in VC(v 1)
network is the final VP of exactly one of its two endpoints.
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 A last useful property that allows to suitably bound the
vk and again by hypothesis the prefix of VC(u) till v 1 coin- maximum size of a chain admitting a h, l, 2-layout is estab-
cides with VC(v 1) and VC(vk ) = s, v1 , v2 , . . . , vk1 , vk . lished in the following lemma.
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 ) = Lemma 4.4. Given a h, l, 2-layout for a chain network
s, v1 , v2 , . . . , vi for every vi with 1 i k. and any j h, let u and v be the last nodes with hop count
40 FLAMMINI ET AL.
h(u) = j 1 and h(v) = j , respectively. Then the last VPs analogous considerations show that it would contradict
of all the VCs reaching the nodes from u+1 to v share a same h(s) = 0.
physical link. 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
Proof. Let us first prove the claim for j = 1, and thus with y, z can be the last one of VC(y).
u corresponding to the source s = 1.
In order to show that the lemma holds for j = 1, it is In conclusion, neither y = w can hold and therefore no
sufficient to prove that no two edge-disjoint VPs exist in the two edge-disjoint VPs exist before v.
VCs of the nodes from u + 1 to v. In order to extend the proof to every j h, observe that
Assume by contradiction that such property does not hold since u is the last node with hop count h(u) = j 1 and
and let x, y,w, z be a pair of closest edge-disjoint VPs in d = 2, all the VCs of the nodes w u have VC(u) as prefix.
such VCs with x < y w < z v. Then y = w or Thus an identical proof shows that all the VPs after u in the
y = w 1, otherwise any other VP used to reach a node VCs of the nodes w with u < w v share a same physical
between y and w would be disjoint from x, y or w, z and edge.
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 It is thus possible to prove the following lemma.
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- Lemma 4.5. Given any h, l, 2-layout for a chain net-
sis h(v) = 1. Thus, if x, y is used to reach y, since work and j h, the last node v with h(v) = j is, such
d(y, w) 2, when moving from y to w x, y must be re- that, v j l + 1.
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, Proof. The claim trivially holds for j = 0, since only the
again contradicting the hypothesis. source s = 1 has hop count h(s) = 0.
By corollary 4.3, x, y is then used to reach x. If y is Assume by induction that the corollary holds for a given
reached by a VP q, y with q < y, then by corollary 4.3 j , such that, 1 j < h and let u j l + 1 the last node
w, z is used to reach z, and thus by lemma 4.1 q, y is con- with hop count h(u) = j and v the last node with hop count
tained in VC(v) again contradicting the hypothesis h(v) = 1. h(v) = j + 1. Since by lemma 4.4 the last VPs of all the VCs
If q > y x, y and y, q would be closer than x, y and reaching the nodes from u + 1 to v share a same physical link,
w, z: a contradiction. there cannot be more than l nodes from u + 1 to v, otherwise
Therefore, y = w 1 cannot hold and it must be y = w. the shared link would have load greater than l. Therefore,
Recalling that by corollary 4.3 every VP is used to reach v u + l (j + 1)l + 1.
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 In conclusion, the following theorem holds.
the following cases:
x, y reaches x and y, z reaches z. Theorem 4.6. For every h, l 0, a h, l, 2-layout for a
By corollary 4.3 y is reached by another VP q, y with chain network Cn is optimal if and only if n = hl + 1.
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, Proof. By lemma 4.5, n hl + 1 holds for any h, l, 2-
thus contradicting h(s) = 0. Similarly, if q < y, q, y is layout for a chain Cn . A layout attaining n = hl + 1 is de-
contained in all the VCs of the nodes after z, z included, picted in figure 1.
thus contradicting h(v) = 1. Therefore, this case cannot
hold. Before concluding the section, we finally observe that the
layout in figure 1 is not the only optimal one. Another exam-
x, y reaches y and y, z reaches z.
ple with h = 4, l = 4 and d = 2 is shown in figure 2.
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 5. Optimal canonic layouts for chain networks
h(v) = 1.
x, y reaches x and y, z reaches y. In this section we provide h, l, d-layouts for chain networks
This case is symmetric to the previous one and completely that are optimal within the class of the canonic layouts. In
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
T r (h j, l 1, d) with 1 j min{h 1, (d 1)/2
}
and the virtual topology induced by its VPs is a tree.
is attached to v in such a way that the highest order node
Let us say that a rooted tree is ordered if a total order of T r (h 1, l, d) coincides with v and the highest order
is defined on its nodes with the root being the lowest 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- (d 1)/2
} coincides with the lowest order node of
outs for chains and ordered trees. Namely, each node of the T r (h j + 1, l 1, d).
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- An example of T (h, l, d) is depicted in figure 3. Infor-
tal order of the nodes of the tree is given by the order of mally speaking, a T (h, l, d) corresponds to the sublayout of
the nodes along the chain. Clearly, not all the ordered trees a canonic layout induced by all the VPs occurring under
yield canonic layouts, as their induced VPs might be cross- a given VP, with the lowest order node being closer to the
ing. However, the one-to-one correspondence between or- source. Thus, T (h, l, d) is the subtree induced by all the VPs
dered trees and canonic layouts is maintained if we restrict whose endpoints occur from the first endpoint of the given VP
to ordered trees in which every subtree contains a subset of until the second endpoint.
nodes that forms an interval according to the node ordering. Directly from the definition, it follows that all the nodes in
In other words, each subtree corresponds to a segment of the T (h, l, d) are at distance at most h from u, and thus at most
chain not touched by the other subtrees. h additional hops from the node corresponding to u in are
Given any ordered tree T , let the reverse tree T r be the sufficient to reach the other nodes corresponding to T (h, l, d)
symmetric ordered tree obtained from T by inverting the or- in the chain. Moreover, the load yielded by T (h, l, d) on its
der of the nodes (hence the root becomes the highest or- segment of the chain is bounded by l. Finally, the channel
der node). We now introduce a new class of ordered trees distance between two consecutive nodes belonging to the sub-
(h, l, d) that allows to completely define the structure of an chain of T (h, l, d) is always at most equal to d. In fact, it
optimal h, l, d-layout. is given by the maximum distance in T (h, l, d) between two
The definition of (h, l, d) is recursive and the solution nodes adjacent in the ordering. Therefore, assuming by induc-
of the associated recurrence gives the exact number of nodes tion that such property holds inside the subtrees T (j, l 1, d)
reached by an optimal canonic h, l, d-layout. Before intro- (and thus T r (j, l 1, d)), in order to show that it holds also
ducing (h, l, d), let us define another ordered subtree that is in T (h, l, d) it is sufficient to prove that the final node of the
exploited in its definition. 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
Definition 5.3. Given any h, l, d, T (h, l, d) is an ordered tree
adjacent pair of nodes, is at distance at most d. But such nodes
recursively defined as follows.
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 FLAMMINI ET AL.
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. Assume then that the claim is true for l 1, that is, for all
Clearly, symmetric considerations hold for each T r (h, l, d). the VPs of a canonic h, l, d-layout with l l + 1 VPs
We are now ready to define the final tree (h, l, d). 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
Definition 5.4. The ordered tree (h, l, d) is formed by the last node with u w v (that is under u, v) reached
chain of h trees T (j, l, d), 1 j h, such that, the lowest by a VC stepping through u and not from v, and consider
order node of T (j, l, d) coincides the highest order node of the subchain of d1 VPs w1 , . . . , wd1 +1 with w1 = u and
T (j + 1, l, d) for 1 j < h (see figure 4). 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
Notice that, if h = 0 or l = 0, (h, l, d) consists of just a connecting v to w + 1. Since w and w + 1 are adjacent and
single node. Moreover, an alternative recursive definition of the maximum channel distance is d, it must be d1 +d2 +1 d.
(h, l, d) is given by a T (h, l, d) attached to a (h 1, l, d) Moreover, since h(w) h and h(w + 1) h, d1 h and
tree (again see figure 4). d2 h 1. Therefore, since such subchains and with all their
Let Tn (h, l, d) denote the number of nodes of T (h, l, d) VPs occur under u, v, by applying the inductive assumption
(and thus of T r (h, l, d)) minus one. Then, directly from it follows that
definition 5.3, Tn (h, l, d) = 0 if h = 0 or l = 0, other- 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
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 0kd/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 0k(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
0kd/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 0kd/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 0kd/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 0k(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
0k(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 0k(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
0k(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 0kd/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 45
[4] J. Burgin and D. Dorman, Broadband ISDN resource management: The [23] G. Parry, Wireless ATM MAC protocols a literature survey, WARP
role of virtual paths, IEEE Communicatons Magazine 29 (1991). Project URL, http://vera.ee.und.ac.za/coe/warp
[5] M. Cheng, S. Rajagopalan, L. Chang, G. Pollini and M. Barton. PCS (1999).
mobility support over fixed ATM networks, IEEE Communications [24] C. Partridge, Gigabit Networking (Addison-Wesley, Reading, MA,
Magazine 35 (1997) 8292. 1994).
[6] C. Chrysostomou, A. Pitsillides and F. Pavlidou, A survey of wireless [25] K.I. Sato, S. Ohta and I. Tokizawa, Broad-band ATM network archi-
ATM handover iusses, in: Proc. of the Internat. Symposium of 3G In- tecture based on virtual paths, IEEE Transactions on Communications
frastructure and Services, 3GIS, Vol. 2/3 (2001) pp. 3439. 38(8) (1990) 12121222.
[7] I. Cidon, O. Gerstel and S. Zaks, A scalable approach to routing in ATM [26] Y. Sato and K.I. Sato, Virtual path and link capacity design for ATM
networks, in: Proc. of the 8th Internat. Workshop on Distributed Algo- networks, IEEE Journal on Selected Areas in Communications 9
rithms, eds. G. Tel and P.M.B. Vitnyi, Terschelling, The Netherlands (1991).
(October 1994), Lecture Notes in Computer Sience, Vol. 857 (Springer, [27] D. Sobirk and J.M. Karlsson, A survey of wireless ATM MAC
New York, 1994) pp. 209222; submitted for publication in IEEE/ACM protocols, in: Proc. of the Internat. Conf. on the Performance
Transactions on Networking. and Management of Complex Communication Networks (PMCCN)
[8] R. Cohen and A. Segall, Connection management and rerouting in (Chapman & Hall, London, 1997).
ATM networks, in: Proc. of the 13th INFOCOM Conf. (1994) pp. 184 [28] A. Srinivasan, Improved approximations for edge-disjoint paths,
191. unsplittable flow, and related routing problems, in: Proc. of the 38th
[9] Y. Dinitz, M. Feighelstein and S. Zaks, On optimal graphs embedded Annual IEEE Symposium on Foundations of Computer Science (FOCS)
into path and rings, with analysis using l1 -spheres, in: Proc. of the 23rd (IEEE Computer Society Press, Los Alamitos, CA, 1997) pp. 416425.
Internat. Workshop on Graph-Theoretic Concepts in Computer Science [29] L. Stacho and I. Vrto, Virtual path layouts for some bounded degree
(WG), Lecture Notes in Computer Science, Vol. 1335 (Springer, New networks, in: Proc. of the 3rd Colloquium on Structural Information
York, 1997) pp. 171183. and Communication Complexity (SIROCCO) (Carleton Univ. Press,
[10] T. Eilam, M. Flammini and S. Zaks, A complete characterization 1996) pp. 269278.
of the path layout construction problem for ATM networks with [30] S. Zaks, Path layouts in ATM networks, in: Proc. of the SOFSEM
given hop count and load, in: Proc. of the 24th Internat. Collo- Conf., Lecture Notes in Computer Science, Vol. 1338 (Springer, New
quium on Automata, Languages and Programming (ICALP), Lecture York, 1997) pp. 144160.
Notes in Computer Science, Vol. 1256 (Springer, New York, 1997)
pp. 527537.
Michele Flammini received the degree in computer
[11] M. Flammini, G. Gambosi and A. Navarra, Wireless ATM layouts for
science at the University of LAquila in 1990 and the
chain networks, in: Proc. of the 17th Internat. Parallel and Distributed
Ph.D. degree in computer science at the University of
Processing Symposium (IPDPS), 3rd Workshop on Wireless, Mobile
Rome La Sapienza in 1995. He is associate profes-
and Ad Hoc Networks (WMAN), IEEE Computer Society (2003)
sor at the computer science Department of the Uni-
p. 220.1.
versity of LAquila since 2000. His research inter-
[12] M. Flammini, A. Gasparini, G. Gambosi and A. Navarra, Dy-
ests include algorithms and computational complex-
namic layouts for wireless ATM, in: Proc. of the 9th Internat.
ity, communication problems in interconnection net-
Conf. on Parallel and Distributed Computing (Euro-Par), Lecture
works and routing. He has authored and co-authored
Notes in Computer Science, Vol. 2790 (Springer, New York, 2003)
more than 50 papers in his fields of interest published
pp. 10561063.
in the most reputed international conferences and journals.
[13] O. Gerstel, I. Cidon and S. Zaks, The layout of virtual paths in
E-mail: flammini@di.univaq.it
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 Giorgio Gambosi received the degree in electronic
network, in: 3rd Annual European Symposium on Algorithms (ESA), engineering at the University of Rome La Sapienza
Corfu, Greece (September 1995), Lecture Notes in Computer Science, in 1980. He is full professor at the Department of
Vol. 979 (Springer, New York, 1995) pp. 508522; to appear in Mathematics of the University of Rome Tor Ver-
Discrete Applied Mathematics. gata. His research interests include distributed sys-
[15] O. Gerstel and S. Zaks, The virtual path layout problem in fast tems, network management algorithms and routing.
networks, in: Proc. of the 13th ACM Symposium on Principles of Dis- He has coauthored about 70 research papers in rele-
tributed Computing, Los Angeles, USA (August 1994) pp. 235243. vant international conferences and journals.
[16] J.D. Gibson, The Mobile Communications Handbook, 2nd ed. (CRC E-mail: gambosi@mat.utovrm.it
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 Alfredo Navarra received the degree in computer
Introduction to ATM-Based Networks (Addison-Wesley, Reading, MA, science at the University of LAquila in 2000. He
1991). spent one year at the research institute INRIA of
[19] ITU recommendation, I series, Blue Book (November 1990). Sophia Antipolis (France) collaborating with the
[20] R.M. Karp, On the computational complexity of combinatorial MASCOTTE project group and now he is ending his
problems, Networks 5 (1975) 4568. Ph.D. degree in computer science at the University of
[21] E. Kranakis, D. Krizanc and A. Pelc, Hop-congestion tradeoffs in Rome La Sapienza. His research interests include
ATM networks, in: Proc. of the 9th IEEE Symposium on Parallel and algorithms and computational complexity, ATM, op-
Distributed Processing (1995) pp. 662668. tical and wireless communication networks.
[22] M. Mouly and M.B. Pautet, The GSM System for Mobile Communica- E-mail: navarra@di.univaq.it;
tions (Cell & Sys, 1993). navarra@sophia.inria.fr
Mobile Networks and Applications 10, 4759, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
Table 1
A few variations of the multicast routing problem and how each node would
compute its cost in MANSI.
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 51
best cost tables in accordance with j and the height field. If by reducing their values by DECAYING_FACTOR at every
the ant is deterministic, the cost that it carries back is the ac- DECAY_INTERVAL time period:
tual cost of the path its originator is currently using to join the
g (i, j, h) = (1 DECAYING_FACTOR)g (i, j, h), (5)
group. Therefore, the best cost corresponding to the height
field is updated to this value. If the ant is not deterministic, where 0 < DECAYING_FACTOR < 1.
however, the best cost is updated only when it is higher than By probabilistically selecting next hops, the majority of
the returned cost, which means that the ant has found a better the F ORWARD A NTs will choose paths with high pheromone
path to join the group from this node. The pheromone inten- intensity, while some of them may explore totally different
sity on this link is also updated to the maximum in order to new paths. If a BACKWARD A NT comes back with a better
encourage subsequent F ORWARD A NTs to use the same link, cost on a new branch, the pheromone amount on that branch
as well as to redirect join request to this link instead. If the will be increased significantly. As a result, a change in multi-
ant comes back with a higher cost, a pheromone amount of cast connectivity (i.e., forwarding set) is triggered due to the
1/(1 + cost) is added instead. In case of deterministic ant, periodic broadcast of J OIN R EQUEST packets, as illustrated
the added amount is reduced by half since this link already in figure 4(d).
has the highest pheromone intensity as it has just been chosen MANSI also takes advantage of broadcast nature of wire-
by a deterministic F ORWARD A NT. Note that we have men- less communication to speed up the learning process as
tioned this procedure before when we explained how a node follows. When a node i overhears a J OIN R EQUEST for
uses it while processing a C ORE A NNOUNCE (line 8 of algo- group g from j but not intended to itself, it invokes Up-
rithm 1). This is because a C ORE A NNOUNCE more or less datePheromoneAndCostg (j, hj , 0, TRUE), where hj is the
serves as a deterministic BACKWARD A NT returning from the height that j reports in its J OIN R EQUEST. This implies
core. 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
Algorithm 5. Node i processing a BACKWARD A NT packet. some members who are not forwarding nodes will broadcast
J OIN R EQUESTs as well and might be mistaken as forwarding
1: Input: nodes by its neighbors.
2: bant incoming BACKWARD A NT
3.4. Multicast data forwarding
3: lastHop the node from which bant
was received Since MANSI is a mesh-based protocol which allows for-
warding nodes and members to accept data packets arriving
4: Begin: from any node, each data packet is assigned a unique se-
quence number when it is transmitted from the source. The
5: g fant.group
sequence numbers are checked by each forwarding node and
6: Invoke UpdatePheromoneAnd member node to make sure that no duplicate data packets are
rebroadcast or delivered to the application. When a node i
Costg (lastHop, bant.height, bant.cost, bant.d) receives a non-duplicate data packet of group g, it checks
7: if i = last entry in fant.visitedNodes then whether it is currently a forwarding node of the group, i.e.,
joing (i) = . If so, it rebroadcasts the packet. Otherwise, the
8: Remove the last entry from bant.visitedNodes packet is silently discarded.
9: if bant.visitedNodes = then
3.5. Handling mobility
10: bant.cost bant.cost + cost(i)
In MANSI, mobility and other network dynamics are handled
11: broadcast bant
inherently rather than as exceptions. With the pheromone lay-
12: end if ing/following behavior of BACKWARD A NTs and F ORWARD
A NTs, each path comprising the forwarding set keeps being
13: end if reinforced as long as no link on the path is broken. How-
ever, network dynamics can cause optimal connectivity to
After updating the pheromone and the best cost tables, change from time to time even though the current connectiv-
k checks if the BACKWARD A NT was intended to itself by ity may still be valid. With the probabilistic nature of F OR -
examining the last entry in the visitedNodes field. If its ID WARD A NT s to explore new paths, the multicast forwarding
matches, it adds its cost into the cost field, removes the last set should be able to evolve into a configuration that is more
entry from visitedNodes, and rebroadcasts as long as there is efficient for the new topology.
at least one entry left in visitedNodes. Algorithm 5 presents When a link currently used by a member or a forwarding
the pseudo code of how a node processes a BACKWARD A NT. node to send J OIN R EQUESTs breaks, the pheromone table
Similar to pheromone evaporation of biological ants, each entries corresponding to that link are also removed. There-
node i updates all the entries g (i, j, h) in its pheromone table fore, all subsequent F ORWARD A NTs will be redirect to other
AD HOC MULTICAST ROUTING ALGORITHM 55
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 toward its best next hop as usual. If nlff exceeds the threshold,
pheromone intensity was the second highest before the link however, it will add another entry for the second best next hop
failure. If this next hop leads to a forwarding node of a higher into its J OIN R EQUESTs. Since all the neighbors are ranked
height, BACKWARD A NTs will return and update pheromone by their goodness in terms of pheromone intensities, the sec-
on the new path, hence reestablishing a connection to the ond best next hop can be easily determined. Formally, if k is
group. However, in case that F ORWARD A NTs fail to find a the best next hop for i to join the group g, as defined in (2),
new path, C ORE A NNOUNCEs flooded periodically will even- then the second best next hop k
is defined as:
tually restore the connectivity.
g (i, n, h)
Although MANSI is considered a mesh-based protocol by k
= arg max . (8)
its way of forwarding data packets, connectivity of the for- nntab(i),n=k h>height (i) bestCostg (i, h) + 1
g
warding set may still be fragile if the network is sparse and
members are far apart from each other, especially with the Figure 8(a) illustrates the forwarding set created by MANSI
presence of mobility. To make data forwarding more effective without the mobility-adaptive mechanism for a multicast
under mobility, while maintaining good efficiency when the group of three members in a network of 50 nodes, where
network is static, we incorporate a mobility-adaptive mecha- each node is moving at 10 m/s. The group connectivity is
nism into MANSI. With this mechanism, each node i keeps almost a straight line and is vulnerable to link failures. With
track of the normalized link failure frequency, denoted by the mobility-adaptive mechanism enabled, most members and
nlff (i), which reflects the dynamic condition of the area sur- forwarding nodes request two of their neighbors to be in the
rounding i in terms of the number of link failures per neigh- forwarding set, as shown in figure 8(b), so that the group con-
bor per second. A calculation of nlff (i) is performed every nectivity becomes more robust.
NLFF_TIME_WINDOW time period as follows:
current_nlff (i)
4. Experimental results and discussion
f
= , (6)
NLFF_TIME_WINDOW |ntab(i)| To study the characteristics and evaluate the performance of
current_nlff (i) + nlff (i) MANSI, we have conducted simulation experiments using the
nlff (i) = , (7)
2 QualNet simulator [10]. Ten random networks were gener-
where f is the number of link failures detected during the last ated with 50 nodes uniformly distributed over a terrain of
NLFF_TIME_WINDOW time period. Initially nlff (i) is set size 1000 1000 m2 . Each node was equipped with a ra-
to zero. dio transceiver which was capable of transmitting signals up
Each member or forwarding node then uses this nlff to de- to approximately 250 meters over a 2 Mbps wireless chan-
termine the stability of its surrounding area. If its nlff is lower nel, using the two-ray path loss model without fading. We
than a threshold NLFF_THRESHOLD, the node will consider used IEEE 802.11DCF as the MAC layer protocol, and IP
its area stable and join the group by sending J OIN R EQUESTs 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
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
Figure 9. Average size of the forwarding set as a function of time for CORE group is high.
and MANSI. Table 3 summarizes the sizes, averaged over the entire sim-
ulation time, of the forwarding sets maintained by MANSI,
unicast routing protocol, no other routing protocols were em- CORE, and FLOOD on each simulated network. (FLOOD
ployed. For each network, a multicast groups of 5 members does not really maintain a forwarding set, but the set con-
was setup, where each member generated a constant bit rate sists of every node in the network.) The results show that in
(CBR) traffic at 2 packets/sec to the group for 20 minutes. all cases, except one, MANSI yields forwarding sets that are
The size of data payload was 512 bytes. The MANSI parame- approximately 15%20% smaller than those of CORE, and
ter values used in our simulation are shown in table 2. Note much smaller than FLOOD. Since the size of the forward-
that NLFF_THRESHOLD is used only when the mobility- ing set indicates how many nodes are involved to relay a data
adaptive mechanism is enabled. packet from one member to the others, this demonstrates the
Our first set of experiments were setup without mobility in efficiency of MANSI in terms of data forwarding.
order to study how MANSI maintains forwarding sets in static We have performed another set of experiments to compare
environments. For comparison purposes, we used two base- the performance of MANSI, in terms of effectiveness and ef-
line protocols: FLOOD and CORE, as references. FLOOD is ficiency, with ODMRP. ODMRP [1] is an on-demand, mesh-
a simple flooding protocol where a data packet is rebroadcast based multicast protocol that attempts to establish a forward-
by every node in the network. And CORE is a generic core- ing group similar to a forward set in MANSI only when
based protocol that operates exactly like MANSI, but with a source of the group has data to send. The nodes in the for-
no ants deployed, where C ORE A NNOUNCEs are periodically warding group form a mesh that connects the group members
flooded as usual. The cost of each node was set to one, which together. When a multicast source has data to send for the
implies that MANSI would attempt to reduce the size of the first time, it broadcasts to its neighbors a J OIN Q UERY packet,
forwarding set. which is a data packet with the query flag set. Upon receiving
We first look at the average size of forwarding sets main- a non-duplicate J OIN Q UERY, each node stores the upstream
tained by CORE and MANSI over time for the ten sample node ID in its routing table and rebroadcasts the packet. When
networks, as shown in figure 9. Due to random delays added a member of the multicast group receives a J OIN Q UERY, it
to avoid packet collisions when broadcasting, the dissemina- constructs and broadcasts a J OIN R EPLY packet containing
tion pattern of a C ORE A NNOUNCE is unpredictable when it the source ID and the upstream node ID to all of its neighbors.
is flooded, which causes a forwarding set to be formed dif- Upon receiving a J OIN R EPLY, a node whose ID matches the
ferently for each announcement. Consequently, the average upstream ID in the packet realizes that it is on the path be-
size of forwarding sets keeps changing from time to time in tween the source and a member, so it becomes a forward-
CORE. In contrast, forwarding sets maintained by MANSI ing node for the group by setting its FG_FLAG (Forwarding
start of at around the same size as that of CORE but keep re- Group Flag). It then constructs and broadcasts its own J OIN
AD HOC MULTICAST ROUTING ALGORITHM 57
hoc multicast routing protocol has been proposed to exploit allel and Distributed Computing and Systems (PDCS98), Las Vegas,
the concept of swarm intelligence. 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
6. Conclusion and future work
A.A. Gray, Adaptive-SDR: Adaptive swarm-based distributed routing,
in: IEEE WCCI 2002, IJCNN 2002 Special Session: Intelligent Sig-
Inspired by swarm intelligence, we have introduced an alter- nal Processing for Wireless Communications, Honolulu, Hawaii (12
native approach to solving the multicast routing problem in 17 May 2002).
mobile ad hoc networks. Our protocol, called MANSI (Mul- [7] S.-J. Lee, W. Su and M. Gerla, On-demand multicast routing
ticast for Ad hoc Networks with Swarm Intelligence), is an protocol (ODMRP) for ad hoc networks, IETF Internet Draft,
on-demand multicast routing protocol that creates a multicast http://www.ietf.org/proceedings/00jul/I-D/manet
mesh shared by all the members within each group. The pro- -odmrp-02.txt (2000).
[8] M. Liu, R.R. Talpade and A. McAuley, AMRoute: Adhoc multicast
tocol uses a core-based scheme, where each member initiates routing protocol, Technical Report 99, The Institute for Systems Re-
a request to the core node to establish multicast connectivity search, University of Maryland (1999).
with other members. Intermediate nodes who receive such a [9] R. Royer and C. Perkins, Multicast using ad-hoc on demand distance
request become forwarding nodes that are used to relay data vector routing, in: MOBICOM99, Seattle, WA (August 1999) pp. 207
packets from one member to the others. Unlike other core- 218.
[10] Scalable Network Technologies, QualNet Simulator, http://www.
based protocols, MANSI does not always rely on the shortest scalable-networks.com.
paths between the core and the members to establish group [11] R. Schoonderwoerd, O. Holland, J. Bruten and L. Rothkrantz, Ant-
connectivity. Instead, each member who is not the core peri- based load balancing in telecommunications networks, Technical Re-
odically deploys a small packet that behaves like an ant to op- port HPL-96-76, Hewlett-Packart Laboraties Bristol, Bristol, UK
portunistically explore different paths. This exploring mech- (21 May 1996).
[12] C.W. Wu and Y.C. Tay, AMRIS: A multicast protocol for ad hoc wire-
anism enables the protocol to discover paths that comprise a less networks, in: IEEE Military Communications Conf. (MILCOM),
better set of forwarding nodes yielding a lower total cost of Atlantic City, NJ (November 1999) pp. 2529.
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
Chien-Chung Shen received his B.S. and M.S. de-
mechanism that allows the protocol to remain effective as
grees from National Chiao Tung University, Taiwan,
mobility increases. The simulation results have shown that and his Ph.D. degree from UCLA, all in computer
MANSI performs both effectively and efficiently in static or science. He was a research scientist at Bellcore Ap-
low-mobility environments, yet still effectively in highly dy- plied Research working on control and management
namic environments. of broadband networks. He is now an assistant pro-
Research is in progress to apply MANSI with other objec- fessor in the Department of Computer and Informa-
tion Sciences of the University of Delaware, and a
tives such as load balancing, energy conservation, and secu- recipient of NSF CAREER Award. His research in-
rity. terests include ad hoc and sensor networks, control
and management of broadband networks, distributed object and peer-to-peer
computing, and simulation.
References E-mail: cshen@cis.udel.edu
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
most no node receives the message or almost all of them do. sume that all wireless nodes have the same maximum trans-
So ideally, we would set the gossiping probability to some mission range and we normalize it to one unit. Throughout
value slightly larger than p0 to reduce the routing messages this paper, a broadcast by a node u means that u sends the
propagated. When the network is sufficiently large, we can set message to all nodes within its transmission range. Notice
p sufficiently close to p0 , thus save about (1 p0 )n messages that, in wireless ad hoc networks, the radio signal sent out
overhead compared with the flooding, since about p0 n nodes by a node u can be received by all nodes within the trans-
will forward the message in gossiping based method com- mission range of u. The main communication cost in wire-
pared with n nodes forwarding in flooding. Hass et al. [24] less networks is to send out the signal while the receiving and
conducted extensive simulations to investigate the extent to processing costs of a message is neglected here.
which this gossiping probability can be lowered. They found
that using gossiping probability between 0.6 and 0.8 suffices 2.1. Location service
to ensure that almost every node gets the message in almost
every routing. They report of up to 35% fewer messages than Several proposed routing algorithms [18,22] assume that the
flooding (close to our previous explanation). Notice that their source node knows the position information (or approximate
experimental setting of the network has some special config- position) of the destination node. Our regional gossip method
urations [24]. also assumes that the source node knows the current posi-
Although gossiping reduces the routing messages com- tion information of the target approximately. Notice that, for
pared with flooding, it still produces lots of unnecessary mes- sensor networks collecting data, the destination node is often
sages in regions that are far from the line between sender fixed, thus, location service is not needed in those applica-
node and receiver node. Notice that, the traditional gossip tions. However, the help of a location service is needed in
will propagate the message to the whole network. To further most application scenarios. Mobile nodes register their loca-
reduce the number of forwarding messages, we propose re- tions to the location service. When a source node does not
gional gossiping, in which essentially only nodes inside some know the position of the destination node, it queries the loca-
region (derived from the source and target) will execute the tion service to get that information. In cellular networks, there
gossiping protocol, and nodes outside the region will not par- are dedicated position severs. It will be difficult to implement
ticipate in the gossiping at all. The region we select in our the centralized approach of location services in wireless ad-
simulations are some ellipses using the source and target as hoc networks. First, for centralized approach, each node has
foci. Notice that here we assume source node knows either to know the position of the node that provides the location
the exact or the approximate location of the destination node, services, which is a chicken-and-egg problem. Second, the
we will discuss this later in section 2 in detail. We also dy- dynamic nature of the wireless ad hoc networks makes it very
namically adjust the forwarding probability based on the node unlikely that there is at least one location server available for
density estimated by the current node. Our results show that, each node. Thus, we will concentrate on distributed location
by using appropriate optimization heuristics, we can save up services.
to 94% messages even compared with the global flooding For the wireless ad hoc networks, the location service pro-
method. vided can be classified into four categorizes: some-for-all,
The remaining of this paper is organized as follows. In sec- some-for-some, all-for-some, all-for-all. Some-for-all service
tion 2, we review some known location services techniques means that some wireless nodes provide location services for
for wireless ad hoc networks. We study our regional gossip all wireless nodes. Other categorizations are defined simi-
method in detail in section 3. We demonstrate its effective- larly.
ness by both theoretical study and extensive simulations in An example of all-for-all services is the location services
section 4 . We also study the effectiveness of the regional gos- provided in the Distance Routing Effect Algorithm for Mo-
siping on constructing multiple paths for any pair of source bility (DREAM) by Basagni et al. [38]. Each node stores a
and destination nodes in section 5. We conclude our paper database of the position information for all other nodes in the
and discuss possible future research directions in section 6. 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
2. Preliminaries the cost of updating and the accuracy of the database.
Using the idea of quorum developed in the databases and
We consider a wireless ad hoc network (or sensor network) distributed systems, Hass and Liang [39] and Stojmenovic
with all nodes distributed in a two-dimensional plane. As- [40] developed quorum based location services for wireless
sume that all wireless nodes have distinctive identities and ad-hoc networks. Given a set of wireless nodes V , a quorum
each static wireless node knows its position information1 ei- system is a set of subset (Q1 , Q2 , . . . , Qk ) of nodes whose
ther through a low-power Global Position System (GPS) re- union is V . These subsets could be mutually disjoint or often
ceiver or through some other way. For simplicity, we also as- have equal number of intersections. When one of the nodes
1 More specifically, it is enough for our protocol when each node knows the requires the information of the other, it suffices to query one
relative position of its one-hop neighbors. The relative position of neigh- node (called the representative node of Qi ) from each quo-
bors can be estimated by the direction of arrival and the strength of signal. rum Qi . A virtual backbone is often constructed between
REGIONAL GOSSIP ROUTING 63
the representative nodes using a non-position-based methods Let G(V , r) be the graph defined on V with edges uv E
such as [4144]. The updated information of a node v is sent if and only if uv r where uv is the Euclidean dis-
to the representative node (or the nearest if there are many) of tance between nodes u and v. Let G (Xn , rn ) be the set of
the quorum containing v. The difficulty of using quorum is graphs G(V , rn ) for n nodes V that are uniformly and in-
that the mobility of the nodes requires the frequent updating dependently distributed in a two-dimensional region . The
of the quorums. The quorum based location service is often problem considered by Gupta and Kumar [5] is then to deter-
some-for-some type. mine the value of rn such that a random graph in G (Xn , rn )
The other promising location service is based on the is asymptotically connected with probability one as n goes to
quadtree partition of the two-dimensional space [45]. It di- infinity, when is a unit disk. Specifically, they showed that
vides the region containing the wireless network into hierar- G(V , rn ) is connected almost surely if nrn2 ln n + c(n)
chy of squares. The partition of the space in [45] is uniform. for any c(n) with c(n) as n goes to infinity, and
However, we notice that the partition could be non-uniform G(Xn , rn ) is asymptotically disconnected with positive prob-
if the density of the wireless nodes is not uniform for some ability if nrn2 = ln n + c(n) and lim supn c(n) < +. It is
applications. Each node v will have the position information unknown whether the same result holds if the geometry do-
of all nodes within the same smallest square containing v. main in which the wireless nodes are distributed is a unit-area
This position information of v is also propagated to up-layer square instead of the unit-area disk.
squares by storing it in the node with the nearest identity to v Independently, Penrose [47] showed that the longest edge
in each up-layer square containing v. Using the nearest iden- Mn of the minimum spanning tree of n points randomly and
tity over the smallest identity, we can avoid the overload of uniformly distributed in a unit area square C satisfies that
some nodes. The query is conducted accordingly. It is easy to
lim Pr nMn2 ln n = ee ,
show that it takes about O(log n) time to update the location n
of v and to query another nodes position information. for any real number . This result gives the probability of the
If the location service is not provided, the nodes can cache network to be connected if the transmission radius is set as a
the location information of some other nodes. When the positive real number r when n goes to infinity. For example,
source node wants to send a message to the target, it directly if we set = ln ln n, we have
uses the region gossip if the target location is known. Oth-
erwise, it will use flooding (with selective forwarding [46] Pr nMn2 ln n + ln ln n = e1/ ln n .
to control the number of messages sent) to send the message
It implies that the network is connected with probability at
to all nodes within k hops, where k is a parameter to be set.
least e1/ ln n if the transmission radius rn satisfies nrn2 =
Then if a node within k hops knows the destination location,
ln n + ln ln n. Notice that e1/ ln n > 1 1/ ln n from ex >
that node then starts the regional gossip to send message to
1 x for x > 0. By setting = ln n, the probability that
the destination.
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
2.2. Random deployment and connectivity only true when n goes to infinity. When n is a finite number,
then the probability of the graph being connected is smaller.
Energy conservation is critical for the life of the wireless net- In [48], Li et al. presented the experimental study of the prob-
work. One approach to save energy is to use the minimum ability of the graph G(V , rn ) being connected for finite num-
power to transmit the signal without disconnecting the net- ber n.
work. The universal minimum power used by all wireless Gupta and Kumar [5] conjectured that if every node has
nodes, such that the induced network topology is connected, probability p of being fault, then the transmission range for
is called the critical power. Determining the critical power resulting a connected graph satisfies prn2 = log n/n. This
for static wireless ad hoc networks is well-studied [5,7,13]. It was recently confirmed by Wan et al. [49]. It is not difficult
remains to study the critical power for connectivity for mo- to see that whether the global gossip can deliver the packet
bile wireless networks. As the wireless nodes move around, it is related to whether a set of randomly deployed nodes in a
is impossible to have a unanimous critical power to guarantee region form a connected graph when each node has a uniform
the connectivity for all instances of the network configuration. faulting probability p. Consequently, given a wireless net-
Thus, we need to find a critical power, if possible, at which work with n nodes , each with transmission range r, the relay
each node has to transmit to guarantee the connectivity of the probability of a gossip routing protocol is p = log n/(nrn2 ),
network almost surely, i.e., with high probability almost one. when n goes to infinity. We conjecture that this is true for any
The wireless nodes are randomly deployed in majority non-flat convex region .
wireless ad hoc networks either due to its massive number,
due to its emergency requirement, or due to harsh environ- 2.3. Fault tolerance and security
ment. For simplicity, we assume that the n wireless devices
are distributed in a unit area square (or disk) according to Fault tolerance is one of the central challenges in designing
some distribution function, e.g., random uniform distribution, the wireless ad hoc networks. To make fault tolerance pos-
denoted by Xn , or Poisson process, denoted by Pn . sible, first of all, the underlying network topology must have
64 LI ET AL.
multiple disjoint paths to connect any two given wireless de- 3. Regional gossip
vices. Here the path could be vertex disjoint or edge disjoint.
Considering the communication nature of the wireless net- Although gossiping reduces the routing messages compared
works, the vertex disjoint multiple paths are often used in the with flooding, it still produces lots of unnecessary messages
literature. A graph is called k-vertex connected (k-connected in regions that are far away from the line between the source
for simplicity) if, for each pair of vertices, there are k mu- and the target node. Notice that, the traditional gossip will
tually vertex disjoint paths (except end-vertices) connecting propagate the message to the whole network. To further re-
them. A k-connected wireless network can sustain the failure duce the number of forwarding messages, we propose re-
of k 1 nodes. gional gossiping, in which essentially only nodes inside some
The connectivity of random graphs, especially the geomet- region (derived from the source and target) will execute the
ric graphs and its variations, have been considered in the ran- gossiping protocol, and nodes outside the region will not par-
dom graph theory literature [50], in the stochastic geometry ticipate the gossiping at all. The region we select in our sim-
literature [47,5154], and the wireless ad hoc network litera- ulations are some ellipses using the source and target as foci.
ture [2,5,5561]. We now describe our regional gossiping routing method
Penrose [53] showed that a graph of G(Xn , r) becomes in detail. Assume that wireless mobile hosts are a set V of
k-connected almost surely at the moment it has minimum de- n points distributed in a two-dimensional space. Each node
gree k. However, this does not mean to guarantee a graph over has a fixed transmission range r: all nodes within distance r
n points is k-connected almost surely, we only have to con- to a node v can receive the signal sent by v. Thus, all mo-
nect every node to its k nearest neighbors. Let V be a set of bile hosts define a communication graph G(V , r) in which
n points randomly and uniformly distributed in a unit square there is an edge uv iff uv r. From now on, we also
(or disk). Xue and Kumar [61] proved that, to guarantee that assume that the source node knows the position of the target
a geometry graph over V is connected, the number of nearest node, the global ellipse factor , in addition to its own posi-
neighbors that every node has to connect must be asymptot- tion. Every mobile host can get its own position through a
ically (ln n). Dette and Henze [51] studied the maximum low-cost GPS. In many applications such as data-centric sen-
length of the graph by connecting every node to its k nearest sor network, there is only a fixed number of destination nodes
neighbors asymptotically. For the unit volume sphere, their (called sink), which is often static, thus every node knows the
result implies that, when k > 2, positions of these possible target nodes. Otherwise, location
service is needed to find the location of the destination node.
2
lim Pr nrn,k ln n + (2k 3) ln ln n 2 ln(k 1)! The geometry information of the source node and the destina-
n tion node and also the current route (i.e., the route from source
2(k 2) ln 2 + ln + 2 = ee . to the sender of the message) is piggybacked along with the
message packet. When a node, say v, receives a message, it
Li et al. [48] showed that, given n random points V over retrieves the geometry position of the source node and the tar-
a unit-area square, to guarantee that a geometry graph get node. Node v then checks if it is inside the ellipse defined
over V is (k + 1)-connected, the number of nearest neigh- by using the source point s and the destination point t as foci.
bors that every node has to connect is asymptotically (ln n+ Notice that, a node v is inside this ellipse iff
(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 vs + vt st,
distributed in C such that the graph G(V , rn ) is k-connected which can be checked trivially. When a node is not inside the
with high probability. ellipse, it will just simply discard this message. Otherwise,
The theoretical value of the transmission ranges gives us with a fixed probability p, the node forwards this message to
insight on how to set the transmission radius to achieve the all nodes within its transmission range. Hereafter, we call p
k-connectivity with certain probability. These results also ap- the relay probability and the ellipse factor of our regional
ply to mobile networks when the moving of wireless nodes gossiping method. Obviously, the probability that the desti-
always generate randomly (or Poisson process) distributed nation node receives the message depends on the relay prob-
node positions. Bettstetter [2] conducted the experiments to ability p, the ellipse factor , the number of nodes n, and the
study the relations of the k-connectivity and the minimum transmission range r.
node degree using toroidal model. Li et al. [48] also con- Gupta and Kumar [5] showed that a random graph G(V , r)
ducted experiments to study the probability that a graph has is connected whenever r is larger than some threshold
minimum degree k and has vertex connectivity k simultane- value rn . It is known that the global gossiping (by simply set-
ously using Euclidean model. Recently, Bahramgiri et al. [8] ting to ) exhibits some bimodal behavior: the destination
showed how to decide the minimum transmission range of node receives the message if and only if the relay probability
each node such that the resulted directed communication is larger than some threshold value. We expect our regional
graph is k-connected. Here it assumes that the unit disk graph gossiping method to have the similar transmission phenom-
by setting each node with the maximum transmission range ena.
is k-connected. Lukovszki [62] gave a method to construct a We then estimate the relay probability for a network of n
spanner that can sustain k nodes or k links failures. nodes. It was shown in [49] that given n wireless nodes dis-
REGIONAL GOSSIP ROUTING 65
tributed in a unit square and each node has transmission range 0.23. When the ellipse factor = , we can estimate the
rn and being off or fault with probability p, then the network relay probability of the regional gossiping as
is connected with high probability if pnrn2 2 ln n. Con- ln n
sider the network of n nodes distributed in a square region p = 0.495.
n r 2
with side length a. Assume that the distance between the
source and the target is d and the ellipse factor is . The The actual relay probability should be larger, so do the per-
number of nodes inside the ellipse is then about centage of vertices involved in global gossiping. The exper-
iments discussed in the following sections verify the above
n 2 1 2 study.
Nd = 2 d .
a 4
Since each node inside the ellipse forwards the message with 4. Experimental studies
probability p after it receives the message, to let the target
receive the message almost surely, the subnetwork composed 4.1. Simulation environment
of the nodes inside the ellipse with fault probability p must
be connected. In other words, the relay probability in our We conducted extensive simulations to study the performance
regional gossiping is at least of our region gossiping method. We model the network by
unit disk graph and the mobile hosts are randomly placed
ln Nd + c(Nd )
p . in a square region. We tried unit disk graphs with different
Nd (r/a)2 number of vertices that are randomly placed in a 15 15
Here r is the transmission range of each wireless node and square. Notice that the density of the graph must be above
c(Nd ) is a number going to when Nd goes to . The some threshold to see the effectiveness of the algorithm oth-
probability that the network (each node is chosen with prob- erwise the properties would be hidden and cannot be seen. In
c(N )
ability p) is connected is ee d . Substituting in Nd , we other words, the algorithm works better for dense graphs than
have sparse graphs with the same parameters p and .
There are different parameters involved in our simulations,
4a 4 ln(n 2 1d 2 /(4a 2 )) ln(n 2 d 2 /4) which are described as follows:
p = .
2 d 2 r 2 2 1 n n 2 2 d 2 r 2 /4
Number of vertices. We tried graphs with 1000, 1500 and
Here 2 = 2 1, d = d/a, and r = r/a. Since for a 2000 vertices. For convenience, we use n to denote the num-
random pair of source and target nodes, d 2a, we have ber of vertices.
ln(n 2 /4)
p . Ellipse factor. In each iteration of the simulation, the source
n 2 2 r 2 /4 vertex and the target vertex are the foci of an ellipse with el-
For example, consider a network of n = 1000 nodes distrib- lipse factor chosen from 1.2, 1.4, 1.6, 1.8 and 2. We also
uted in a square of side length a = 15, and each node has consider the case where the ellipse factor is which is just
transmission range r = 1. For ellipse factor = 1.2, we can the traditional global gossiping method. The smaller the el-
calculate the relay probability p such that the regional gossip- lipse factor is, the narrower the ellipse will be. Notice that
ing routing can deliver the packets almost surely as ellipse factor must be greater than one.
Beside the above parameters there are two more constant the message. It is important because sending message con-
metrics used in our simulations as follows: sumes energy and energy is a bottleneck for wireless nodes.
In all the figures of this paper the Y-axis is either the mes-
Source-target pairs. To compute the exact value of the av- sage delivery rate or the percentage of vertices that are in-
erage delivery rate, we have to try all possible pairs for each volved in message delivery, and the X-axis is one of the pa-
graph, which is n(n1), where n is the number of vertices. It rameters with respect to another parameter which is shown
is not feasible to test all pairs when n is large. Instead we ran- in the legend and the remaining two parameters are fixed.
domly select 100 pairs for each graph and conduct regional For example, we can show message delivery rate as a func-
gossiping based routing for each pair. Although we are not tion of relay probability p for different values of ellipse fac-
testing all possible pairs, choosing 100 random pairs would tor , while the transmission range r and the number of ver-
give the results close enough to exact values. tices n are fixed (see figure 1). Each point in each figure rep-
resents the average of the 100, 000 iterations since we will
Number of trys. The delivery probability (called delivery test 100 different source-target pairs, and each pair is tested
rate also) of our regional gossiping method for a pair of nodes 1000 times, when all four parameters are fixed.
is defined as the probability that the destination node receives We believe that the relay probability and the graph density
the message. To compute the delivery rate, we tried sending are two major factors of message delivery rate. On the other
the message 1000 times for each pair and then the delivery hand, the ellipse factor and the relay probability are the major
rate is approximated by the total number of times that the factors determining the number of vertices that are involved
message reached the target divided by the total number that in message delivery. Here a node is said to be involved if it
the message is sent (which is 1000 in out simulations). relays the message. In other words, when the Y-axis is the
message delivery rate and X-axis is either relay probability,
There are four different types of nodes in each iteration of number of vertices or transmission range, we expect to see a
our simulations: jump in the figures. It means that when the X-axis exceeds
some threshold, then the regional gossiping method almost
(1) Not in ellipse. Nodes that are out of the ellipse region.
surely guarantees that the message arrives at the target. When
(2) Blocked. Nodes that receive the message and do not relay the X-axis is less than some threshold, the target almost never
it. gets the message.
(3) Relayed. Nodes that receive and relay the message.
4.2. Message delivery rate as a function of relay probability
(4) Initial hops nodes. The nodes within the initial hops al-
ways receive the message and from those, the ones whose We first conduct extensive simulations to study the effect of
distance from source is less than some fix initial hops pa-
the relay probability on the message delivery rate. Intuitively,
rameter, always relay the message. Other nodes inside the
if we increase the relay probability, the message delivery rate
ellipse relay the message with the given relay probability.
increases. Besides the relay probability, we vary either the el-
Here we want to involve as little nodes as possible. In other lipse factor , or the number of vertices n, or the transmission
words, we want to minimize the number of nodes that relay range r. Now we discuss them one by one as follows.
(a) (b)
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.
REGIONAL GOSSIP ROUTING 67
1. Message delivery rate as a function of relay probability for when the relay probability is 0.3 with ellipse factor of 1.4, the
different values of ellipse factor. As can be seen in figure 1, delivery rate is about 95% for n = 2000, while we cannot
when the probability exceeds some threshold the delivery rate achieve this rate when n = 1000 (see figure 1(a)).
jumps from near 0% to near 100%. In each figure, this thresh-
old decreases as the ellipse factor increases because the bigger 2. Number of nodes involved in message delivery as a func-
the ellipse factor is, the more vertices contribute in message tion of relay probability for different values of ellipse factor.
delivery, and consequently, the probability of the message to So far, we have concentrated on the transition phenomena of
reach the target, which is nothing but the message delivery the delivery rate over the relay probability. Not only the deliv-
rate, increases. For both figures the transmission range is ery rate is important for the network performance, but also the
fixed to 1 unit and the number of vertices is also fixed to 1000 number of vertices involved in the message delivery is impor-
and 2000, respectively. tant for the network life since the wireless devices are often
From figure 1, we observe that when the graph becomes
powered by the batteries only.
denser (more vertices in this case), the curve jumps earlier,
The challenge is to find an ellipse factor and a relay prob-
and the reason is each time a vertex relays the message,
ability such that not only the delivery rate is high (close to
more nodes get the message (due to more neighbors in dense
100%) but also the number of vertices involved in the message
graphs) so the probability that the message reaches the target
delivery is as small as possible. Actually the ellipse factor
increases.
and the number of vertices involved in sending the message
One important observation is as follows: as we increase
from source to target, work against each other. It means that
the ellipse factor, the message delivery rate does not increase
proportionally. Surprisingly, when the ellipse factor is around if we choose a bigger ellipse factor, a higher delivery rate is
1.8, the message delivery rate is almost as good as the one achieved, on the other hand, lots of vertices will be involved
using global gossiping (i.e., the ellipse factor constraint is in route discovery. In reverse, if we choose a small ellipse
relaxed to ). The reason is where a bigger ellipse factor factor then fewer vertices will be involved but it may not have
is used we are actually considering the vertices that are less a good delivery rate.
helpful than the vertices which are already considered. Intu- As can be seen in figure 2, the relation between the num-
itively, the vertices, which are far away from the line connect- ber of vertices involved and the relay probability with respect
ing the source and target, do not help improving the message to ellipse factors is close to linear. The bigger the relay prob-
delivery rate. ability, the more number of vertices will be involved in the
We also observe that, for a fixed relay probability, when the message delivery. The exact relation between the number of
graph is dense, even a narrow ellipse could guarantee a good vertices and relay probability is not simple. Clearly, the far-
rate of message delivery. Achieving the same delivery rate ther it is from the source, the less probability it will get the
using the same relay probability, for a sparser graph, might the message to relay.
not be possible, even if the ellipse factor is relaxed to infin- In figure 2 when the ellipse factor is infinity, we are ac-
ity. In other words, the ellipse factor does not compensate the tually flooding the network with a uniform relay probability,
description of the graph density. For example, in figure 1(b), and when this relay probability is 1, the network is completely
(a) (b)
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.
68 LI ET AL.
flooded, i.e., traditional flooding, so all nodes have the chance volved (see figures 1 and 2) by choosing ellipse factor 1.2
to contribute in message delivery. and relay probability 0.5.
Assume that we want to have the delivery rate more than So far the transmission rang was fixed to 1. We were mo-
99%, first consider the case in which we have 1000 nodes, tivated to study the effect of transmission range as well. We
illustrated in figures 1(b) and 2(b). then tried different values of transmission range. Obviously
We build the table 1 as follows: for each ellipse factor, we the larger the transmission range is, the denser the graph will
can find the needed relay probability to guarantee the message be and as mentioned before that causes the jump to occur ear-
delivery above 99% from figure 1, and then by knowing the lier.
values of ellipse factor and the relay probability we can find In figure 3 the transmission range is 2. See how similar
the percentage of vertices that are involved from figure 2. figure 1 and figure 3 are, the only difference between these
For example, to achieve this rate (above 99%) when ellipse two figures is the probability at which the jump occurs for
factor is 1.2, the relay probability must be at least 0.9 (see fig- any fixed ellipse factor. Since in delivery rate happens earlier
ure 1). Then having these two values fixed, we can find the and quicker when the transmission range increases, we plot
number of nodes that are involved from figure 2, which would
the figures using relay probability range [0, 0.3] for r = 2,
be about 15% of all vertices. Doing the same thing for differ-
instead of [0, 1] for r = 1.
ent values of ellipse factor, we get table 1.
Again assume that we want to have the delivery rate more
The first column is the different ellipse factors we simu-
than 99%. Consider the case in which we have 1000 nodes,
lated, and the second column is the corresponding relay prob-
but the transmission range is 2 (figures 3(a) and (b)).
ability in our regional gossip method to guarantee this fixed
delivery rate 99%, and the third column is the percentage of We build table 2 as we built table 1: for each ellipse factor.
vertices that are involved in our regional gossiping (i.e., re- We can find the relay probability that guarantees the message
laying the message). delivery rate above 99% from figure 3(a), and then by know-
Table 1 shows that we could involve only 15% of vertices ing the values of ellipse factor and the relay probability we
to guarantee the message delivery rate above 99% when the can find the percentage of vertices involved in message deliv-
ellipse factor is 1.2. If we do the same calculations where ery from figure 3(b).
there are 2000 nodes then only 10% of vertices will be in- For example, to achieve this rate (above 99%) when el-
lipse factor is 1.2, the relay probability must be 0.3 (see fig-
Table 1
Percentage of the vertices involved in message delivery.
ure 3(a)). Then having these two values fixed, we can find the
number of nodes involved from figure 3(b), which would be
Ellipse factor Relay probability Vertices involved (%)
about 8%. Doing the same thing for different values of ellipse
1.2 0.9 15 factor, we get table 2.
1.4 0.8 22 Table 2 shows that we could involve only 8% of vertices
1.6 0.7 25
to guarantee the message delivery rate above 99% for net-
1.8 0.7 30
infinity 0.7 70 works of 1000 nodes and with transmission range equal to 2.
If we do the same calculations for networks of 2000 nodes
(a) (b)
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.
REGIONAL GOSSIP ROUTING 69
with transmission range equal to 2, then only 6% of vertices (see figure 4(a)). Then having these two values fixed, we
will be involved (figures are not shown here). can find the percentage of vertices involved from figure 4(b),
which would be about 30%. We get table 3 by doing the same
3. Message delivery rate as a function of relay probability calculation for different values of transmission range.
for different values of transmission range. So far we plotted Table 3 illustrates the number of vertices involved in the
the message delivery rate as a function of relay probability for
regional gossip routing to guarantee a fixed delivery rate 99%
different values of ellipse factor. Let us replace the ellipse fac-
for networks of 1000 nodes with ellipse factor 1.6. Observe
tor parameter with transmission range and see how the graph
that, all these curves intersect in a common point when the re-
behaves.
lay probability is 1. Because the ellipse factor is fixed, chang-
As you can see in figure 4, transmission range plays a very
ing the transmission range does not change the number of
important role in message delivery (see how far the curves are
nodes that are inside ellipse, which is total number of vertices
from each other). As the transmission range is increased, the
delivery rate improves significantly as opposed to the situa- involved in message delivery when the relay probability is 1.
tion we had earlier with ellipse factor. The reason is when the Actually it is possible to have a node in the ellipse which does
transmission range is bigger then each node will be connected not contribute in message delivery even when the relay prob-
to more nodes, in other words the graph density increases. ability is 1, but that is very unlikely. It happens only when
Thus, each time a node relays the message, more nodes will a node in the ellipse doesnt have any neighbor inside the el-
get it and the probability that the message dies out becomes lipse. In our simulations this scenario happened 2 times out
smaller. Here in figure 4, the ellipse factor is fixed to 1.6. of 180,000,000 iterations.
We built table 3 as follows: for each transmission range, Another observation is that we get different curves for dif-
we can find the relay probability that guarantees the mes- ferent transmission ranges. Typically, when the transmission
sage delivery rate above 99% from figure 4(a), and then by range is larger, more nodes inside this ellipse will be involved
knowing the values of transmission range and the relay prob- in the message delivery.
ability we can find the percentage of vertices involved from
figure 4(b). 4. Message delivery rate as a function of relay probability
For example, to achieve this rate (above 99%) when trans- for different number of nodes. In our simulations we stud-
mission range is 1, the relay probability must be at least 0.8, ied networks with different densities in two different ways.
Table 2 Table 3
Percentage of the vertices involved in message delivery. Percentage of the vertices involved in message delivery.
Ellipse factor Relay probability Vertices involved (%) Transmission range Probability Vertices involved (%)
(a) (b)
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.
70 LI ET AL.
(a) (b)
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- 4.3. Message delivery rate as a function of ellipse factor
works with fixed number of vertices and different transmis-
sion ranges. Now we study networks with fixed transmission We can look at the problem from a totally different point of
view. So far we have concentrated on the transition phenom-
range and different number of vertices placed in a 15 15
ena of the delivery rate over the relay probability. In other
square. In both cases we expect the similar results if the net-
words, in all figures the X-axis was the relay probability. Now
work densities are similar.
let us see how the network behaves if we use different ellipse
As you can see in figure 5, the number of vertices plays factors while some other parameters are fixed. We found that,
an important role in message delivery (see how far the curves regardless of the network density and relay probability, in-
are from each other). Here we have the same reasoning as the creasing the ellipse factor does not improve the message de-
previous section. As the number of vertices is increased, the livery rate significantly.
delivery rate improves significantly. The reason is when there
are more vertices in the same area, the graph becomes denser. 1. Message delivery rate as a function of ellipse factor for
Thus, each time a node relays the message more nodes will different values of transmission range. First let us fix the re-
get it and the probability that the message dies out becomes lay probability and the number of vertices. Remember that to
change the message delivery rate dramatically we can either
smaller.
increase the relay probability or increase the network density.
Now let us look at the percentage of nodes that are involved
As can be seen in figure 6 there is no jump. In other words,
in message delivery as a function of relay probability for dif-
increasing the ellipse factor does not improve the message de-
ferent number of nodes (see figure 5). Remember that in this livery rate dramatically.
case ellipse factor and transmission range are fixed. Here we Figure 6 shows when the relay probability is fixed, regard-
have the same ellipse with different number of vertices in- less of the value of ellipse factor, the graph density must be
side them. When there are more vertices in the same area above some threshold to guarantee a high message delivery.
the message is delivered with higher probability since more As you can see in figure 6(a) when the transmission range
nodes will relay the message. Notice that, given a fixed relay is less than 1.5 then the delivery rate is always below 20%
probability, when the node density exceeds some threshold even if the ellipse factor constraint is relaxed (the case where
(depending on the relay probability) almost all nodes inside ellipse factor constraint is relaxed and not shown in figure 6).
the ellipse will receive the message, thus, have the chance to As it is expected if we set the relay probability to a higher
relay the massage. In other words, if the relay probability is value then the delivery rate would be higher. This is illus-
low, high message delivery rate still can be achieved if the 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
graph is dense enough and if the graph is sparse, high mes-
shifted up.
sage delivery rate still can be achieved by increasing the relay
probability. On the other hand, larger relay probability will 2. Message delivery rate as a function of ellipse factor for
involve more nodes in message delivery (the number of nodes different number of vertices. As mentioned earlier, the net-
involved is almost linear to the relay probability as shown in work density can be increased either by increasing the trans-
right figure of figure 5). 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 the message delivery rate (as a function of ellipse factor for
number of vertices and we expect similar results. In other different values of relay probability) by fixing the number of
words, let us fix the relay probability and the transmission nodes and the transmission range.
range to see the delivery rate as a function of ellipse factor for In figure 8 when the relay probability is below some
different number of vertices. threshold, a high delivery rate cannot be achieved even when
Again, as can be seen in figure 7 there is no jump. In other the ellipse factor constraint is relaxed. Figure 8 is similar to
words, increasing the ellipse factor does not improve the mes- figures 6 and 7 due to the fact that a high relay probabil-
sage delivery rate dramatically.
ity can compensate the sparseness of the network and vice
3. Message delivery rate as a function of ellipse factor for versa.
different values of relay probability. In the previous two sec- Intuitively, all the discussions of the two previous sections
tions, we studied the effect of ellipse factor in networks with apply to this section too. For example, when the network
different densities, in this section instead of changing the net- density is larger than some threshold, the number of vertices
work density, we change the relay probability. Thus, in this involved is almost linear to the ellipse factor, see figures 7
section, the network density is fixed. Specifically, we study 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 ellipse factor and the number of vertices. We expect to see
range jump because in each curve the graph density changes and
also we expect to see curves that are far from each other due
We can look at the problem from a totally different point of to the fact that for each curve the relay probability is fixed.
view. So far the X-axis was the relay probability or the ellipse As you can see in figure 9 when the relay probability is
factor. Thus, for each curve in figures discussed in previous bigger the jump occurs earlier. This figure is similar to fig-
sections, the network density was fixed. But if we choose the ure 4 due to the fact that the relay probability and transmis-
transmission range or number of vertices as the X-axis then sion range both improve the message delivery rate signifi-
the graph density changes for each curve. We first study the cantly.
case where the X-axis is the transmission range and in the
2. Message delivery rate as a function of transmission range
next section we study the case where the X-axis is the the
for different values of ellipse factor. Let us fix the number
number of vertices.
of vertices and the relay probability to see the delivery rate as
1. Message delivery rate as a function of transmission range a function of transmission range for different values of ellipse
for different values of relay probability. First let us fix the factor. As you can see in figure 10, like figure 1, as we in-
REGIONAL GOSSIP ROUTING 73
(a) (b)
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- 4.5. Message delivery rate as a function of number of
crease proportionally. The only difference between figure 10 vertices
and figure 1 is: in figure 10 the network density changes in
each curve but in figure 1 the relay probability changes in The last parameter is the number of vertices. Since both trans-
each curve. Since increasing either the relay probability or mission range and number of vertices affect the network den-
transmission range improves the message delivery, exchang- sity, we expect similar results like the previous section.
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 1. Message delivery rate as a function of number of vertices
global gossiping. for different values of relay probability. Now let us fix the
Observe that, in figure 10, the number of vertices involved ellipse factor and the transmission range to see delivery rate
in message delivery is almost linear after the transmission as a function of number of vertices for different values of re-
range is larger than some threshold (almost 2). When the lay probability. As shown in figure 12, if we use a big enough
transmission range is small, the number of nodes involved is relay probability, a high delivery rate is guaranteed. But when
small since the message quickly dies out (the relay probability the relay probability is small then we need a large number of
is 0.2 here). vertices to compensate this small relay probability to guaran-
tee a high delivery rate.
(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- chance to receive the message more than once. Observe that
pens. figure 15 is a little bit misleading. It shows that with a narrow
ellipse and the replay probability fixed to 1 the probability
5. Fault tolerance that the target receives the message more than once is below
95%. The reason is in our simulations, the sourcetarget pairs
To study the fault tolerance of the ad-hoc networks, we simu- are chosen randomly, so in some cases the target is only one
lated the cases in which the target receives the message more
hop away from the source, thus the target gets the message for
than once. The figure 15 shows the number of times that the
sure but at the same time, due to the closeness of source and
message is delivered to the target at least twice as a function
of relay probability for different values of ellipse factor. If target, there might not be another neighbor inside the ellipse
target has h neighbors inside the ellipse in the best case (i.e., for target. Thus the target has no chance to receive the mes-
all neighbors of the target receive the message) we expect the sage more than once. In other words, in some cases, although
message to be delivered p h times. Note that if the target the message delivery rate is 100%, the chance that the target
has only one neighbor inside the ellipse, then the target has no 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 performance. We doubt this due to two reasons: (1) the back-
bone formed by clusterheads are already very sparse, and to
We proposed a regional gossip approach, where only the guarantee that all nodes receive the messages, the gossiping
nodes within some region forward the routing message with probability is very high; and (2) the communication cost to
some probability, to reduce the overhead of the routing pro- maintain the backbone will also offset the benefit gained by
tocol imposed on the network. We showed how to set the for- global gossiping, if there is any. We will conduct simulations
warding probability based on the region and the estimated to study this.
network density both by theoretical analysis and by exten- One of the main questions remaining to be studied is to use
sive simulations. Our simulations showed that the number of non-uniform ellipse factors. In our simulations, the ellipse
messages generated using this approach is less than the sim- factor is uniform regardless of the distance between source
ple global flooding (up to 94%), which already saves many and target. We believe that using a bigger ellipse factor, when
messages compared with global flooding. the source and target are close, will get better results.
Hass et al. [24] expected that the global gossiping com- Another question is studying networks with different den-
bined with the cluster-based routing can further improve the sities, meaning that instead of trying different transmission
76 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).
Figure 15. The number of times that the message receives the target more [14] R. Wattenhofer, L. Li, P. Bahl and Y.-M. Wang, Distributed topology
than once as a function of relay probability for different values of ellipse control for wireless multihop ad-hoc networks, in: IEEE INFOCOM01
factor. Here transmission range is 1 and number of vertices is 1000. (2001).
[15] S. Singh, M. Woo and C. Raghavendra, Power-aware routing in mobile
ranges and different number of nodes, networks with differ- ad hoc networks, in: Proc. of MOBICOM (1998) pp. 181190.
ent densities can be studied. To generate a network with [16] I. Stojmenovic and X. Lin, Power aware localized routing in ad hoc
wireless networks, IEEE Transactions on Parallel and Distributed Sys-
a given density with respect to transmission range, we can tems 12(11) (2001) 11221133.
keep adding nodes to the network until the desired density is [17] J. Wu, F. Dai, M. Gao and I. Stojmenovic, On calculating power-
reached. aware connected dominating sets for efficient routing in ad hoc wireless
We had assumed that two nodes can always communicate networks, IEEE/KICS Journal of Communication and Networks 4(1)
if their distance is no more than the transmission range. How- (2002) 5970.
[18] P. Bose, P. Morin, I. Stojmenovic and J. Urrutia, Routing with guar-
ever, this is not totally true practically. Some pair of nodes
anteed delivery in ad hoc wireless networks, ACM/Kluwer Wireless
cannot communicate at all even if they are close. We can Networks 7(6) (2001) 609616, also in: 3rd Internat. Workshop on
model this by assigning another link probability pl : a link ex- Discrete Algorithms and Methods for Mobile Computing and Commu-
ist with probability pl . Here probability pl could be uniform nications (1999) pp. 4855.
or dependent on the distance between the pair of nodes. [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
References distance-vector routing (DSDV) for mobile computers, Computer
Communications Review (October 1994) 234244.
[1] S. Capkun, M. Hamdi and J.P. Hubaux, GPS-free positioning in mobile
[21] D.B. Johnson and D.A. Maltz, Dynamic source routing in ad hoc
ad-hoc networks, in: Proc. Hawaii Internat. Conf. on System Sciences
wireless networks, in: Mobile Computing, eds. Imielinski and Korth,
(2001).
Vol. 353 (Kluwer Academic, Dordrecht, 1996).
[2] C. Bettstetter, On the minimum node degree and connectivity of a wire-
[22] B. Karp and H.T. Kung, Gpsr: Greedy perimeter stateless routing for
less multihop network, in: 3rd ACM Internat. Symposium on Mobile Ad
wireless networks, in: ACM/IEEE Internat. Conf. on Mobile Comput-
Hoc Networking and Computing (MobiHoc02) (June 2002).
ing and Networking (2000).
[3] G. Calinescu, I. Mandoiu and A. Zelikovsky, Symmetric connectivity
[23] Y.-B. Ko and N.H. Vaidya, Using location information to improve rout-
with minimum power consumption in radio networks, in: IFIPTCS
ing in ad hoc networks, Technical Report, Department of Computer
(2002) to appear.
Science, Texas A&M University (1997).
[4] A. Clementi, P. Penna and R. Silvestri, The power range assignment
[24] Z. Haas, J. Halpern and L. Li, Gossip-based ad hoc routing, in: IEEE
problem in radio networks on the plane, in: XVII Symposium on The-
INFOCOM (2002).
oretical Aspects of Computer Science (STACS00), Lecture Notes in
Computer Science, Vol. 1770 (Springer, New York, 2000) pp. 651660. [25] S. Murthy and J. Garcia-Luna-Aceves, An efficient routing protocol for
[5] P. Gupta and P. R. Kumar, Critical power for asymptotic connectivity in wireless networks, ACM Mobile Networks and Applications (Special
wireless networks, in: Stochastic Analysis, Control, Optimization and Issue on Routing in Mobile Communication Networks) 1(2) (1996).
Applications: A Volume in Honor of W.H. Fleming, eds. W.M. McE- [26] V. Park and M. Corson, A highly adaptive distributed routing algorithm
neaney, G. Yin and Q. Zhang (1998). for mobile wireless networks, in: IEEE INFOCOM (1997).
[6] J. Monks, V. Bharghavan and W.-M Hwu, A power controlled multi- [27] C. Perkins, Ad-hoc on-demand distance vector routing, in: MILCOM
ple access protocol for wireless packet networks, in: IEEE INFOCOM 97 (November 1997).
(2001). [28] C. Perkins and P. Bhagwat, Highly dynamic destination-sequenced
[7] M. Sanchez, P. Manzoni and Z. Haas, Determination of critical trans- distance-vector routing, in: Proc. of the ACM SIGCOMM (October
mission range in ad-hoc networks, in: Multiaccess, Mobility and Tele- 1994).
traffic for Wireless Communications (MMT99) (1999). [29] P. Sinha, R. Sivakumar and V. Bharghavan, CEDAR: Core extraction
[8] M. Bahramgiri, M.T. Hajiaghayi and V.S. Mirrokni, Fault-tolerant distributed ad hoc routing algorithm, IEEE Journal on Selected Areas
and 3-dimensional distributed topology control algorithms in wireless in Communications 17(8) (1999) 14541465.
REGIONAL GOSSIP ROUTING 77
[30] P. Sinha, R. Sivakumar and V. Bharghavan, Enhancing ad hoc rout- [52] M. Penrose, Extremes for the minimal spanning tree on normally dis-
ing with dynamic virtual infrastructures, in: Proc. of IEEE INFOCOM tributed points, Advances in Applied Probability 30 (1998) 628639.
2001, Vol. 3 (2001) pp. 17631772. [53] M. Penrose, On k-connectivity for a geometric random graph, Random
[31] J. Wu and H. Li, A dominating-set-based routing scheme in ad Structures and Algorithms 15 (1999) 145164.
hoc wireless networks, Telecommunication Systems (Special Issue on [54] M. Penrose, A strong law for the longest edge of the minimal spanning
Wireless Networks) 3 (2001) 6384. tree, Annals of Probability 27 (1999) 246260.
[32] M. Joa-Ng and I.-T. Lu, A peer-to-peer zone-based two-level link state [55] D.M. Blough, M. Leoncini, G. Resta and P. Santi, On the symmetric
routing for mobile ad hoc networks, IEEE Journal on Selected Areas in range assignment problem in wireless ad hoc networks, in: Proc. of the
Communications 17(8) (1999) 14151425. 2nd IFIP Internat. Conf. on Theoretical Computer Science (2002).
[33] E. Royer and C. Toh, A review of current routing protocols for ad- [56] C. Cooper and A. Frieze, On the connectivity of random kth nearest
hoc mobile wireless networks, IEEE Personal Communications (April neighbour graphs, Combinatorics, Probability and Computing 4 (1995)
1999). 343362.
[34] S. Ramanathan and M. Steenstrup, A survey of routing techniques for [57] M. Grossglauser and D. Tse, Mobility increases the capacity of ad-hoc
mobile communication networks, Mobile Networks and Applications wireless networks, in: INFOCOMM, Vol. 3 (2001) pp. 13601369.
(1996) 89104. [58] P. Gupta and P. Kumar, Capacity of wireless networks, Technical Re-
[35] M. Mauve, J. Widmer and H. Harenstein, A survey on position-based port University of Illinois, Urbana-Champaign (1999).
routing in mobile ad hoc networks, IEEE Network Magazine 15(6) [59] O.D. Patrick, Connectivity in ad-hoc and hybrid networks, in: IEEE
(2001) 3039. INFOCOM (2002).
[36] P. Hall, On continuum percolation, The Annals of Probability 13(4) [60] P. Santi and D.M. Blough, An evaluation of connectivity in mobile
(1985). wireless ad hoc networks, in: Proc. of IEEE DSN (2002) pp. 8998.
[37] R. Meester and R. Roy, Continuum Percolation (Cambridge Univ. [61] F. Xue and P.R. Kumar, The number of neighbors needed for connec-
Press, Cambridge, 1996). tivity of wireless networks, Wireless Networks (2002) submitted.
[38] S. Basagni, I. Chlamtac, V.R. Syrotiuk and B.A. Woodward, A dis- [62] T. Lukovszki, New results of fault tolerant geometric spanners, in:
tance routing effect algorithm for mobility (DREAM), in: Proc. of Workshop on Algorithms and Data Structures (1999) pp. 193204.
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. Xiang-Yang Li has been an Assistant Professor of
[40] I. Stojmenovic, A routing strategy and quorum based location update Computer Science at the Illinois Institute of Tech-
scheme for ad hoc wireless networks, Technical Report TR-99-09, nology since 2000. He hold M.S. (2000) and Ph.D.
Computer Science, SITE, University of Ottawa (1999). (2001) degree in computer science from University
[41] I. Chlamtac and A. Farago, A new approach to design and analysis of of Illinois at Urbana-Champaign. He received his
peer to peer mobile networks, Wireless Networks 5 (1999) 149156. Bachelor degree in computer science and Bachelor
[42] P.-J. Wan, K.M. Alzoubi and O. Frieder, Distributed construction of degree in business management from Tsinghua Uni-
connected dominating set in wireless ad hoc networks, in: INFOCOM versity, P.R. China in 1995. His research interests
(2002). span the computational geometry, wireless ad hoc
[43] K.M. Alzoubi, P.-J. Wan and O. Frieder, New distributed algorithm networks, optical networks, and cryptography. He
for connected dominating set in wireless ad hoc networks, in: HICSS, is a member of the ACM and IEEE.
Hawaii (2002). E-mail: xli@cs.iit.edu
[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 Kousha Moaveninejad received the Bachelor in
Computing (MobiHoc02) (2002). computer soft engineering from Sharif University of
[45] K.N. Amouris, S. Papavassiliou and M. Li, A position based multi-zone Technology, Tehran, Iran, in 1997. He joined the
routing protocol for wide area mobile ad-hoc networks, in: Proc. of the Department of Computer Science of Illinois Insti-
49th IEEE Vehicular Technology Conf. (1999) pp. 13651369. tute of Technology in 2000, as a M.S. student and
[46] G. Calinescu, I. Mandoiu, P.-J. Wan and A. Zelikovsky, Selecting for- received the Masters degree in 2002. He then contin-
warding neighbors in wireless ad hoc networks, in: ACM DialM (2001). ued his study as a Ph.D. student at Illinois Institute of
[47] M. Penrose, The longest edge of the random minimal spanning tree, Technology and his current research interests include
Annals of Applied Probability 7 (1997) 340361. wireless ad-hoc networks, computational geometry,
[48] X.-Y. Li, Y. Wang, C.-W. Yi, P.-J. Wan and O. Frieder, Robust wireless algorithm design, and mobile computing.
ad hoc networks, in: IEEE ICC (2003) accepted for publication. E-mail: moavkoo@iit.edu
[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. Ophir Frieder is the IITRI Professor of Computer
[50] B. Bollobs, Random Graphs (Cambridge Univ. Press, Cambrige, Science at the Illinois Institute of Technology. His
2001). research interests span the general area of distributed
[51] H. Dette and N. Henze, Some peculiar boundary phenomena for ex- information systems. He is a member of ACM and a
tremes of rth nearest neighbor links, Statistics & Probability Letters 10 fellow of the IEEE.
(1990) 381390. E-mail: ophir@cs.iit.edu
Mobile Networks and Applications 10, 7988, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
To optimally resolve the competition between service cov- algorithms based on simulated annealing [1], tabu search
erage and financial cost, we introduce a multiple objective [19,20] and genetic algorithms [23] are far more popular.
optimization framework that does not require a priori knowl- Simulated annealing has been adopted for the APP in
edge of the relative importance of service coverage versus [2,3,6,26,33], and tabu search for the APP in [4,21,32,47].
cost. This is achieved by providing a range of alternative Both these techniques operate by ranking solutions using a
site selections which approximate the best possible trade-offs cost function. Given a solution, small changes are made to
(i.e., Pareto front) between cost and coverage. This means create a neighborhood of solutions from the current solution.
that unlike the current convention for cell planning, which The meta-heuristic then guides the acceptance of new solu-
generally seeks to generate a single cell plan given informa- tions available in the neighborhood. The advantage of this
tion on the relative importance of objectives, a radio engineer approach is that it has the ability to escape from local minima
will be able to choose from a range of alternative cell plans, in the search space (regarding the cost function), thereby im-
given visual and detailed information regarding each. This is proving performance. Differences in the application of these
particularly beneficial when there is a nonlinear relationship approaches involve how the cell planning problem is mod-
or unknown dependency between the objective functions as elled, the formulation of rules to create neighborhoods, and
for the general APP. the cost function used to rank solutions.
As far as we are aware [50], this method has only been con- Genetic algorithms have also become increasingly popu-
sidered for cell planning in [36], where a genetic algorithm lar for the APP [8,21,25,31,32,36,37,40]. With the excep-
was developed specifically for the APP; however, the total tion of [36], these approaches predominantly seek to opti-
network cost was not considered. Unlike [36], the framework mize a single function (or a linear combination of multiple
we propose considers financial cost and is flexible because objective functions) to create a population of high quality so-
it is possible to plug-in any multiple objective optimiza- lutions. These algorithms mimic evolution and natural se-
tion algorithm (MOA) which seeks to approximate a Pareto lection through fitness assignment, selection, recombination,
front. This flexibility is achieved by making the cell plan rep- and mutation on a population of solutions. For a genetic al-
resentation independent from the task of the MOA, which, in gorithm to succeed, a suitable representation of the problem
this case, is to find optimal orderings of candidate site loca- needs to be used. The most popular representation for the
tions which optimize the two objective functions. Exploring APP in the literature is a binary string with crossover opera-
a search space in this manner is common practice in many tions defined using geographic information between individ-
discrete optimization problems (e.g., the knapsack problem) ual base stations. However, as we show in this paper, genetic
and has also been successfully applied to the frequency as- algorithms need not be restricted to this representation. Our
signment problem [27]. study is unique in that it combines an integer string represen-
Using this approach, we explore the relative trade-offs be- tation with a multiple objective approach.
tween financial cost and service coverage using four state-of-
the-art genetic MOAs, which were selected as they produce 2.1. Resolving conflicting multiple objectives
many alternative solutions in parallel. The performance of
SEAMO, SPEA2, NSGA-II, and PESA is considered using Regardless which optimization technique is adopted to solve
a range of synthesized test problems. We argue that the ap- the APP, it is necessary to resolve the conflict between com-
proach we take is more likely to lead to an efficient opera- peting multiple objectives, such as service coverage and fi-
tional cellular network, as the two fundamental aspects are
nancial cost. The following definition is useful in this context.
considered in isolation from other cell planning tasks.
Definition 1 (Pareto optimality). Let o1 , o2 , . . . , on be ob-
2. Solving the APP jective functions which are to be maximized. Let S be the set
of all possible solutions. s S is dominated by t S (de-
The first published paper on optimizing antenna placement noted t s) if j , j {1, . . . , n}, such that oj (t) > oj (s)
dates back to 1994 [6]. Since then a large number of ap- and i, 1 i n, oi (t) oi (s). A non-dominated solution
proaches and scenarios have appeared in the literature. Al- is said to be Pareto optimal.
though exact approaches are only feasible for relatively small
test problems, they have been applied in a range of papers Pareto optimal cell plans are non-dominated in the sense
[3739,49]. However, in some papers, they have been re- that it is not possible to improve the value of any objective
laxed or selectively applied. Sequential, or greedy, algo- without simultaneously degrading the quality of one or more
rithms have been less well used and then predominantly for of the other objectives. The set of all possible Pareto optimal
comparison purposes (see [5,37]). Deterministic heuristic solutions in the entire search space is called the Pareto front.
algorithms have also been proposed by a range of authors In figure 1, a hypothetical Pareto front is indicated for the ob-
[9,17,18,28,34,37,38,44,45,53]. Frequently, these approaches jectives of cost and coverage. The most desirable cell plan in
exploit observations (e.g., density of base station locations) the Pareto front depends on which objective is most impor-
about the APP and incorporate them to enhance the perfor- tant. However, in the absence of such a relative ranking of
mance of the cell plans obtained. However, meta-heuristic objectives, solutions from the Pareto front must be regarded
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS 81
each base station to each STP. As well as Hatas model, many At iteration j (1 j n), (j ) is considered for addi-
other propagation models would have been equally suitable. tion to the set L .
Handover between c(j ) and L is feasible if the han-
2.2.2. Handover regions and objectives
dover percentage for c(j ) is less than the maximum
The subset of service test points covered by a particular an-
permitted.
tenna A is the cell served by A, denoted cA . Note that cells
served by different antennae are not necessarily disjoint since The largest power setting, denoted pmax , is identified
an STP can potentially be covered by more than one antenna. from the list p0 , p1 , p2 , . . . , pk such that handover is
Such an STP is referred to as a handover STP. A handover feasible between c(j ) and L .
STP which is contained in more than two distinct cells is re- If pmax = p0 , then (j ) is added to L , and the trans-
ferred to as a soft handover STP. For a cell cA , the subset of mission power of (j ) is recorded as pmax . Otherwise
handover STPs is denoted hA . For cA , the handover percent- (j ) is not added to L .
age is defined as
A number of observations can be made regarding this ap-
|hA | proach. Firstly, the approach is greedy in the sense that once
100.
|cA | 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
Controlling the size and distribution of handover regions is
transmission power be adjusted. Secondly, for a particular list
crucial for both operational and financial reasons. Handover
of potential site locations, characteristics (e.g., cost and cov-
regions are a prerequisite for seamless call transfer between
erage) of the resultant cell plan L is entirely dependent on
cells for mobile users. However, if very large handover re-
the order (i.e., permutation ) in which the base stations are
gions are permitted, there is a greater potential for inter-
considered for selection. It is our aim to find the best permu-
ference due to strong signals being received from multiple
tations, which lead to Pareto optimal cell plans, using genetic
sources. In frequency division multiple access systems, large
algorithms.
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 3. Genetic algorithms
cost of the network by increasing the total number of base
stations required to cover a given area. Only over the last decade have genetic algorithms (GAs) been
The objectives we are concerned with relate to financial successfully adapted to solve multiple objective problems. An
cost and area coverage of service test points. The cost of a excellent overview of this area is given in [12]. The general
cell plan L , denoted costL , is defined as: principle is to breed a new population of solutions (i.e., cell
plans) through a process of selection and recombination. This
costL = $(Li ).
occurs over a number of generations to try to improve the per-
Li L
formance of the population, as shown in figure 1. The expec-
The coverage of a cell plan is expressed as the proportion of tation is that desirable characteristics in solutions from one
STPs which are covered. Handover is not considered as an generation will combine to produce better solutions for the
objective in our problem formulation, but imposed as a con- next generation. Introduced by Holland [23], GAs are sup-
straint controlled by via the decoder. ported by theory which identifies the conditions under which
solutions converge to a high performing set of solutions. GAs
2.3. Cell plan representation which approximate the Pareto front seek to find, ideally, a di-
verse set of solutions spread evenly over the entire range of
The potential base station location Li is referred to as the ith the Pareto optimal front. In this study, we consider the abil-
base station. We use a permutation of the potential base ity of four state-of-the-art GAs, namely, SPEA2, NSGA-II,
station locations to represent cell plans. Each permutation PESA, and SEAMO to perform this function. Each algorithm
orders the potential base station locations. Under the permu- is briefly described below.
tation , the ith base station location listed is denoted (i).
We introduce a decoder which translates a permutation into 3.1. Brief description of each GA
a cell plan. This approach mimics the way in which the prob-
lem might be attempted manually. The decoder is effectively The Strength Pareto Evolutionary Algorithm version II
a greedy, sequential algorithm for creating a cell plan, which (SPEA2) is an enhancement of that originally proposed in
is dependent on the order of inspection for commissioning po- [54], and is described in detail in [55]. SPEA2 has been used
tential sites occurring in . The decoder adds cells iteratively in numerous studies (e.g., [7]) where good performance, in
to create a cell plan L as follows: comparison to other MOAs, has been reported. In SPEA2,
the most fit individuals from the union of archive and child
Initially L = . populations are determined by computing a fitness value for
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. raw fitness value based on how many solutions it dominates,
COMPARISON AND EVALUATION OF MULTIPLE OBJECTIVE GENETIC ALGORITHMS 83
and the second is a density estimate based on the its proximity 3.2. Recombination and mutation
to other solutions in objective space. At each generation, the
most fit n solutions are saved to the archive, and genetic oper- Each of the algorithms considered has a specific method for
ators are applied to form a new child population. This process selecting parents. SPEA2 bases selection on fitness, NSGA-
is repeated until termination. II on the crowded comparison operator, PESA on non-
NSGA-II is a fast elitist non-dominated sorting genetic al- dominated members of its archive set, and SEAMO uni-
gorithm (see [13] for a full description), which has been well formly. However, common recombination and mutation op-
studied (e.g., [14,29]). NSGA-II is similar to SPEA2, but uses erators have been used to maintain a fair comparison between
slightly different mechanisms. For example, in NSGA-II the the algorithms. The well-known cycle crossover [43] has been
most fit individuals from the union of archive and child popu- used as the recombination operator and the mutation operator
lations are determined by a ranking mechanism (or crowded involves the simple transposition of candidate base station lo-
comparison operator) composed of two parts. The first part cations in a randomly selected pair of positions. This was
peels away layers of non-dominated fronts, and ranks solu- governed by a mutation rate (set to 1%) to restrict the fre-
tions in earlier fronts as better. The second part computes a quency of mutation.
dispersion measure, the crowding distance, to determine how
close a solutions nearest neighbors are, with larger distances 3.3. Measuring the relative performance of GAs
being better. At each generation, the best n solutions with
regard to these two measures are saved to the archive, and ge- Comparing the performance of multiple objective algorithms
is problematic because a set of solutions rather than a single
netic operators applied to form a new child population. This
solution is obtained. Although several alternatives have been
process is repeated until termination.
proposed (e.g., [16,48,56,57] ) no single approach seems most
The Pareto Envelope-based Selection Algorithm, PESA, is
prevalent. We adopt the approach first given in [57] to calcu-
described in [10]. It uses different mechanisms than SPEA2
late a set coverage metric. This involves the concept of weak
and NSGA-II. The main differences are that its archive pop-
domination. Solution A weakly dominates solution B if A
ulation is not of fixed size and only allows non-dominated
and B have the same performance across all objectives or A
solutions to be members, which is a more limited set than
dominates B. For two sets of solutions SA and SB , the set
the previous two GAs allowed. If the archive ever exceeds
coverage metric of set SA with respect to SB is the percent-
n solutions, a squeeze factor is calculated for all members of
age of solutions in SB which are weakly dominated by at least
the archive. The squeeze factor is the total number of mem-
one solution from SA . The higher the set coverage metric ob-
bers in the same subregion of a hyper-grid (which partitions
tained, the greater the superiority of SA over SB .
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 4. Results
factor are then removed until the size of the archive is reduced
to n. Genetic operators are then applied to archive members The performance of the algorithms have been compared using
to form a new child population. This process is repeated until a wide range of synthesized test problems, each of which has
termination. been randomly generated. Each test problem gives the loca-
Finally, the Simple Evolutionary Algorithm for Multi- tion and cost of candidate sites. Test problems are classified
objective Optimization, known as SEAMO, has performed in two ways: the size of area in which they are positioned and
particularly well on the benchmark test knapsack optimiza- the density of candidate sites, as documented in figure 2.
tion problem [46]. The main difference between SEAMO and Combining the size of regions and the density of sites leads
the other algorithms, is that it is steady-state and has only one to a total of nine test problem classes, as indicated in figure 2.
population (of constant size n) to maintain. The main advan- For each test problem class, we produce five incidences on
tage of SEAMO is the simple approach it uses to dispose of which each algorithm is tested. This means that average algo-
all selection mechanisms based on fitness or rank. Instead, rithm performance is estimated and compared using five prob-
the search progresses based on three simple rules: lem instances from each of nine classes, with four different al-
gorithms, leading to a total of 180 experiments. All problem
(1) Parents can only be replaced by their own offspring. instances are available at: http://www.cs.cf.ac.uk/
user/L.Raisanen/downloads.html.
(2) Duplicates in the population are deleted.
Region size km2 Density of sites per km2
(3) Offspring can only replace parents if superior elitism. 0.03 0.06 0.12
15 15 7 14 28
30 30 27 54 108
Genetic operators are applied to each parent in turn to form
45 45 61 122 244
a new child, which is considered for substitution into the par-
ent population based on the three rules. This process is re- Figure 2. Number of candidate sites in nine problem classes defined by region
peated until termination. 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.
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 Watts erage and cost, closely followed by SPEA2 (89.67%), then
p1 30 1000 PESA (64.01%) and, finally, SEAMO (59.53%). See table 1
p2 27 501 for details.
p3 24 251 In figure 4 we plot the Pareto fronts (i.e., non-dominated
p4 21 125
solutions from the final population) achieved by each algo-
p5 18 63
rithm on the large region problem at each density. Despite
Figure 3. Power settings used in tests. the differences in relative algorithm performance, the Pareto
fronts obtained are closely clustered in real terms. Generally,
To maintain a fair comparison between algorithms, com- the plots show that that as candidate site density increases, so-
mon parameter settings have been adopted for each experi- lutions with a higher level of coverage are achievable. Also, in
ment. Five nonzero power settings (displayed in units of dBW the most dense problem instances, lower cost solutions with
and Watts) have been used, as specified in figure 3. Unless higher coverage are achievable due to more freedom in site
otherwise specified, a population size of 100 is adopted us- selection and cost.
ing 500 generations. Additionally, the same random starting In figure 5, we display an example of cell plans with the
populations have been used for each problem class. highest coverage for the large region problem at each density,
We consider the performance of each GA in four ways: with density increasing left to right.
(1) the average performance (in terms of the objective values 4.2. Measure of solution distribution
of members in the final population) compared to other
GAs across all test problems using the set coverage met- To measure the distribution, or spread, of solutions along the
ric, with diagrams to show obtained Pareto fronts and cell Pareto front, a metric proposed in [41] has been implemented.
plans, The spacing measure is based on the range of values for di ,
(2) the average measure of population spread, which is the distance (in terms of solution space) between the
ith element of the solution set and its nearest neighbour. The
(3) the speed of convergence to solutions in the final popula- average of di values for a solution set of size n is denoted d.
tion, and Then the measure of spread is defined as:
(4) the average speed of execution.
1 n
2
S= d di (1)
4.1. Average performance across test problems n1
i=1
In terms of the average performance of each GA compared for the n members in the final population. Note that S = 0
to other GAs across all test problems using the set coverage indicates all members of the Pareto front are spaced equidis-
metric, it was found that NSGA-II achieves the best perfor- tantly in the solution space.
mance by weakly dominating an average of 91.15% of solu- It was found that PESA performed the best on this mea-
tions obtained by other algorithms in terms of service cov- 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.
Figure 6. Average spacing values by algorithm for each test problem class.
SPEA2 at 20.91, NSGA-II at 21.46, and SEAMO at 26.84 tions, and intermediate populations (produced every 250 gen-
(see figure 6 for details). It is little surprise that the algorithm erations) have been compared against each other, using the
which is specifically designed to encourage spacing in objec- set coverage metric (see section 3.3). The results indicate
tive space, PESA, performed the best, and that the algorithm that PESA and SEAMO converge quickly, by generation 1000
with no direct measure to control dispersion, SEAMO, per-
dominating 88.88% and 77.77% of solutions in the final gen-
formed the worst.
eration (1500), respectively, whereas SPEA2 and NSGA-II
converge more slowly dominating 22.22% and 33.33% re-
4.3. Convergence
spectively (see figure 8 for details). However, considering
The ability of an algorithm to rapidly converge to the final that SPEA2 and NSGA-II outperform PESA and SEAMO in
solution set is desirable. This has been investigated for each terms of final solutions (see section 4.1), this may also indi-
algorithm using the largest problem class with the highest site cate that these algorithms are better at improving solutions
density. Each algorithm has been applied for 1500 genera- 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 87
[19] F. Glover, Heuristics for integer programming using surrogate con- IEEE Journal on Selected Areas in Communications 14(4) (1996) 662
straints, Decision Sciences 8 (1977) 156166. 673.
[20] F. Glover, E. Taillard and D. de Werra, A users guide to tabu search, [43] T. Starkweather, S. McDaniel, K. Mathias, D. Whitley and C. Whitley,
Annals of Operations Research 41 (1993) 328. A comparison of genetic sequencing operators, in: Proc. of the Fourth
[21] J.K. Han, B.S. Park, Y.S. Choi and H.K. Park, Genetic approach with a Internat. Conf. on Genetic Algorithms, eds. R. Belew and L. Booker
new representation base station placement in mobile communications, (Morgan Kaufman, San Mateo, CA, 1991) pp. 6976.
in: Proc. of the 54th IEEE Conf. on Vehicular Technology, Vol. 4 (2001) [44] D.-W. Tcha and Y.-S. Myung, Base station location in a cellular CDMA
pp. 27032707. system, Telecommunication Systems 14 (2000) 163173.
[22] M. Hata, Empirical formula for propogation loss in land mobile ra- [45] K. Tutschku, Interference minimization using automatic design of cel-
dio services, IEEE Transactions on Vehicular Technology 29(3) (1980) lular communication networks, in: Proc. of the IEEE VTC98 Conf.
317325. (1998) pp. 634638.
[23] J.H. Holland, ed., Adaptation in Natural and Artificial Systems (Uni- [46] C.L. Valenzuela, A simple evolutionary algorithm for multi-objective
versity of Michigan Press, Ann Arbor, 1975). optimisation(SEAMO), in: IEEE Congress on Evolutionary Computa-
[24] I. Howitt and S.-Y. Ham, Base station location optimization, in: Proc. tion (2002) pp. 717722.
of the IEEE VTC99 Conf., Vol. 4 (1999) pp. 20672071. [47] M. Vasquez and J.-K. Hao, A heuristic approach for antenna positioning
[25] X. Huang, U. Behr and W. Wiesbeck, Automatic cell planning for a in cellular networks, Journal of Heuristics 7 (2001) 443472.
low-cost and spectrum efficient wireless network, in: Proc. of Global [48] D.A. Veldhuizen and G.B. Lamont, On measuring multiobjective evo-
Telecommunications Conf. (GLOBECOM), Vol. 1 (2000) pp. 276 lutionary algorithm performance, in: Congress on Evolutionary Com-
282. putation (2000) pp. 204211.
[26] S. Hurley, Planning effective cellular mobile radio networks, IEEE [49] R.M. Whitaker and S. Hurley, Omni-directional cell planning, in:
Transactions on Vehicular Technology 51(2) (2002) 243253. Telecommunications Network Design and Management, eds. G. Anan-
[27] S. Hurley, D.H. Smith and C.L. Valenzuela, A permutation-based ge- dalingam and S. Raghavan (Kluwer Academic, Dordrecht, 2002) chap-
netic algorithm for minimum span frequency assignment, in: Lecture ter 2, pp. 2541.
Notes in Computer Science, Vol. 1498 (Springer, New York, 1998) [50] R.M. Whitaker and S. Hurley, The state-of-the-art in automatic cell
pp. 907916. planning, submitted for publication.
[28] L.J. Ibbetson and L.B. Lopes, An automatic base station placement al- [51] J.K.L. Wong, M.J. Neve and K.W. Sowerby, Optimisation strategy for
gorithm, in: Proc. of the IEEE VTC97 Conf., Phoenix, AZ (May 1997) wireless communications system planning using linear programming,
pp. 770774. IEE Electronics Letters 37(17) (2001) 10861087.
[29] V. Khare, X. Yao and K. Deb, Performance scaling of multi-objective [52] M.H. Wright, Optimization methods for base station placement in wire-
evolutionary algorithms, KanGAL Report No. 2002009 (2002) pp. 1 less systems, in: Proc. of the IEEE VTC98 Conf. (1998) pp. 387
15. 391.
[30] J.D. Knowles, Local-search and hybrid evolutionary algorithms for [53] J. Zimmermann, R. Hons and H. Muhlenbein, ENCON: Evolutionary
Pareto optimization, Ph.D. thesis, The University of Reading (2002). algorithm for the antenna placement problem, Computers and Industrial
[31] I. Laki, L. Farkas and L. Nagy, Cell planning in mobile communication Engineering 44 (2003) 209226.
systems using SGA optimization, in: Proc. of Internat. Conf. on Trends [54] E. Zitzler, Evolutionary algorithms for multiobjective optimization:
in Communications, Vol. 1 (2001) pp. 124127. Methods and applications, Ph.D. thesis, Swiss Federal Institute of Tech-
[32] C.Y. Lee and H.G. Kang, Cell planning with capacity expansion in mo- nology, Zurich, Switzerland, TIK-Schriftenreihe Nr. 30 (1999).
bile communications: A tabu search approach, IEEE Transactions on [55] E. Zitzler, M. Laumanns and L. Thiele, SPEA2: Improving the strength
Vehicular Technology 49(5) (2000) 16781691. Pareto evolutionary algorithm, Technical Report 103, Computer En-
[33] R.M. Mathar and T. Niessen, Optimum positioning of base stations for gineering and Networks Laboratory (TIK), ETH Zurich, Switzerland
cellular radio networks, Wireless Networks 6 (2000) 421428. (2001).
[34] R.M. Mathar and M. Schmeink, Optimal base station positioning and [56] E. Zitzler, M. Laumanns, L. Thiele, C. Fonseca and G. da Fonseca, Per-
channel assignment for 3G mobile networks by integer programming, formance assessment of multiobjective optimizers: An analysis and re-
Annals of Operations Research 107 (2001) 225236. view, Technical Report 139, Computer Engineering and Networks Lab-
[35] R.M. Mathar and M. Schmeink, Integrated optimal cell site selection oratory (TIK), ETH Zurich, Switzerland (2002).
and frequency allocation for cellular radio networks, Telecommunica- [57] E. Zitzler and L. Thiele, Multiobjective optimization using evolutionary
tion Systems 21 (2002) 339347. algorithms a comparative case study, in: Parallel Problem Solving
[36] H. Meunier, E. Talbi and P. Reininger, A multiobjective genetic algo- from Nature (1998) pp. 292301.
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 Larry Raisanen is a second year Ph.D. student
spread for cellular network planning using the combination algorithm studying at Cardiff University at the Centre for Mo-
for total optimisation, in: Proc. of the 1st Internat. Conf. on 3G Mobile bile Communications based in the School of Com-
Communication Technologies (2000) pp. 171175. puter Science. The focus of his work is on the devel-
[40] P. Reininger, S. Iksal, A. Caminada and J.J. Korczak, Multi-stage op- opment, testing, and analysis of algorithms to resolve
timization for mobile radio network planning, in: Proc. of the IEEE the base station placement problem in wireless mo-
VTC99 Conf., Vol. 3 (1999) pp. 20342038. bile communication systems using a multi-objective
[41] J. Schott, Fault tolerant design using simple multicriteria genetic algo- optimization framework. He graduated in 1993 with
rithms, Ph.D. thesis, M.S. thesis, Department of Aeronautics and As- a BA magna cum laude from Marquette Univer-
tronautics, MIT, Cambridge, MA (1995). sity (WI, USA), and in 2003 with a M.Sc. in com-
[42] H.D. Sherali, C.H. Pendyala and T.S. Rappaport, Optimal location puting distinction from Cardiff University (Wales, UK).
of transmitters for micro-cellular radio communication system design, 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.
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
the value of the channel separation among adjacent stations induce a clique in the augmented graph of the graph. We will
as witnessed by optimal colourings. 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
2. Preliminaries this paper follow the pattern: for a given graph G, and for a
Formally, the Channel Assignment Problem with Separation given ,
(CAPS) can be modelled as an appropriate colouring prob- (1) identify T (G, ),
lem on an undirected graph G = (V , E) representing the net-
(2) find the number of vertices in T (G, ), and hence a lower
work topology, whose vertices in V correspond to stations,
and edges in E correspond to pairs of stations that can hear bound for the given assignment problem,
each others transmission [2]. The colour assigned to a par- (3) describe a colouring scheme to colour all the vertices of
ticular vertex corresponds to the frequency channel assigned T (G, ),
to the corresponding station. For a graph G, we will denote
(4) demonstrate a tiling of the entire graph made up of
the distance between any two vertices in the graph, i.e., the
T (G, ) to show that the colouring scheme described
number of edges in a shortest path between the two vertices,
colours the entire graph, and
by dG (, ). (When the context is clear, we will denote the
distance as simply d(, ).) CAPS is then defined as: (5) show that the colouring scheme satisfies the given reuse
distance and channel separation constraints.
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 3. A characterisation of optimal colourings
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 . We first introduce the conventions we follow to represent
square grids and cellular grids. We explain tilings in both
This assignment is referred to as a g-L(1 , 2 , . . . , grids, and define some notation. Then we present our charac-
1 ) colouring of the graph G [7], and CAPS (G, , ) is terisation of optimal colourings in cellular and square grids.
sometimes referred to as the L() colouring problem for G. For any d-dimensional lattice L, the minimal distance in
Note that a g-L(1 , 2 , . . . , 1 ) uses only the (g + 1) the lattice is denoted by (L). The infinite graph, denoted
colours in the set {0, . . . , g}, but does not necessarily use all G(L), corresponding to the lattice L consists of the set of lat-
the (g + 1) colours. A g-L(1 , 2 , . . . , 1 ) colouring of G tice points as vertices; each pair of lattice points that are at a
is optimal iff g is the smallest number witnessing a solution distance (L) constitute the edges of G(L).
for CAPS (G, , ). The lattice Zd is the set of ordered d-tuples of integers, and
Finding the optimal colouring for general graphs has Ad is the hyperplane that is a subset of Zd+1 , and is charac-
been shown to be NP-complete. The problem remains terised as the set of points in Zd+1 such that the coordinates
NP-complete even if the input graphs are restricted to planar of each point add up to zero. (Zd ) = 1, and the minimal
graphs, bipartite graphs, chordal graphs, and split graphs [4]. length vectors in Zd are the unit vectors in each dimension.
Most of the work on this problem has dealt with specific For each d > 0, for each i, j, 0 i, j d, i = j , define
graphs such as grids and rings, for small reuse distance ( ) dij = (x0 , . . . , xd ), where xi = 1, xj = 1, and for each
values, and for small channel separation (i ) values, e.g., op- k, 0 k d, k = i, j, xk = 0. Then, (Ad ) = 2, and
timal L(1, 1) colourings for rings and bidimensional grids [1], the set of minimal length vectors in Ad is {dij | i, j, 0 i,
optimal L(2, 1) and L(2, 1, 1) colourings for hexagonal, bidi- j d, i = j }. (See [5,10] for more on these lattices.)
mensional, and cellular grids [2], etc. Recently, Bertossi The infinite 2-dimensional square grid is, then, G(Z2 ), and
et al. [3] exhibited optimal L(1 , 1, . . . , 1) colourings, for the infinite 2-dimensional cellular grid is G(A2 ).
1 /2, for bidimensional grids and rings. (See [3] for a
succinct literature survey of this problem.) Below, we refer to 3.1. Cellular grids
L(, 1, . . . , 1) colourings by L(, 1k ) colourings.
As pointed out in [2], a lower bound for the L(1, 1 k ) For a given value of , two kinds of tiles can be identified:
colouring problem is also a lower bound for the L(, 1 k ), triangular and hexagonal. The tiles are shown in figure 1(a).
> 1. Given an instance of CAPS, consider the augmented It can be easily shown that:
graph obtained from G by adding edges between all those
pairs of vertices that are at a distance of at most 1. Clearly, Lemma 1. 1. The number of vertices in a triangular tile cor-
then, the size (number of vertices) of any clique in this aug- responding to reuse distance , denoted by cT ( ), is given by
mented graph places a lower bound on an L(1, 1 1 ) colour- cT ( ) = ( + 1)/2.
ing for G; the best such lower bound is given by the size of a 2. The number of vertices in a hexagonal tile corre-
maximum clique in the augmented graph. sponding to reuse distance , denoted by cH ( ), is given by
In each graph G, for each , we identify a canonical sub- cH ( ) = (3 2 +( mod 2))/4. Thus, when = 2k +1(odd),
graph, T (G, ), of the graph so that the vertices of T (G, ) c( ) = 3k 2 + 3k + 1.
A CHARACTERISATION OF OPTIMAL CHANNEL ASSIGNMENTS 91
= 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
(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 93
Figure 8. Cellular grid bounding box for = 4 with position numbers. Figure 11. Square grid bounding box for = 5 with position numbers.
A colouring scheme is optimal if it uses the smallest possible Proof. Consider the edge d0 of the bounding box around
number of colours. In other words, a colouring which uses p, B(p). Consider the k different tiles, each of whose edge
colours from the set {0, 1, . . . , g} will be optimal if it uses the t0 is a part of the edge d0 of B(p), where = 2k + 1 for odd
smallest possible value for g. From lemmas 1 and 2, we know and = 2k for even . Refer to figures 811.
that c( ) is a lower bound on the number of colours used. We Let P (i) denote the sequence of position numbers on d0 of
are concerned only with such colouring schemes which use B(p) that are on the edge t0 of the ith of these k tiles. In the
exactly c( ) different colours. case of odd , P (i) are given by:
We already know that is the minimum distance at which
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) Theorem 2. Given , and given a tiling of a cellular or
..
. square grid by tiles (for ), a colouring with reuse distance
is optimal iff all the tiles in the tiling are identical in their
P (k) = 2 + (k 2), 2 + (k 1), . . . , 2k 1 .
colour assignment.
Since the colouring is optimal, the colour c, that the Recall the definition of tilings A and B from section 3.2
point p is coloured in, must appear somewhere on each of (see figures 4 and 6). From the proof of theorem 1, we make
these tiles. Except for the edge t0 , each of these tiles is com- the following observation.
pletely contained within B(p). Thus, the colour c must appear
on the edge t0 of each of these tiles, otherwise, the reuse con- Corollary 1. Suppose = 2k + 1, and we have an optimal
straint is violated. Since no pair of vertices with position num- colouring of the cellular (square) grid. If for any point p in
bers 1, 2, . . . , 2k on the edge d0 of B(p) are at a distance , the grid, the vertex corresponding to position number k on an
it must be the case that the colour c is assigned to some vertex edge of the bounding box of p has the same colour as that as-
that is common to all the above tiles. In case of odd , as seen signed to p, then the tiling of the grid, by identically coloured
from equation (1), the only two common vertices are the ones tiles, corresponds to tiling B; if the position number is k + 1,
with position numbers k and k + 1, and hence, one of these then the tiling of the grid corresponds to tiling A.
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 4. Optimal L(1 , 1 2 ) colourings for G(A2 ) and G(Z2 )
be assigned the colour c. A similar argument establishes that
on each edge of B(p), the vertices corresponding to position In this section, we deal with optimal frequency assignment
numbers k and k + 1 in case of odd and k in case of even schemes for wireless networks modelled as cellular grids
are the only possible candidates for being assigned colour c. and square grids. We first present an L(1 , 1 2 ) colouring
Now, in case of odd , let q be the vertex, corresponding scheme of G(A2 ) for the case where reuse distance is odd
i.e., = 2k + 1, k {1, 2, . . .}. This is followed by an
to position number k on the edge d0 of B(p), that is assigned
L(1 , 1 2 ) colouring scheme of G(Z2 ) for all values of .
colour c (see figures 9 and 11). Suppose, by way of contra-
The colouring schemes presented here correspond to tiling A.
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
4.1. Cellular grids
will now consider cellular and square grids separately in two
different cases. We present a colouring scheme where 1 varies as the square
Case 1 (Cellular grids). Consider the bounding box B(q). of , for 5, odd. We note that the colouring of the
The edge d2 of B(q) passes through the vertex with position entire cellular grid is achieved by colouring one tile and re-
number k on the edge d1 of B(p) as shown in figure 9. By producing the same colouring in all the tiles present in the
the above argument, one of the two vertices with position grid. Recall that the number of vertices c( ), in a tile cor-
numbers k or k + 1 on this edge d2 of B(q) must be as- responding to an odd reuse distance = 2k + 1 is equal to
signed colour c. But both these vertices are at a distance less 3k 2 + 3k + 1 (see lemma 1). From the above fact and from
than from the vertex x. Therefore, x cannot be assigned theorem 2, we make the following observations (refer to fig-
colour c, implying that the vertex with position number k on ure 12).
the edge d1 must be assigned colour c.
Case 2 (Square grids). Consider the bounding boxes B(q) Lemma 4.
and B(x). Let r be the point of inter-section of the edges d1
1. Colouring c( ) points starting from the vertex of a tile
of B(q) and d0 of B(x) (see figure 11). If both q and x are
along the direction j is equivalent to colouring all the di-
coloured c, it follows that r should be assigned the colour c.
agonals of a tile in the following order:
This is not possible because r lies within the bounding box
B(p) of point p which is also coloured c. This implies that L0 , Lk+1 , L1 , Lk+2 , . . . , Lk1 , L2k , Lk .
the vertex with position number k on the edge d1 must be
2. Along a line i = m, where m is a constant, any pair of
assigned colour c.
points which are at a distance c( ) apart will have the same
Similar arguments in both cases above establish that, if the
colour assigned to them.
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 3. Consider a point (p, q) on the line i = p. The point
B(p), the vertex with position number k is also coloured the (p + 1, q 3k 1) on the line i = p + 1 will have the
same as p. same colour as (p, q).
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)
(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)
(1) Find optimal colouring schemes for cellular grids with M.V.S. Shashanka obtained a B.E. (Hons.) from the
high 1 values for the case when is even. 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
(2) Find and prove the existence of tight upper bounds for University, Boston, MA, USA.
1 , 1 2, . . . for a general . 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.
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
however, may be integrated easily with zone routing proto- savings for high query rates in large-scale networks; a drastic
cols to compose a general routing solution. improvement in performance.
Nodes in ad hoc networks are usually portable devices with The rest of this document is organized as follows. Sec-
limited battery power. Therefore to save power the resource tion 2 discusses related work. Section 3 describes our design
discovery mechanism should be efficient in terms of commu- goals and provides an overview of our architecture, CARD,
nication overhead. Our architecture is designed to meet re- and introduces the contact selection, maintenance and query
quirements for power-efficient resource discovery and small algorithms. Section 4 presents analysis of CARD, and com-
transfers in large-scale ad hoc networks with (potentially) pares it to flooding, smart flooding and bordercasting. We
thousands of wireless devices. Scalability is one of our main conclude in section 5.
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 2. Related work
in highly clustered networks results in significant reduction
in the average path length (or degrees of separation) to ap- Related research lies in the areas of routing and resource dis-
proach that of random networks. In our architecture we adopt covery in ad hoc networks. Due to lack of infrastructure in ad
a hybrid approach in which a node uses periodic updates to hoc networks, resource (and route) discovery is a challeng-
reach its vicinity within a limited number of hops, R, and re- ing problem. Most of the routing protocols can be broadly
active querying beyond the vicinity via contacts. Contacts classified as: proactive (table-driven), reactive (on-demand),
act as short cuts that attempt to transform the network into a hybrid, or hierarchical.
small world by reducing the degrees of separation between the Proactive schemes such as DSDV [21], WRP [18] and
source and destination of the transfer. They help in providing GSR [2] flood periodic updates throughout the network. This
a view of the network beyond the vicinity during resource dis- is resource consuming, especially for large-scale networks.
covery. Each node maintains state for a few contacts beyond Reactive schemes such as AODV [22] and DSR [13] attempt
its vicinity. Contacts are polled periodically to validate their to reduce the overhead due to periodic updates by maintaining
presence and routes. For discovering resources efficiently, state only for the active resources and using route caching. In
queries are sent to the contacts that leverage the knowledge these schemes a search is initiated for new discovery requests.
of their vicinity. As the number of contacts increases, the net- However, the search procedure generally involves flooding (or
work view (or reachability) increases. However, at the same expanding ring search), which also incurs significant over-
time the overhead involved in contact selection and mainte- head. Furthermore, the performance of on-demand routing
nance also increases. Our results show this trade-off. We with caching has been shown to degrade significantly with
introduce and study alternative mechanisms for contact selec- small transfers in large-scale mobile networks.
tion and identify a novel scheme (called the edge method for Hybrid schemes such as the zone routing protocol (ZRP)
contact selection) that is able to achieve a balanced trade-off [5,20] try to combine the benefits of both the proactive and re-
and good performance in terms of increased reachability and active schemes. ZRP limits the overhead of periodic updates
reduced overhead. to a limited number of hops (called the zone radius). Re-
Once the contacts are selected by a node they are used in sources beyond the zone are discovered in a reactive manner
the resolution of resource discovery queries. Only the contact by sending queries through nodes at the edges of the zones
nodes are queried without the need for flooding, resulting in (bordercasting). The zone concept is similar to the vicin-
drastic reduction in per-query communication overhead. The ity concept in our study. However, instead of bordercasting
total overhead, however, is the resultant of we use contact queries. The design principles upon which
our CARD architecture was designed employing contacts
(i) the query overhead, which is a function of the per-query
as short cuts to create a small world, and trading off optimal
overhead and the query rate,
paths for energy efficiency are fundamentally different from
(ii) the vicinity establishment and maintenance overhead, those used for ZRP bordercast. In our study, through detailed
which is a function of the node mobility, and comparison we show that the contact-based approach is much
more efficient than bordercasting for our purposes. Further-
(iii) the contact selection and maintenance overhead.
more, CARD maybe easily integrated with ZRP to provide a
Our study elaborates on the interplay between these var- complete routing protocol in which ZRP is used to discover
ious overhead components, the query rate, and the mobility routes for long-lived flows and CARD is used for resource
rate using the call-to-mobility-ratio (CMR) metric. discovery and small transfers.
Extensive simulation-based comparisons with flooding Hierarchical schemes, such as CGSR [3,15], tend to have
and bordercasting [5,20] show our architecture to be more ef- good scalability, but involve election of cluster-heads, that
ficient, especially for high query rates. Simulation results also have greater responsibilities than other nodes. A cluster-head
show that our protocol is scalable and can be configured to is responsible for routing traffic in and out of the cluster.
provide good performance for various network sizes. Over- Cluster-based hierarchies rely on complex coordination and
head savings are function of the query rate, reaching 93% thus are susceptible to major re-configuration due to mobility
(vs. flooding) and 80% (vs. bordercasting) in communication and node failure, leading to serious performance degradation
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 101
in highly dynamic networks. Also, a cluster head may be a 3. CARD architectural overview
single point of failure and a potential bottleneck. In our ar-
chitecture each node has its own view of the network, and In this section we provide an overview of the CARD archi-
hence there is very little coordination between various nodes. tecture. In particular, we describe the design requirements for
This enables our architecture to adapt gracefully to network our architecture, present definitions and terminology used in
dynamics. GLS [14] provides a location-discovery service for this document, and introduce and investigate alternative con-
tact selection, maintenance and query mechanisms.
geographic routing. GLS requires nodes to know of a network
grid map and assumes knowledge of node locations (via GPS 3.1. Design requirements
or other). CARD does not require location information.
Related work on smart or efficient flooding has been pro- The design requirements of our CARD resource discovery ar-
posed in [4,6,16,19]. These techniques attempt to reduce the chitecture for large-scale Ad hoc networks include (a) scala-
redundancy inherent in flooding, and may be integrated in our bility, (b) power-efficiency, (c) robustness, (d) decentralized
work to provide more efficient vicinity establishment instead self-organization, and (e) independence of location informa-
of regular link state protocol. One major difference between tion.
smart flooding and CARD is that smart flooding reduces the (a) Scalability. Applications of large-scale ad hoc networks
redundant messages in querying every node in the network, include military and sensor network environments that may
whereas CARD attempts to create a small world and only include thousands of nodes. Therefore the resource discovery
queries a small number of nodes on the order of the degrees of mechanism should be scalable in terms of control overhead
separation from source to target. In relatively sparse networks with increase in network size. We shall show that CARD may
(some of which we include in our study) smart flooding will be configured to perform very well over a wide array of net-
not be very effective since there is no significant redundancy work sizes and conditions.
in flooding anyway. Section 4.3 discusses this issue further.
(b) Power and communication efficiency. Ad hoc networks
In [8] we have shown the relationship between small include portable devices with limited battery power. There-
worlds and wireless networks. In this paper, we build fore, resource discovery mechanisms should be power-effi-
upon that relationship by introducing the contacts to act as cient. CARD achieves dramatic reduction in communica-
short cuts in the highly clustered multi-hop wireless network, tion overhead (in terms of transmitted and received messages)
proposing and evaluating in details two proactive contact- over the several existing schemes considered in our study.
selection mechanisms. We first introduced the high level idea
of using contacts in [7]. The initial work on the CARD archi- (c) Robustness. The mechanism should be robust in the face
tecture was presented in [11]. This work extends the analysis of network dynamics. A periodic soft-state mechanism is pro-
vided to handle node failures and frequent link failures due to
of the CARD architecture and explores the important inter-
mobility.
play between the query rate and mobility rate. The MARQ
architecture [9] provides a mobility-assisted contact selec- (d) Decentralized operation. For the network to be rapidly
tion mechanism, the efficiency of which increases with mo- deployable, it should not require any centralized control.
bility. In cases of static networks (e.g., sensor networks), or CARD does not require or assume any centralized entity or
when mobility is low, CARD may be used in conjunction with special infrastructure.
MARQ for efficient query resolution. TRANSFER [10] pro-
(e) Independence of location information. GPS (or other lo-
vides a reactive (on-the-fly) contact selection mechanism to cation information) may not be available in many context
reduce node-contact vicinity overlaps, but does not explicitly (e.g., indoors, or in simple devices and sensors). Hence, as-
reduce the contact-contact vicinity overlap because contacts suming availability of location information limits the applica-
are selected in parallel. CARD, by virtue of selecting contacts bility of the proposed scheme. We avoid such limitation in our
proactively and using the edge method for contact selection design and do not assume or require any location information.
in serial is able to guarantee non-overlapping node-contact
vicinities and reduce the contact-contact vicinity overlap, but 3.2. Definitions
may incur more overhead for periodic contact maintenance.
An overview of the CARD architecture is shown in figure 1.
ACQUIRE [23,24] is an architecture for multi-variable query
Following are some terminology definitions we use through-
resolution in sensor networks, that uses the look-ahead tech-
out this document.
nique to optimize overhead. The query is forwarded from one
querying node to another d hops away, randomly. The work Vicinity (of a node). All nodes within a particular number
provides an analytical framework to get optimal d for given of hops (R) from the node. R is the radius of the vicinity.
level of network and event dynamics. A variant of CARDs Edge nodes (of a nodes vicinity). All nodes at a distance
contact selection may be used to reduce the overlap between of exactly R hops away from the node.
the look ahead zones for successive querying nodes to im- Maximum contact distance (r). The maximum distance (in
prove the performance of ACQUIRE. hops) from the source within which a contact is selected.
102 HELMY ET AL.
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 Each node proactively (e.g., using a link state protocol) main-
common nodes between their vicinities. tains state for resources within its vicinity. Alternatively, a
Number of Contacts (NoC). NoC specifies the value of smart flooding scheme (e.g., based on dominating sets) may
the maximum number of contacts to be selected by each be used to reduce the vicinity establishment and maintenance
source node. The actual number of contacts chosen is usu- overhead. For comparison reasons, however, in this study
ally less than this value. This is due to the fact that for we shall use a link state protocol similar to that used in ZRP
a particular value of R and r, there is only a limited re- to maintain vicinity information. The overhead of such link
gion available for choosing contacts. Once this region has state protocol increases with node mobility and the number of
been covered by vicinities of the chosen contacts, choos- nodes within the vicinity. Such overhead under mobility sce-
ing more contacts in the same region is not possible, as narios will be thoroughly studied later in section 4.3, and will
their vicinities would overlap with the vicinities of the al- be factored into the overall overhead of CARD. As we shall
ready chosen contacts. This is according to our contact see, when the vicinity overhead is amortized over a reasonable
selection policy to minimize overlap. number of queries the overall gain is still quite significant.
Each node also maintains state for (a few) nodes that lie
Depth of search (D). D specifies the levels of contacts
outside the vicinity. These nodes serve as contacts for ac-
(i.e., contacts of contacts) queried by a source.
cessing resources beyond the vicinity. Contacts are selected,
Reachability. The reachability of a source node refers to maintained and queried using the mechanisms described be-
the number of nodes that can be reached by the source low.
node. This includes the nodes within the vicinity that can
be reached directly and the nodes that lie in the contacts 3.4. Contact selection, maintenance and query mechanisms
vicinities, and their contacts vicinities, and so on, up to D
levels of contacts. This is also considered a measure of the Contacts are key to the efficient resolution of resource discov-
discovery success rate. ery queries. In CARD, the contacts are selected proactively in
anticipation of queries, and paths to these contacts are main-
3.3. Establishing and maintaining vicinity information tained using a periodic soft state mechanism to capture net-
work dynamics and mobility effects. Since the contacts are
Our architecture employs a hybrid of proactive and reactive selected proactively, the contact selection delays become less
approaches for resource discovery. As shown in figure 1, all of a concern (than they would otherwise in a reactive, on-the-
nodes within R hops from a node form the nodes vicinity. fly, scheme, for example). Hence, contacts are selected in
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 103
(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
Figure 5. Overhead for (1) PM and (2) EM.1
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
this scheme guarantees non-overlap between the node and the
out of the vicinity (i.e., d > R). Therefore, it is possible
contacts vicinities, as follows. Any node that lies at a dis-
that a contact may be selected at a location where the CS has
tance of R hops or less from the edge will have an overlap-
traversed more than 2R hops, but the contact may in fact be
ping vicinity with the ss vicinity, and hence will have at least
closer than 2R hops from the source, as shown in figure 2(a)
one of ss edge nodes in its vicinity. Thus, checking for non-
route 2, leading to heavy overlap.
overlap with the edge nodes ensures that a contact is chosen at
More seriously, the probabilistic method for contact selec-
least 2R +1 hops away from s. This eliminates the possibility
tion can be expensive in terms of the amount of traffic gener-
of an overlap due to the lack of direction. The Edge_List may
ated by the CS. This is due to the extra traffic generated due
be added to the CS in a communication-efficient manner by
to backtracking, and lost opportunities when the probability
using bloom filters [17] to represent membership in the edge-
fails, even when there is no overlap. To reduce the possibil-
ity of such a situation, the probability equations (1) and (2) list. Figures 4 and 5 show a comparison of the probabilistic
are not used. The probability equations were formulated to and edge methods. As can be seen from figure 4 the reach-
have a higher possibility of choosing the contact that lies ei- ability saturates in both PM and EM. However the saturation
ther between R and r hops (equation (1)) or between 2R and occurs much earlier in the case of probabilistic method. Also
r hops (equation (2)). To maintain this non-overlapping prop- as compared to EM, the reachability achieved is less for PM,
erty without the probability equations, the contact selection for the same values of NoC. Figure 5 shows the backtrack-
procedure is modified as follows. ing overhead for PM and EM. Due to the reasons explained
The list of all edge nodes (Edge_List) of s is added to earlier, overhead is significantly reduced for EM.
the CS. Also, the query and source IDs are included to pre-
vent looping. Note that the Edge_List is readily available 3.4.3. Contact maintenance mechanism
through the vicinity information and obtaining it does not re- Node mobility may cause the path to a contact to change.
quire any extra overhead. Upon receiving a CS, in addition Therefore a node needs to keep track of its contacts and their
to checking for overlap with ss vicinity and the vicinities of paths. This is done using soft-state periodic polling of the
all the already-selected-contacts (Contact_List), the receiving contacts as follows.
node also checks for overlap with the vicinities of any of the 1 Shown: 500 nodes, 710 m 710 m, Tx range = 50 m, R = 3, r = 20,
nodes on the Edge_List as well. It can be easily proven that D = 1. Similar trends were obtained for other simulation scenarios.
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 105
(1) Each node periodically sends a validation message to- possible, without having to initiate new searches from s. Thus
wards each of its contacts. These validation messages local recovery provides an efficient mechanism for validating
contain the path from a node s to the contact. contacts and recovering from changes in the contact paths. If
(2) Each node on the path that receives the validation mes- the next hop on the path (node c) is missing, the node that
sage checks if the next hop in the path is a directly con- received the validation message (node b) looks for the next
nected neighbor. If so, it forwards the validation message hops (c, d and e) in its vicinity routing table. If any of the
to the next hop node. If the next hop is missing, the node next hops (c, d or e) is found the vicinity, the path is updated
tries to salvage the path using local recovery, discussed and the validation message is forwarded to that next hop. If
later in this subsection. the lookups for all next hops fail, an error message is returned
to the source s, and another contact selection is initiated. Fig-
(3) If a path cannot be salvaged using local recovery, the con- ure 6 further illustrates an example of local recovery when
tact is considered to be lost. two nodes along the path to the contact (nodes c and d in this
(4) If the path to a contact is validated but the number of hops case) move.
to the contact does not lie between 2R and r, the contact
is considered to be lost. 3.4.4. Query mechanism
When a source node s (potentially any node), needs to reach
(5) After validating all the contacts, if the number of contacts a destination or target resource T it first checks its vicinity
left is less than the specified NoC, then a new contact se- table to see if T exists in its own vicinity. If T is not found
lection procedure is initiated. in the vicinity, s sends a Destination Search Query (DSQ) to
The local recovery mechanism is illustrated using an ex- its contacts. The DSQ contains the following information:
ample of a contact path (a b c d e). Assuming (1) depth of search (D), and (2) target resource ID (T ). Upon
reasonable values of node velocities and validation frequency receiving a DSQ, each contact checks the value of D. If D
(section 1 in our study), there is a high probability that if a is equal to 1, the contact performs a lookup for T in its own
node (say c) has moved out of a contact path (i.e., moved out vicinity. If T exists, then the path to T is returned to s, and
of transmission range of b), that it is still within the vicinity the query is considered successful. Otherwise, if D > 1, the
of the previous hop (b) in the path. Even in the case when contact receiving the DSQ decrements D by 1 and forwards
the moving node (c) is completely lost (because it has moved the DSQ to its contacts. In this way the DSQ travels through
out of the vicinity of the previous hop, b), some other node multiple levels of contacts until D reduces to 1.
further down the path (say dor e) might have moved into the The source node s first sends a DSQ with D = 1 to its
vicinity of the previous node (b). Local recovery takes advan- contacts. So only the first level contacts are queried with this
tage of these cases to recover from changes in the path when DSQ. After querying all its contacts if the source does not
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.
106 HELMY ET AL.
Table 1
Description of the various scenarios used for simulating CARD.
Scenario Nodes Area Transmission range No. of links Aver. node degree Network diameter Aver. path length (hops)
receive a path to the target within a specified time, it creates as flooding and bordercasting in terms of query overhead and
a new DSQ with D = 2 and sends it again to its contacts. query success rate.
Each contact observes that D = 2 and recognizes that this Table 1 shows the scenarios used in our simulations. These
query is not meant for itself. So it reduces the value of D in scenarios vary in number of nodes, network size, node den-
the DSQ by 1 and forwards it to its contacts. These contacts sity and transmission range. The variation is considered to
serve as second level contacts for the source. Upon receiving capture the effect of these factors on CARD. As was shown
the DSQ, a second level contact observes that D = 1 and it in figures 4 and 5, the edge method outperforms the proba-
does a lookup for the target T in its own vicinity and returns bilistic method. Therefore, we use the edge method (EM) for
the path to T , if found. In this way the value of D is used contact selection in the rest of our study.
to query multiple levels of contacts in a manner similar to the
expanding ring search. However, querying in CARD is much 4.1. Analysis of reachability
more efficient than the expanding ring search as the queries
are not flooded with different TTLs but are directed to indi- Analysis of the reachability, or query success rate, was con-
viual nodes (the contacts). Contacts leverage knowledge of ducted to understand how contacts help in increasing the view
their vicinity (gained through the proactive scheme operating of the network. Here we present results for a topology of 500
within the vicinity) to provide an efficient querying mecha- nodes spread over area of 710 m 710 m. The details can
nism. be seen from table 1, scenario number 5. Similar trends were
observed for other scenarios.
4. Evaluation and analysis
4.1.1. Varying vicinity radius (R)
In this section we present detailed simulation based evalua- Figure 7 shows the effect of increasing the vicinity radius (R)
tion and analysis of our architecture. NS-2 [1] along with our on reachability. As R increases, the reachability increases
CARD extensions and other utilities were used to generate and the reachability distribution in figure 7(a) shifts to the
various scenarios of ad hoc networks. The mobility model right; i.e., more nodes achieve higher percentage of reacha-
used for these simulations was the random way-point model. bility. This increase in reachability with the increase in R is
Our simulations so far did not consider MAC-layer issues. In due to increase in the number of nodes within the vicinity.
random way point model a node is assigned a random velocity As the value 2R approaches the maximum contact distance r
from [0, Vmax ] and assigned a destination location randomly. (r = 16 in this experiment), the region available for contact
Once the node reaches its destination it is assigned a random selection (between 2R and r) is reduced. This results in less
velocity and random destination again, so on. In the reacha- number of contacts being chosen. In figure 5, when R = 7,
bility analysis experiments the mobility was set to 0 to un- contacts can only be selected between 2R = 14 and r = 16
derstand the basic effects of the various architectural parame- hops from the source. This small region for contact selec-
ters on reachability characteristics. For the maintenance over- tion significantly reduces the number of contact and hence
head, total overhead and comparison experiments, continuous the reachability reduces as seen in figure 7(b). At this point
mobility was used (with no pauses) with Vmax = 20 m/s. most reachability is due to the vicinity of the source.
First we try to understand the effect of various parameters
such as vicinity radius (R), maximum contact distance (r), 4.1.2. Varying maximum contact distance (r)
the number of contacts (NoC), the depth of search (D) and Figure 8 shows the effect of increasing r on reachability.
network size (N) on reachability and overhead. Reachability Since contacts are selected between 2R and r hops from the
here is defined as the percentage of nodes that are reachable source, higher values of r provide a wider region for contact
from a source node. For overhead we consider the number selection. The mechanisms for the edge method for contact
of control messages; the contact selection (CS) messages and selection described earlier provide selection of contacts that
the periodic contact maintenance validation messages. Hav- have vicinities with reduced overlaps. This implies that as r
ing developed an understanding of the various parameters in increases a larger number of contacts can be selected with-
our architecture, we then compare it to other schemes such 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- 4.1.3. Varying number of contacts (NoC)
erage contact path length would increase (as more contacts NoC specifies the maximum number of contacts to be selected
are chosen at larger distances from the source). However, for each node. The actual number of contacts chosen may
once the vicinities of the contacts and the source become non- be less than this value. This is because of the limited re-
overlapping, for r > (2R + 8), we see no significant increase gion available for choosing contacts for given R and r accord-
in reachability with further increase in r. ing to the contact selection mechanism. Once this region has
108 HELMY ET AL.
been covered by vicinities of chosen contacts, choosing more 4.1.5. Varying network size
contacts in the same region is not possible as their vicinities Figure 11 shows the reachability distribution for three differ-
would overlap with the vicinities of the already chosen con- ent network sizes, N = 250, 500 and 1000 nodes. The area of
tacts. Therefore the contact selection mechanism prevents the the three networks has been chosen so that the node density
selection of more contacts. This can be seen in figure 9, in is almost same across the three networks. Figure 11 shows
which the reachability initially increases sharply as more and that for a given network (specified by the values of N and the
more contacts are chosen. However, the increase in reacha- area), the values of R and r can be configured to provide a
bility saturates beyond NoC = 6 as the actual number of con- desirable reachability distribution in which most of the nodes
tacts chosen saturates due to the effect of overlapping vicini- have a high value of reachability.
ties.
4.2. Contact selection and maintenance overhead analysis
4.1.4. Varying depth of search (D)
D specifies the levels of contacts that are queried in a breadth The overhead analysis measures the number of control mes-
first manner. When D = 1, a source node looking for a re- sages required for contact selection and maintenance. The
source beyond its vicinity, queries its first level contacts only. query overhead is considered in the next section. The over-
When D = 2, if none of the first level contacts contain the re- head considered in this section includes:
source in its vicinity, second level contacts (contacts of the
1. Contact selection overhead: This is the amount of CS traf-
first level contacts) are queried through the first level con-
fic generated for selecting new contacts. This includes
tacts. As can be seen from the figure 10, reachability increases overhead due to backtracking as described earlier.
sharply as the depth of search, D is increased. The depth of
search, D, results in a tree-like structure of contacts, improv- 2. Contact maintenance overhead: This is the traffic gener-
ing the reachability and success rate of CARD. ated by the contact path validation messages. Local recov-
Figure 10. Effect of depth of search (D) on reachability. N = 500, area = Figure 11. Reachability for different network sizes (D = 1).
710 m 710 m, Tx range = 50 m, R = 3, NoC = 10, r = 10.
(a) (b)
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.
CARD: A CONTACT-BASED ARCHITECTURE FOR RESOURCE DISCOVERY 109
ery, as described earlier, helps in reducing this part of the backtracking occurs when a node receiving a CS cannot be-
total overhead. come a contact due to overlap with already existing contacts.
As r increases, the possibility of this overlap decreases due
Results are shown for scenario number 5 in table 1 (N = to availability of a wider area for contact selection. This de-
500, area = 710 m 710 m). Similar trends were observed crease in back-tracking overhead is significantly more than
for other scenarios. the increase in overhead due to increased number of contacts
and contact path length. Therefore, the total contact selection
4.2.1. Varying number of contacts (NoC) and maintenance overhead decreases.
As shown in figure 12, as the number of contacts increases
the maintenance overhead increases sharply as more nodes 4.3. Maintenance overhead over time
are periodically maintained through the validation scheme.
Figure 15 shows the maintenance overhead per node over a
4.2.2. Varying maximum contact distance (r) 20 s period for Vmax = 20 m/s. The maintenance overhead
As r increases the number of selected contacts increases. The decreases steadily with time. However, the number of con-
increase in the number of contacts is due to the availabil- tacts increases slightly. This suggests that the source nodes
ity of a wider area for choosing contacts. Moreover, with find more stable contacts over time. Stable contacts may be
higher values of r, contacts may lie at greater distances from defined as those nodes that have low velocity relative to the
the source. That is, the contact path length is expected to be source node. For example, a node moving in the same direc-
higher for larger values of r. This suggests that the mainte- tion as source node with similar velocity could prove to be
nance overhead should increase with increase in r. However, a stable contact. Hence, over time, CARD leads to source
as shown in figure 13, the overhead actually decreases with in- nodes finding more stable contacts.
crease in r. Figure 14 explains this decrease in maintenance
overhead. Figure 14 shows that as the value of r increases 4.4. Comparison with related schemes (query overhead and
the backtracking overhead decreases significantly. Recall that total overhead)
(a) (b)
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).
110 HELMY ET AL.
Figure 15. Variation of overhead with time. N = 250, area = 710 m710 m, Figure 16. Query overhead for CARD, flooding and bordercasting.
Tx range = 50 m, NoC = 6, R = 4, r = 16, D = 1.
q = 0.005 query/m, and 8694% savings for high query rates
ing we investigated several techniques (probabilistic flood- q = 0.05 to 0.5 query/m).
ing, minimum dominating set, counter based methods) and For high mobility, in figures 18(a), (b) the savings are less
we show the results for those settings that achieved success than low mobility scenarios, nonetheless they are still signif-
rate of 90%. This was equivalent to probabilistic flooding as icant for moderate to high query rates (2275% savings for
in [19] with p = 0.65. For CARD the values of R and r used q = 0.05 query/m, 7993% savings for q = 0.5 query/m
were chosen as the values that gave maximum reachability for over flooding or bordercast). For low query rates and high
that particular network size. This information was obtained mobility however, e.g., for 20 m/s and q = 0.005 query/m,
from previous results shown under the analysis of CARD with CARD and bordercasting perform worse than flooding, where
respect to various parameters (see figure 11. Reachability for maintenance overhead dominates and only very few queries
different network sizes). Flooding and bordercasting result in are triggered (an unlikely scenario in mobile ad hoc net-
100% success in queries, smart flooding achieved 90% suc- works). For high mobility, large-scale, high query rates (1000
cess rate, and CARD showed a 95% success rate with D = 3. nodes, 20 m/s, 0.5 query/m), we get savings between 79%
CARDs success rate can be increased by increasing D, or (vs. bordercasting) and 87% (vs. flooding).
with resource replication. No replication is assumed in our To further understand the effect of query rate and mobility
study. As can be seen from figure 16, CARD leads to sig- on the total overhead we investigate the overhead ratio (OR)
nificant savings in communication overhead over the other metric for CARD over the total overhead of bordercast and
two approaches. CARD incurs, on average, around 5% of flooding. This metric enables us to have a more comprehen-
the query overhead for flooding, and around 10% or more of sive view of the operating conditions under which CARD is
the query overhead of bordercasting or smart flooding. We favorable. Let OR(C/B) be the overhead ratio for CARD over
note that smart flooding achieves the least success rate. To bordercast, and OR(C/F) and OR(C/S) be the overhead ratio
increase the success rate for smart flooding the overhead ap- of CARD over flooding and smart flooding. Let CSM be the
proaches that of flooding. contact selection and maintenance overhead, and let ZO be
What is not shown in figure 16, however, is the effect the zone (or vicinity) maintenance overhead, both in packets
of contact and vicinity maintenance. For that we show the per node per m/s. Also, let CQO be the CARD query over-
following total overhead comparison results. Maintenance head in packets per query, hence q CQO is the overhead in
overhead (for contacts and vicinity) is a function of mobil- packets per node per m/s. Define BQO as the query overhead
ity and simulation time. Its cost is amortized over the num- for bordercast. Hence, we get
ber of queries performed during that period. Hence, we CSM + ZO + q CQO
present our results as function of the query rate per mobil- OR(C/B) = .
ZO + q BQO
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 Similarly, we have
node. We show results for simulations with Vmax = 1 m/s CSM + ZO + q CQO
and 20 m/s, for various query rates q for 20 seconds of sim- OR(C/F) = and
q FQO
ulated time. These results take into consideration the con-
CSM + ZO + q CQO
tact selection and maintenance overhead, the vicinity estab- OR(C/S) = ,
lishment and maintenance overhead and the query overhead. q SQO
As can be seen from figures 17 and 18, the advantage of us- where FQO and SQO is the flooding and smart flooding over-
ing contacts becomes clearer for higher query rates, where head in packets per query, respectively.
the cost of maintenance is amortized over a large number of OR(C/B) and OR(C/F) were evaluated for q = 0.01 to
queries. For low mobility, in figures 17(a) and (b), the main- 100 query/sec/(m/s) per node. Figure 19 shows results for
tenance overhead is low and the advantages of using con- OR(C/B) and figure 20 shows results for OR(C/F). From the
tacts are the clearest (4685% savings for low query rates 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.
References
[15] J. Liu, Q. Zhang, W. Zhu, J. Zhang and B. Li, A novel framework for Ahmed Helmy received his Ph.D. in computer sci-
QoS-aware resource discovery in MANets, in: IEEE Internat. Conf. on ence (1999), M.S. in electrical engineering (1995)
Communications (ICC) (May 2002). from the University of Southern California, M.S.
[16] W. Lou and J. Wu, On reducing broadcast redundancy in ad hoc wire- Eng. Math. (1994) and B.S. in electronics and com-
less networks, IEEE Transactions on Mobile Computing 1(2) (2002). munications engineering (1992) from Cairo Univer-
[17] M. Mitzenmacher, Compressed bloom filters, in: The Twentieth ACM sity, Egypt. Since 1999, he has been an Assistant
Symposium on Principles of Distributed Computing (PODC) (August Professor of Electrical Engineering at the University
2001). of Southern California. In 2002, he received the Na-
[18] S. Murthy and J.J. Garcia-Luna-Aceves, An efficient routing protocol tional Science Foundation (NSF) CAREER Award.
for wireless networks, Mobile Networks and Applications (Special Is- In 2000 he received the USC Zumberge Research
sue on Routing in Mobile Communication Networks) (October 1996). Award, and in 2002 he received the best paper award from the IEEE/IFIP
[19] S. Ni, Y. Tseng, Y. Chen and J. Sheu, The broadcast Storm problem International Conference on Management of Multimedia Networks and Ser-
in a mobile ad hoc network, in: Proc. of the ACM MOBICOM Conf. vices (MMNS). In 2000, he founded and is currently directing the
(August 1999) pp. 151162. wireless networking laboratory at USC. His current research interests lie
[20] M. Pearlman and Z. Haas, Determining the optimal configuration for in the areas of protocol design and analysis for mobile ad hoc and sensor
the zone routing protocol, IEEE Journal on Selected Areas in Commu- networks, mobility modeling, design and testing of multicast protocols, IP
nications 8 (1999) 13951414. micro-mobility, and network simulation.
[21] C.E. Perkins and P. Bhagwat, Highly dynamic destination-sequenced E-mail: helmy@usc.edu
distance-vector routing (DSDV) for mobile computers, ACM Computer
Communications Review (October 1994) 234244. Saurabh Garg received his M.S. in computer science from the University of
[22] C.E. Perkins, E.M. Royer and S.R. Das, Ad hoc on-demand distance Southern California in May 2003 and B.S. in engineering from University of
vector routing, IETF Internet draft (October 1999). Delhi, India. He is currently working as a Programmer for National Center
[23] N. Sadagopan, B. Krishnamachari and A. Helmy, Active query for- for Ecological Analysis and Synthesis (NCEAS) affiliated with University
warding in sensor networks (ACQUIRE), Ad Hoc Networks Journal of California, Santa Barbara. He has worked with researchers in fields of
(2004) to appear. wireless networking, linguistics dialogue management and ecology. Before
[24] N. Sadagopan, B. Krishnamachari and A. Helmy, The ACQUIRE joining NCEAS he was working for Institute of Creative Technologies at the
mechanism for efficient querying in sensor networks, in: First IEEE University of Southern California in the field of dialogue management. In
Internat. Workshop on Sensor Network Protocols and Applications wireless networking, he has worked on network simulation, protocol design
(SNPA), in conjunction with IEEE ICC, Anchorage (May 2003) and analysis.
pp. 149155. E-mail: sgarg@usc.edu
[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
Nitin Nahata received his M.S. in computer science from the University of
ACM MOBICOM Conf. (The Ninth Annual Internat. Conf. on Mobile
Southern California in December 2002. He is currently working as a software
Computing and Networking), San Diego, CA (September 2003).
engineer at Dynamix Technologies Inc.
[26] D.J. Watts, The dynamics of networks between order and randomness,
E-mail: nnahata@usc.edu
in: Small Worlds (Princeton Univ. Press, Princeton, 1999).
[27] D. Watts and S. Strogatz, Collective dynamics of small-world net-
Priyatham Pamu received his M.S. in computer science from the University
works, Nature 393 (4 June 1998).
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.
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 The state of the art in WSN design is largely ad-hoc sys-
tem planning and resource management are done without a
Wireless sensor networks (WSNs) are being developed for a systematic methodology. This can lead to inefficient utiliza-
wide range of civil and military applications, such as target tion of the system. The main motivation of our efforts is to
tracking, infrastructure monitoring, habitat sensing, and bat- develop techniques for systematic and rapid design and de-
tlefield surveillance [6,10]. WSNs usually contain a number ployment of WSN applications [3,25,31].
of networked sensor nodes with each sensor node consisting We focus on the development of energy-efficient collabo-
of computation, communication, and sensing devices. These rative algorithms for WSNs based on high-level computation
sensor nodes collaborate with each other to realize certain ap- models of WSNs. Such high-level models allow designers to
plications. make informed decisions regarding energy and time tradeoffs
For instance, in a target tracking application, up to thou- at the node and network level creating a modular, layered
sands of sensor nodes are dispersed over a specific area of paradigm for application development. Toward such a goal,
interest. The sensor nodes are usually organized into clus- we study the following problem in this paper.
ters [13,31] with each cluster consisting of tens of sensor
nodes. Distributed signal detection and collaborative data Energy-balanced task allocation problem. We consider a
processing are performed within each cluster for detecting, single-hop cluster of homogeneous sensor nodes connected
identifying, and tracking vehicles. Some of the operations through multiple wireless channels. Each sensor node is
involved in such data processing include the LU factoriza- equipped with dynamic voltage scaling (DVS) [30]. The tar-
tion [5] and the Fast Fourier Transformation (FFT) [7]. get application consists of a set of communicating tasks.
Energy efficiency is a key concern in WSNs. The large Throughout the paper, the term activity refers to either a com-
number of sensor nodes involved in the system and the need putation task or a communication request. We consider an
to operate over a long period of time require energy-aware epoch-based scenario [18], where an instance of the appli-
design and operation at all levels of abstraction, from the cation is executed during the beginning of each epoch and
physical layer to the application layer. However, while many must be completed before the end of the epoch. Such a re-
hardware techniques [1,14], network protocols [13,16], and quirement is usually called the latency constraint. We use
data processing algorithms [18,19] have been proposed for the term period to indicate the length of each epoch. Also,
energy-aware design, systematic mechanisms for designing we assume that time-synchronization schemes (e.g., [9]) are
energy-aware collaborative processing between sensor nodes available within the cluster.
still need to be addressed. We consider the exclusive access constraint. Specifically, a
This work is supported by the DARPA Power Aware Computing and Com- non-preemptive scheduling policy is employed by each sensor
munication Program under contract no. F33615-C-00-1633. A preliminary node and each wireless channel. Also, at any time, a sensor
version of this paper appears in ACM LCTES 2003. node can receive or send data by using at most one channel.
116 YU AND PRASANNA
The underlying network protocol is assumed to be capable of given in section 4. The 3-phase heuristic is described in sec-
scheduling a communication activity over a specified chan- tion 5. Techniques, such as modulation scaling, are incor-
nel according to the start and finish time of the activity. Such porated into our approaches in section 6. Simulation results
a scheduling policy requires coarse-level bandwidth reserva- are demonstrated in section 7. Finally, we give concluding
tion mechanisms, which can be provided by, for example, a remarks in section 8.
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 2. Related work
allocated onto different sensor nodes. Extensive research efforts have studied the problem of energy-
A task allocation is defined as (1) the assignment of tasks efficient task allocation and scheduling with DVS in uni-
onto sensor nodes, (2) the voltage settings of tasks, (3) the processor real-time systems, including [2,15,24,30]. Re-
assignment of communication activities onto channels, and cently, research interests have been shifted to multi-processor
(4) the scheduling of computation and communication activ- systems. A list-scheduling based heuristic is proposed in [12],
ities. Our general goal is to find an allocation in order to to dynamically recalculate the priority of communicating
maximize the lifetime of the cluster. Toward such a goal, we tasks. In [17], static and dynamic variable voltage schedul-
propose an energy-balanced task allocation such that the max- ing heuristics for real-time heterogeneous embedded systems
imal energy dissipation among all sensor nodes during each are proposed. An approach based on critical-path is used for
period is minimized, subject to the latency, exclusive access, selecting the voltage settings of tasks. However, both [12]
and task placement constraints. and [17] assume that the task assignment is given. A sim-
Our contributions. The idea of energy-balanced task alloca- ilar problem to the one studied in this paper is investigated
tion to a single-hop cluster in WSNs is proposed. As we shall in [33]. A two-phase framework is presented to first deter-
see in section 2, most of the previous efforts in energy-aware mine the allocation of tasks onto processors and then the volt-
task allocation or resource management try to minimize the age settings of tasks using convex programming. In [34], a
overall energy dissipation of the system. This strategy may dynamic processor voltage adjustment mechanism for a ho-
not be suitable in the context of WSNs, since each sensor node mogeneous multi-processor environment is discussed. How-
is equipped with its own energy source. Moreover, for event- ever, the time and energy costs for communication activities
driven systems, applications often need to be executed after are not addressed in any of [12,33,34].
the system has been working for sometime. In such a case, an The goal of all the above works is to minimize the overall
energy-balanced task allocation should also consider the fact energy dissipation of the system. While such a goal is reason-
that the remaining energy can vary among sensor nodes. able for tightly coupled systems, it does not capture the nature
To the best of the authors knowledge, this is the first work of WSNs. The reason is that to minimize the overall energy
for task allocation in WSNs that considers the time and en- dissipation can lead to heavy use of energy-effective sensor
ergy costs of both the computation and communication ac- nodes, regardless of their remaining energy. The consequent
tivities. We first present an integer linear programming (ILP) short lifetime of such sensor nodes will very likely hinder the
formulation of our problem. The optimal solution of the prob- system from delivering required performance. This weakness
lem can be obtained by using a commercial software pack- is a major motivation of the proposed energy-balanced task
age such as [26], though the running time of such a software allocation.
can be large. Next, we propose a polynomial time 3-phase Our work considers the energy and time costs of both com-
heuristic. Finally, we incorporate techniques that explore the putation and communication activities. As indicated by sev-
latency-energy tradeoffs of communication activities, such as eral researches, wireless communication is a major source of
modulation scaling [23]. energy dissipation in WSNs. By incorporating techniques
Our simulation results show that for small scale problems, such as modulation scaling, we can greatly improve the
up to 5x lifetime improvement is achieved by the ILP-based energy-efficiency of the system.
approach, compared with the case where no DVS is used. Energy-balanced task allocation bears some resemblance
Also, the 3-phase heuristic achieves up to 63% of the sys- to load-balance in distributed computing. However, the com-
tem lifetime obtained by the ILP-based approach. For large munication activities over the same wireless channel need to
scale problems, the 3-phase heuristic achieves up to 3.5x life- be serialized such that run-time contentions can be avoided.
time improvement when only DVS is used. By incorporat- The serialization imposes new challenges that distinguish our
ing modulation scaling, up to 10x lifetime improvement was problem from most of the existing works for load-balance or
observed. Simulations were also conducted for application real-time scheduling in distributed systems.
graphs from two real world problems LU factorization and
FFT. We observed a lifetime improvement of up to 8x for the 3. Problem definition
LU factorization algorithm and up to 9x for FFT.
3.1. System model
Paper organization. We discuss the related work in sec-
tion 2. The energy-balanced task allocation problem is de- We consider a set of m homogeneous sensor nodes, PE =
fined in section 3. The ILP formulation of the problem is {PEi : i = 1, . . . , m}, connected by a single-hop wireless
ENERGY-BALANCED TASK ALLOCATION 117
network with K communication channels. The homogene- a source task. A task with no outgoing edges is called a sink
ity refers to the same processor and radio capabilities. Each task.
sensor node is equipped with D discrete voltage levels, listed For most applications in WSNs, the source tasks are used
as V = {Vi : i = 1, . . . , D} in decreasing order. Each voltage for sensing or gathering raw data. For ease of analysis, the
level in V corresponds to a specific computation speed (given task placement constraint is defined as that no two source
in cycles per second) of the processor. Let SPj denote the tasks can be assigned to the same sensor node. Nevertheless,
speed of Vj . Let Ri denote the remaining energy of PEi . For our models and approachs can be extended to handle the gen-
ease of analysis, we assume that the processors consume zero eral case that any pair of tasks must be or must not be assigned
power during idle state. to the same sensor node.
Regarding the exclusive access constraint, we assume that For any task Ti T , let Ci denote its workload in terms of
a non-preemptive scheduling policy is employed by each sen- the worst-case number of required computation cycles. The
sor node and each wireless channel. In other words, the execution time of Ti on any voltage level Vj V , tij , can
time duration scheduled for different computation (commu- be calculated as tij = Ci /(SPj ). The voltage level of a sen-
nication) activities over the same sensor node (wireless chan- sor node is assumed to be dynamically switched, if necessary,
nel) cannot overlap with each other. Moreover, the under- upon the arrival of a task instance. Because at most one switch
lying communication protocols are assumed to be capable is needed for executing a task instance, the associated time
of scheduling communication activities according to the start overhead is assumed to be included in the workload of the
time of each activity in order to avoid run-time contentions. task. From [4], the power consumption for executing a task
We assume all channels have the same bandwidth. Let de- follows a monotonically increasing and strictly convex func-
note the time for transmitting one data unit between two sen- tion of the computation speed, gi (), which can be represented
sor nodes over any channel. For ease of analysis, we assume as a polynomial function of at least second degree. Hence, the
that such a transmission costs the same amount of energy at energy dissipation for executing Ti on Vj , eij , can be calcu-
both the sender and the receiver, denoted by . Let s and s lated as eij = gi (SPj )tij . The exact forms of gi () can vary
denote the startup time and energy costs for communication. for different tasks based on their instruction components.
The data transmission between two tasks on the same sensor The communication load of any edge Ei E is rep-
node is performed through the local memory with zero time resented by its weight wi , as the number of data units to
and energy costs. be transmitted. We assume that all the data of an edge is
For ease of analysis, we assume that the radios are com- transmitted in one data packet with variable size. For an
pletely shutdown in idle state. The energy cost for shutting edge Ei = (j, k), let ti and ei denote the time and energy
down and restarting the radio is assumed to be included in s . costs of the corresponding communication activity, if tasks
Low power paging or signaling channel mechanisms can be Tj and Tk are not assigned to the same sensor node. We have
used for synchronization between sensor nodes when the ra- ti = s + wi and ei = s + wi .
dios are shutdown. However, the modeling of the power con-
sumption for such mechanisms is beyond the scope of this 3.3. Task allocation
paper. We also assume that computation and communication
Based on the above system and application models, a task al-
activities can be parallelly executed on any sensor node.
location is defined as (1) the assignment of tasks onto sensor
nodes, (2) the voltage settings of tasks, (3) the assignment of
3.2. Application model communication activities onto channels, and (4) the schedul-
ing of computation and communication activities. Each task
An epoch-based application [18] consisting of a set of com- can be assigned to exactly one sensor node with a fixed volt-
municating tasks is considered. Let P denote the period of the age setting. Also, each communication activity can be as-
application, which is the length of each epoch. An instance of signed to exactly one channel. An allocation is feasible if
the application is activated at time kP , and must be completed it satisfies the latency, exclusive access, and task placement
by the relative deadline, (k + 1)P , where k = 0, 1, 2, . . . . constraints.
The structure of the application is represented by a directed The system lifetime is defined as the time duration from the
acyclic graph (DAG), G = (T , E), where node set T denotes time when the application starts execution to the time when
the set of n tasks, {Ti : i = 1, . . . , n}, and edge set E de- any sensor node in the cluster fails due to depleted energy.
notes the set of e directed communication activities between A general solution to maximize the system lifetime is to allow
tasks, {Ei : i = 1, . . . , e}. Every edge in E pointing from variable task allocations in different periods. Consequently,
node Ti to Tj , denoted as (i, j ), means that the output of the energy cost for each sensor node may vary in different pe-
task Ti needs to be transmitted to Tj before Tj can start com- riods. However, due to the high complexity raised by such a
putation. There is a precedence constraint on two tasks Ti solution, we assume that the task allocation remains the same
and Tj , if there is a path of alternate nodes and edges from Ti for all application periods. That is, the behavior of the system
to Tj in the DAG. Similarly, there is a precedence constraint repeats for each period and every sensor node spends the same
on two communication activities, (i, j ) and (i , j ), if there is energy duration each period. Let Ei denote the energy dissi-
a path from Tj to Ti . A task with no incoming edges is called pation of PEi PE during each application period. Given an
118 YU AND PRASANNA
allocation, the system lifetime (in number of periods) can be enforced by constraint set 1. However, for activities that do
calculated as mini {Ri /Ei }. A feasible allocation is optimal not have precedence constraints between them, an extra set of
if the corresponding system lifetime is maximized among all constraints are needed (constraint set 2 in figure 2) to enforce
the feasible allocations. the exclusive access constraint. In addition, the task place-
Note that a more complex definition of the system lifetime ment constraint is captured by the constraint set 3 in figure 2.
would be the time period from the beginning of the applica- The complete ILP formulation is given in figure 3, where
tion execution to the time when not enough sensor nodes are E is an auxiliary variable. In the figure, the factor |xik xj k |
alive to deliver required performance. However, such a defin- means that the energy cost for (i, j ) is counted if exactly one
ition is quite application-specific. Thus, a simple but general of Ti or Tj is assigned to PEk , but not both. Clearly, the pre-
definition of the system lifetime is adopted in this paper. Now, sented formulation is nonlinear. It can be transformed into
our task allocation problem can be informally stated as: an ILP formulation by standard linearization techniques [29].
Due to the space limitation, we omit the details of lineariza-
Find an allocation of a set of communicating tasks onto a
tion in this paper.
single-hop cluster that minimizes the maximal energy dis-
sipation among all sensor nodes during each application
period, normalized by their remaining energy. 5. Heuristic approach
Constraint set 1:
Ti 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
Constraint set 2:
Ti , Tj T , such that i = j and Ti ||Tj
pij = 1 pj i , // pij is the inverse of pj i
(j ) pij rij (i), // if Ti and Tj are assigned to the same sensor node, Ti
// completes before Tj starts execution iff pij = 1
(i) pj i rij (j ), // if Ti and Tj are assigned to the same sensor node, Tj
// completes before Ti starts execution iff pj i = 1
Ei , Ej E, such that Ei = (a, b), // Communication activities from the same sensor node
Ej = (a, c), b = c
qij = 1 qj i , // qij is the inverse of qj i
(j ) qij (1 rab )(1 rcd )(i), // Ei completes before Ej starts transmission iff qij = 1
(i) qj i (1 rab )(1 rcd )(j ), // Ej completes before Ei starts transmission iff qj i = 1
Ei , Ej E, such that Ei = (a, b), // Communication activities to the same sensor node
Ej = (c, b), a = c
qij = 1 qj i , // qij is the inverse of qj i
(j ) qij (1 rab )(1 rcd )(i), // Ei completes before Ej starts transmission iff qij = 1
(i) qj i (1 rab )(1 rcd )(j ), // Ej completes before Ei starts transmission iff qj i = 1
Ei , Ej E, such that Ei = (a, b),
Ej = (c, d), a = c, b = d, and Ei ||Ej
qij = 1 qj i , // qij is the inverse of qj i
sij = 1 iff k = 1, . . . , K, zik = zj k , //sij equals one if Ei and Ej are assigned to the same channel
(j ) qij (1 rab )(1 rcd )sij (i), // if Ei and Ej are assigned to the same channel, Ei completes
// before Ej starts transmission iff qij = 1
(i) qj i (1 rab )(1 rcd )sij (j ). // if Ei and Ej are assigned to the same channel, Ej completes
// before Ei starts transmission iff qj i = 1
Constraint set 3:
Ti , Tj T , such that Ti and Tj are source tasks and i = j
rij = 0 // any two source tasks cannot be assigned to the same sensor node
plication can be reduced by merging C(i) with C(j ) with- The pseudo code for Traverse() is shown in figure 5. In
out violating the task placement constraint, we perform the the code, we maintain a queue of activities, Qact , that stores
merge. Otherwise, Ti and Tj remain in two different clusters. all the ready computation or communication activities in their
In lines 3 and 6, the function Traverse() is called to traverse expected execution order. We also maintain a timestamp for
the DAG in order to determine the schedule of the tasks and each task cluster that indicates the finish time for all scheduled
hence L. tasks within the cluster. Similarly, we maintain a timestamp
120 YU AND PRASANNA
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- Phase 3. The voltage levels of tasks are adjusted in this phase
tially, is sorted into a non-increasing order of energy dis- with the goal to maximize the system lifetime. An iterative
sipation of clusters. Then, for each cluster in , we calcu- greedy heuristic is used (shown in figure 7). Let E denote
lated the norm-energy of every sensor node as if the cluster the maximum of the norm-energy among all sensor nodes.
is assigned to the sensor node (called expected norm-energy). The sensor node that determines E is called the critical node.
We then assign the cluster to the sensor node that gives the In each iteration, we find the task such that by lowering its
minimal expected norm-energy. In the code, function Tra- current voltage level to the next level, E can be decreased the
verseAssigned() is used to find the execution time of the ap- most. The increased latency caused by lowering the voltage
ENERGY-BALANCED TASK ALLOCATION 121
8. L L + tdj
takes O(n + e) time. Thus, phase 1 needs O(e(n + e)) time.
9. Lower the voltage of Tj to the next level In phase 2 (figure 6), the ordering in line 1 takes O(c log c)
10. Update edj in EDr ; resort EDr if necessary time. The outer iteration is executed c times. The results
of m possible assignments are compared in line 5. The tra-
11. Find the new critical sensor node, PEr ; update E
verse in line 8 takes O(n + e) time. Hence, phase 2 takes
12. If r = r O(c log c + mc + n + e) time. In phase 3 (figure 7), the sorting
13. r r ; i 1 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
14. Else i i + 1
needs O(log n) time. Let p denote the number of times for
15. L TraverseAssigned() calling TraverseAssigned() in line 12. The time complexity
16. Until E can not be reduced any more 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
Figure 7. Pseudo-code for phase 3. 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),
is added to L. Since the schedule of activities can be changed which is O(dn(n + e + log n) + e2 + mn) in the worst
by the latency increment, L is re-computed by traversing the case.
DAG every time it reaches P (in line 15).
An illustrative example. We illustrate the execution of the
In figure 7, edj denotes the energy gain by lowering the above heuristic through a simple example. We assume a clus-
current voltage of Tj to the next level, while tdj denotes the ter of 3 sensor nodes connected by 2 channels. Each sensor
incurred increment in latency. The array composed by edj s node have two voltage levels, Vh and Vl , with SPh = 1 and
for all tasks assigned to PEi is denoted as EDi . SPl = 0.3. We assume that it costs one time and energy unit
for transmitting one data unit over any channel. The applica-
Time complexity analysis. In phase 1 (figure 4), the While tion graph is shown in figure 8(a), with each circle represent-
iteration is executed e times. Function Traverse() in line 6 ing a task. The number close to each circle is the required
122 YU AND PRASANNA
workload, while the number on each edge is the weight of the a packet of si bits, assuming a fixed symbol rate Ri , the trans-
edge. The time and energy costs for executing tasks at the mission time can be calculated as [23]:
two voltage levels are given in figure 8(b). We assume that si
P = 250 time units. i = , (1)
bi Ri
The clustering steps in phase 1 is shown in figure 9. In this where bi is the modulation level in terms of the constellation
phase, the voltage levels of all tasks are set to Vh . The sorted size (number of bits per symbol). The corresponding energy
edge list with respect to edge weights is {(T4 , T6 ), (T1 , T2 ), dissipation can be modeled as a function of i , denoted as
(T3 , T6 ), (T6 , T7 ), (T2 , T7 ), (T5 , T6 ), (T1 , T3 )}. The table in fi (i ). We have [23]
figure 9 traces the execution of the algorithm, where Li is the
execution time of the application at the completion of step i. fi (i ) = Ci 2si /(i Ri ) 1 + Di i Ri , (2)
The subfigures (a) through (e) correspond to the application
where Ci is determined by the quality of transmission (in
graph at the completion of steps 0, 1, 2, 3 and 5, respectively.
terms of bit error rate) and the noise power, and Di is a
The clusters are marked with polygons in dash line. Note that
device-dependent parameter that determines the power con-
in steps 6 and 7, the clustering is not performed due to the
sumption of the electronic circuitry of the sensor nodes. The
task placement constraint.
energy-latency tradeoffs for transmitting 1 bit is plotted in fig-
During phase 2, we first calculate the energy dissipa-
ure 10. The settings for Ci , Di and Ri are extracted from [23].
tion for each cluster 190 energy units for cluster 1 =
{T1 , T2 , T7 }, 100 for the cluster 2 = {T3 , T4 , T6 }, and 50 for Also, we may estimate the energy dissipation for receiving the
cluster 3 = {T5 }. Since the remaining energy for the three packet as
sensor nodes are the same, we simply assign 1 to PE1 , 2 to fi (i ) = Di Ri i . (3)
PE2 , and 3 to PE3 .
Finally, we adjust the voltage levels of tasks. Since PE1 is In practice, the value of bi is typically set to positive even
the critical node, we first set the voltage level of T2 to Vl , integers, resulting in discrete values of i . For any communi-
which reduces E1 to 106 and increases L from 80 to 219. cation activity Ei E, let tij denote the time cost with bi set
Next, we set the voltage level of T1 to Vl , which further de- to the j th modulation level. Also, let eijs and eijr denote the
creases E1 to 92 and increases L to 242. After this step, the corresponding sending and receiving energy costs. We can
critical node becomes PE2 with E2 = 0.1. Since the latency calculate the values of tij s, eijs s, and eijr s based on equa-
constraint is 250, our heuristic terminates. tions (1)(3).
In the above example, we decreases the norm-energy of To modify our ILP formulation, a set of 01 variables {uij }
the critical sensor node from 0.19 to 0.1, implying a system are needed to indicate the modulation level of the communica-
lifetime improvement by a factor around 2. 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
6. Incorporating energy-latency tradeoffs for one, which states that the transmission time of Ei = (a, b)
communication activities depends on the modulation level for Ei and the locations of
Ta and Tb :
While DVS has been widely applied into various applications
for energy saving in computation activities, techniques for ex- Ei = (a, b) E, (i) = (i) + uij tij (1 rab ).
ploring the energy-latency tradeoffs of communication activ- j
ities are gaining interest. An important observation [11] is
that in many channel coding schemes, the transmission en- Moreover, we change the constraint on the auxiliary vari-
ergy can be significantly reduced by lowering the transmis- able E in figure 3 as follows:
sion power and increasing the duration of the transmission.
PEk
Techniques such as modulation scaling [23] have been pro-
posed for implementing such tradeoffs. Recently, algorithms Ti T {xik j (yij eij )}
+
for applying such techniques in the context of packet trans- Rk
Ei =(a,b)E {xak (1 xbk ) j (uij eij ) + (1 xak )xbk
missions or data gathering in wireless networks have been s r
j (uij eij )}
studied in [11,23,32]. Rk
Our approaches can be extended to incorporate the above E.
tradeoffs. In the following, we discuss through the example
of modulation scaling that explores the tradeoffs by adapting For the 3-phase heuristic, we assume that both voltage and
the modulation level to match the traffic load. modulation levels of the system are set to the highest options
For ease of analysis, we focus on the Quadrature Am- in phases 1 and 2. We then slightly modify phase 3, such
plitude Modulation (QAM) scheme [28]. The techniques that the energy savings achieved by lowering the modulation
presented in this paper are extendible to other modulation levels of communication activities are also examined. The
schemes as well. Given a communication activity Ei with modified pseudo code is shown in figure 11. One concern
ENERGY-BALANCED TASK ALLOCATION 123
(a) (b)
(c) (d)
(f)
(e)
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
The time and energy costs of communication activities are itively, a larger value of u implies a tighter latency constraint
determined by the number of data units to transmit and the and hence less latency laxity.
values of and . Based on the data for WINS, we set The remaining energy of sensor nodes follows a uniform
= 10 sec and = 1 J. To focus on the main issues, distribution between Emean (1 0.3), where Emean is a fairly
we set the startup energy dissipation of the radio to be zero. large number.
To study the effect of different communication load with re-
spect to the computation load, the number of bits per com- Small scale problems. We first conducted simulations for
munication activity follows a uniform distribution between small scale problems, with 3 sensor nodes, 3 voltage levels,
200 CCR (1 0.2), where CCR (communication to compu- 2 channel, and 710 tasks. The number of source tasks in the
tation ratio) is a parameter indicating the ratio of the average application graph is set to 2, while the maximal in-degree and
execution time of the communication activities to that of the out-degree for each node are set to 3. A commercial software
computation activities. Intuitively, a larger value of CCR im- package, LINDO [26], was used to solve the ILP problems.
plies a relatively heavier communication loads compared with Due to the large running time for solving some problem in-
the computation loads. Note that by varying CCR, we abstract stances, LINDO was interrupted after two hours of execution
if the optimal solution was not yet found. Then, the best so-
not only the variations in the amount of transmitted data, but
lution obtained so far was returned. We observed that in most
also the variations in the relative speed of computation and
cases, LINDO was able to find the optimal solution within
communication devices. In our simulations, CCR was varied
two hours.
within [0, 20].
The data shown in figure 12 is averaged over more than
The period of the application, P , was generated in the fol-
70 instances so that each data point has a 95% confidence in-
lowing way. We first define the distance of a node in the appli-
terval with a 10% precision. In figure 12(a), we illustrate the
cation DAG as the number of edges in the longest path from
lifetime improvement achieved by the ILP-based approach,
the source to the node. Nodes are then divided into layers,
which is calculated as LT ILP /LT raw 1. We can see an im-
with nodes in each layer having the same value of distance. provement around 3x5x. Figure 12(b) shows the perfor-
Since the average time to execute a task in the highest speed mance ratio of the 3-phase heuristic over the ILP-based ap-
is 2 msec, the computation time required for a layer is esti- proach, i.e., LT heu /LT ILP . We can see that the 3-phase heuris-
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 based approach for the conducted simulations.
in executing the tasks at each layer. In addition, the expected While the running time of the heuristic is around zero, the
number of communication activities initiated by a task is es- average running time of the ILP-based approach ranges from
timated as its out-degree subtracted by 1. Assuming there are 550 sec (n = 7, u = 0.5) to 5900 sec (n = 10, u = 0.8)
in total q communication activities requested by all the tasks on a Sun Blade1000 machine with a UltraSparc III 750 Mhz
in a specific layer, the corresponding time cost is estimated CPU.
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, Large scale problems. A set of simulations were conducted
where u [0, 1] is a parameter that approximates the overall for evaluating the performance of the 3-phase heuristic for
utilization of the system. The setting of u is important as it problems with 10 sensor nodes, 8 voltage levels, 4 channels,
determines the latency laxity for trading against energy. Intu- 60100 tasks, CCR [0, 20], and u [0, 1]. The number of
(a) (b)
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.
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).
much. This is understandable since the energy behaves as a to even numbers between 2 and 6. We set Ri = 1.7 104 so
monotonically increasing and strictly convex function of the that when bi = 6, it roughly takes 10 sec and 1 J to trans-
computation speed. The first derivative of the energy function mit a bit (as shown in figure 10).
tends to when the speed tends to . Thus, the most portion The simulations were conducted with 10 sensor nodes,
of energy saving is obtained by changing the speed from the 8 voltage levels, 3 modulation levels ({2, 4, 6}), 60 tasks,
highest option to some lower options, which can be efficiently u {0.0, 0.2, 0.5, 0.8, 1.0}, and CCR [0, 20]. Com-
achieved with 4 voltage levels per sensor node. pared with figure 13, we can observe a significant amount of
When u = 0.2, the latency laxity is so large that the volt- performance improvement in figure 16. For example, when
age level of most tasks can be set to the lowest option. Thus, u = 0.5, the highest lifetime improvement increases from
there is almost no improvement by increasing the number of 3x in figure 13(a) to 6x in figure 16(a) and even 10x in fig-
voltage levels beyond 2. ure 16(b). The difference in performance improvement of fig-
ures 16(a) and (b) is because that a larger Ci leads to larger
Incorporating modulation scaling. We used modulation energy/time ratio of communication activities, which in turn
scaling to illustrate the energy-latency tradeoffs for commu- gives more advantage in reducing the communication energy
nication activities. We assume that all sensor nodes have the by utilizing modulation scaling.
identical settings for parameters Ci , Di and Ri . From [23], Similar to figure 13, larger improvement is observed when
we set Di = 107 [23]. To investigate the impact of different u becomes smaller. In addition, the miss rate of the heuristic
energy/time ratio for data transmission, we set Ci to 107 and exhibits a similar trend as the cases with DVS only.
106 for different instances. The modulation level, bi , was set
7.2. Application graphs from real world problems
(a) (b)
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 ).
128 YU AND PRASANNA
(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
8. Concluding remarks
(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 YU AND PRASANNA
to 8x for the LU factorization algorithm and an improvement [18] S.R. Madden, M.J. Franklin, J.M. Hellerstein and W. Hong, TAG: a
of up to 9x for the FFT algorithm. Tiny AGgregation service for ad-hoc sensor networks, in: Symposium
on Operating Systems Design and Implementation (OSDI) (December
In the future, we would like to validate our approaches us-
2002).
ing real systems. We are particularly interested in advanced [19] C. Meesookho, S. Narayanan and C.S. Raghavendra, Collaborative
applications for WSNs, where systematic methodologies for classification applications in sensor networks, in: 2nd IEEE Sensor Ar-
task allocation are mostly required for rapid and automated ray and Multichannel Signal Processing Workshop (August 2002).
[20] P. Meja-Alvarez, E. Levner and D. Moss, An integrated heuristic ap-
system design.
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
References 558.
[22] V. Sarkar, Partitioning and Scheduling Programs for Execution on Mul-
[1] G. Asada, M. Dong, T.S. Lin, F. Newberg, G. Pottie and W.J. Kaiser, tiprocessors (MIT Press, Cambridge, MA, 1989).
Wireless integrated network sensor: Low power systems on a chip, in: [23] C. Schurgers, O. Aberhorne and M.B. Srivastava, Modulation scaling
ESSCIRC 98 (1998). for energy-aware communication systems, in: ISLPED (2001) pp. 96
[2] H. Aydin, R. Melhem, D. Moss and P.M. Alvarez, Determining opti- 99.
mal processor speeds for periodic real-time tasks with different power [24] Y. Shin, K. Choi and T. Sakurai, Power optimization of real-time em-
characteristics, in: 13th Euromicro Conf. on Real-Time Systems (June bedded systems on variable speed processors, in: IEEE/ACM Internat.
2001). Conf. on Computer-Aided Design (2000) pp. 365368.
[3] A. Bakshi, J. Ou and V.K. Prasanna, Towards automatic synthesis of [25] M. Singh and V.K. Prasanna, A hierarchical model for distributed col-
a class of application-specific sensor networks, in: Internat. Conf. on laborative computation in wirelss sensor networks, in: 5th Workshop
Compilers, Architecture, and Synthesis for Embedded Systems (CASES) on Advances in Parallel and Distributed Computational Models (April
(October 2002). 2003).
[4] T.D. Burd, T.A. Pering, A.J. Stratakos and R.W. Brodersen, A dynamic [26] The LINDO System Inc., http://www.lindo.com.
voltage scaled microprocessor system, IEEE Journal of Solid-State Cir- [27] The WINS Project, Rockwell Science Center, http://wins.rsc.
cuits 35(11) (2000) 15711580. rockwell.com.
[5] M. Conard, M. Marrakchi, Y. Robert and D. Trystram, Parallel [28] T. Ue, S. Sampei, N. Morinaga and K. Hamaguchi, Symbol rate and
Gaussian elimination on an MIMD computer, Parallel Computing 6 modulation level-controlled adaptive modulation/TDMA/TDD system
(1988) 275295. for high-bit rate wireless data transmission, IEEE Transactions on Ve-
[6] S. Conner, L. Krishnamurthy and R. Want, Making everyday life easier hicular Technology 47(4) (1998) 11341147.
using dense sensor networks, in: ACM UBICOMP (2001). [29] H.P. Williams, Model Building in Mathematical Programming (Wiley,
[7] T.H. Cormen, C.E. Leiserson and R.L. Rivest, Introduction to Algo- New York, 1999).
rithms (MIT Press, Cambridge, MA, 1990). [30] F. Yao, A. Demers and S. Shenker, A scheduling model for reduced
[8] R.P. Dick, D.L. Rhodes and W. Wolf, TGFF: Task graphs for free, CPU energy, IEEE Annual Foundations of Computer Science (1995)
in: Internat. Workshop on Hardware/Software Codesign (March 1998) 374382.
pp. 97101. [31] Y. Yu, B. Krishnamachari and V.K. Prasanna, Issues in designing mid-
[9] J. Elson, L. Girod and D. Estrin, Fine-grained network time synchro- dleware for wireless sensor networks, IEEE Network Magazine (Spe-
nization using reference broadcasts, in: Symposium on Operating Sys- cial Issue on Middleware Technologies for Future Communication Net-
tems Design and Implementation (OSDI) (December 2002). works) 18(1) (2004) 1521.
[10] D. Estrin, L. Girod, G. Pottie and M.B. Srivastava, Instrumenting the [32] Y. Yu, B. Krishnamachari and V.K. Prasanna, Energy-latency tradeoffs
world with wireless sensor networks, in: Internat. Conf. on Acoustics, for data gathering in wireless sensor networks, in: Proceedings of IEEE
Speech and Signal Processing (ICASSP) (May 2001). INFOCOM (March 2004).
[11] A.E. Gamal, C. Nair, B. Prabhakar, E. Uysal-Biyikoglu and S. Zahedi, [33] Y. Zhang, X. Hu and D.Z. Chen, Task scheduling and voltage selection
Energy-efficient scheduling of packet transmissions over wireless net- for energy minimization, in: Design Automation Conf. (DAC) (2002).
works, in: IEEE INFOCOM (2002). [34] D. Zhu, R. Melhem and B. Childers, Scheduling with dynamic volt-
[12] F. Gruian and K. Kuchcinski, LEneS: Task scheduling for low-energy age/speed adjustment using slack reclamation in multi-processor real-
systems using variable supply voltage processors, in: Design Automa- time systems, in: IEEE Real-Time Systems Symposium (RTSS) (Decem-
tion Conf. (DAC) (2001) pp. 449455. ber 2001).
[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- Yang Yu is a Ph.D. degree candidate in the Depart-
niques for low-power hard real-time systems on variable voltage ment of Electrical Engineering at the University of
processors, in: IEEE Real-Time Systems Symposium (RTSS) (Decem- Southern California (USC). He received both BS and
ber 1998). MS degrees in computer science from Shanghai Jiao-
[16] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion: Tong University in China. His research interests in-
A scalable and robust communication paradigm for sensor networks, clude energy-aware resource management for wire-
in: ACM/IEEE Internat. Conf. on Mobile Computing and Networking less sensor networks, especially in algorithmic so-
(MOBICOM) (2000). lutions for energy-aware communication scheduling
[17] J. Luo and N.K. Jha, Static and dynamic variable voltage scheduling and task allocation. He is a student member of the
algorithms for real-time heterogeneous distributed embedded systems, IEEE.
in: VLSI Design (January 2002). E-mail: yangyu@usc.edu
ENERGY-BALANCED TASK ALLOCATION 131
Viktor K. Prasanna is a Professor of Electrical En- sity. He has published extensively and consulted for industries in the above
gineering and Computer Science at the University areas. He is the steering committee co-chair of the International Parallel
of Southern California (USC). He is also a mem- and Distributed Processing Symposium (IPDPS) (merged IEEE International
ber of the NSF supported Integrated Media Sys- Parallel Processing Symposium (IPPS) and Symposium on Parallel and Dis-
tems Center (IMSC) and an associate member of the tributed Processing (SPDP)). He is the steering committee chair of the In-
Center for Applied Mathematical Sciences (CAMS) ternational Conference on High Performance Computing (HiPC). He serves
at USC. His research interests include high perfor- on the editorial boards of the Journal of Parallel and Distributed Computing.
mance computing, parallel and distributed systems, He is the editor-in-chief of the IEEE Transactions on Computers. He was
network computing, and embedded systems. He re- the founding chair of the IEEE Computer Societys Technical Committee on
ceived his BS in electronics engineering from Ban- Parallel Processing. He is a fellow of the IEEE.
galore University, MS from the School of Automation, Indian Institute of E-mail: prasanna@usc.edu
Science, and Ph.D. in computer science from the Pennsylvania State Univer-
Mobile Networks and Applications 10, 133149, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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 books, people), the factory floor (motors, small robotic de-
vices).
Networked sensors (or Smart Dust) are very large systems, There are many possible models for such networked sen-
comprised of a vast number of homogenous ultra-small fully sors. In this work, we consider networked sensors where
autonomous computing and communication devices that co- (a) all nodes in the network are homogenous and constrained
operate to achieve a large sensing task. Each device has one by low availability of resources (energy, communication) and
or more sensors, embedded processors and low-power radios, (b) the data being sensed by the nodes must be transmitted
and is normally battery operated. Examining each such sin- to a fixed control center located far away from the sensors.
gle device individually, might appear to have small utility. Thus direct communication between the sensor nodes and the
The realization of Smart Dust, however, lies in using and co- control center is impossible and/or expensive, since there are
coordinating a vast number of such devices. no high-energy nodes through which communication can
Smart Dust is a useful case of dynamic environments of proceed. This is the general framework for MITs AMPS
networked sensors that are spread over a global system and project [20], which focuses on innovative energy-optimized
try to communicate and compute efficiently and quickly, hav- solutions at all levels of the system hierarchy, from the phys-
ing only partial knowledge of the global conditions and hav- ical layer and communication protocols up to the application
ing poor energy and computing resources. Typically, these layer.
networked sensors coordinate to perform a common task. De- To motivate the challenges in designing such sensor net-
signing protocols to coordinate such systems (i.e., create a dy- works, we can consider the following scenario where local
namic and efficient network of these sensors) and monitoring detection and fast propagation to the authorities of the real-
their behavior as they operate in complex and dynamic global ization of a crucial event can be achieved using smart dust.
environments is of great importance for information gathering Think about thousand of disposable sensors scattered (e.g.,
and processing in many practical situations. thrown from an aircraft) over a forest. Each of these sensors
As an example, [11] points that integrated low-power sens- can monitor the temperature at a single, very small geograph-
ing devices will permit remote object monitoring and track- ical area. The sensors coordinate to establish an efficient, dy-
ing in inhospitable physical environments such as remote namic and short-lived communication network, dividing the
geographic regions or toxic urban locations. They will also task of monitoring the terrain and offering continuous moni-
enable low maintenance sensing in the field (vehicles, equip- toring of the environment in order to alert the authorities as
ment, personnel), the office buildings (projectors, furniture, soon as possible after a forest fire is detected by some sensor.
Several aspects of such systems of autonomous networked
Preliminary versions of this work have appeared in the 2nd ACM Work-
entities emerge, which are quite different from those posed
shop on Principles of Mobile Computing (POMC, 2002 [9]) and the 3rd by standard computer networks. Such aspects include the
Workshop on Mobile and Ad-hoc Networks (WMAN, 2003 [6]).
This work has been partially supported by the IST Program of the Eu- very poor and highly restricted resources (e.g., very low bat-
ropean Union under contract numbers IST-1999-14186 (ALCOM-FT) and tery power, low computing capabilities, total absence of syn-
IST-2001-33116 (FLAGS). chrony and anonymity). Network protocols must be designed
134 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
to achieve fault tolerance in the presence of individual node analytic and extensive experimental means, we investigate
failure while minimizing energy consumption. Another im- the relation between (a) the success probability and (b) the
portant aspect is the scalability to the change in network size, time efficiency of the protocol, to the maximum sleeping
node density and topology. The network topology changes time period, for various values of other parameters, such
over time as some nodes may die, or possibly because new as particle density, particle distribution and angle . We
nodes join later. interestingly note that the new protocol is efficient, despite
This work, continuing our line of research on communica- the fact that the particles are allowed to enter a sleeping
tion in ad-hoc mobile networks [7,8], is an attempt towards mode in order to save energy.
capturing the underlying foundational and algorithmic issues
in the design of such systems, abstracting accurately enough All protocols mentioned above are shown to be robust in
the real technological specifications involved and providing the following sense: (a) the protocols use the search and the
some first concrete results for the efficiency of a variety of backtrack phases to explore the active (non-faulty) next par-
smart dust protocols using an average case analysis. We focus ticles. Thus, the fact (demonstrated both by analysis and sim-
in this paper on the efficient use of smart dust in local detec- ulation) that the protocols succeed with high probability, ex-
tion and propagation protocols. We first provide an abstract hibits also fault-tolerance properties of the protocols; (b) our
model for smart dust systems which is simple enough to allow findings showing that the protocols succeed even in the case
an analysis to develop, being however at the same time quite of low densities also implies robustness.
realistic, in terms of the technological specifications of real
smart dust systems it captures. Then we define the problem Discussion of selected related work
of local detection and propagation using smart dust and also In the last few years, Sensor Networks have attracted a lot of
propose some concrete performance and robustness measures attention from researchers at all levels of the system hierarchy,
for the average case analysis of protocols for this problem. from the physical layer and communication protocols up to
the application layer.
Our results At the MAC level, many researchers have done research
For the local detection and propagation problem using smart work in an effort to minimize the power consumption. [27]
dust, we provide three protocols. All protocols are simple presents a contention-based protocol that tries to minimize en-
enough to be implemented in real smart dust systems despite ergy consumption due to node idle listening, by avoiding the
the severe energy and computing power limitations of such overhearing among neighboring nodes. A recent work [30]
systems. Furthermore, we give a rigorous average case analy- exploits a similar method for energy savings, and further re-
sis for the efficiency of these protocols. We consider a variety duce idle listening by avoiding any use of out-of-channel sig-
of performance and robustness criteria, such as propagation naling. Additionally, their protocol trades off per-node fair-
time, number of particle to particle transmissions (which also ness for further energy savings.
characterizes energy consumption and time efficiency, assum- For establishing communication and routing information
ing an efficient MAC protocol) and fault-tolerance: to the control center, mobile ad-hoc routing protocols [24]
may be used in sensor networks. However, although proto-
1. Our first protocol, which we call the local target protocol cols for mobile ad-hoc networks take into consideration en-
(LTP), uses a fast and cheap search phase which is assumed ergy conservation issues, most of them are not really suitable
to always return a nearby particle towards the authorities, for sensor networks. [19] presents a routing protocol suitable
uniformly in some range. We show that LTP is efficient, for sensor networks that makes greedy forwarding decisions
in the sense that it achieves a propagation time and an en- using only information about a nodes immediate neighbors
ergy consumption whose expected ratio over the optimal in the network topology. This approach achieves high scala-
solutions is at most /2 1.57. bility as the density of the network increases. [14] presents a
2. Our second protocol, the min-two uniform targets proto- clustering-based protocol that utilizes randomized rotation of
col (m2TP), applies the simple idea of getting at least two local cluster heads to evenly distribute the energy load among
particles towards the authorities and selecting the best in the sensors in the network. [21] introduces a new energy ef-
terms of propagation progress. It is, in fact, an optimized ficient routing protocol that does not provide periodic data
and more efficient version of the local target protocol, and monitoring (as in [14]), but instead nodes transmit data only
has an expected time and energy ratio over the optimal so- when sudden and drastic changes are sensed by the nodes. As
lutions which is at most 2 /8 1.24. such, this protocol is well suited for time critical applications
and compared to [14] achieves less energy consumption and
3. Next we provide tight upper bounds to the distribution of
response time.
the number of particle to particle data transmissions (and
A family of negotiation-based information dissemination
thus the efficiency) of a generalized target protocol.
protocols suitable for wireless sensor networks is presented
4. We propose a new protocol which we call the Sleep in [15]. Sensor Protocols for Information via Negotiation
Awake protocol (SWP), that explicitly uses the energy (SPIN) focus on the efficient dissemination of individual sen-
saving characteristics, such as the alteration of sleep and sor observations to all the sensors in a network. However,
awake time periods, of smart dust particles. By using both in contrast to classic flooding, in SPIN sensors negotiate with
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 135
each other about the data they possess using meta-data names. work, during the entire execution of the data propagation pro-
These negotiations ensure that nodes only transmit data when tocol. This property is important since it prolongs the net-
necessary, reducing the energy consumption for useless trans- works lifetime by avoiding early energy depletion of sen-
missions. sors. They propose a new algorithm that in each step decides
A data dissemination paradigm called directed diffusion whether to propagate data one-hop towards the final destina-
for sensor networks is presented in [18], where data-generated tion (the sink), or to send data directly to the sink. This ran-
by sensor nodes is named by attributevalue pairs. An ob- domized choice balances the (cheap) one-hop transimssions
server requests data by sending interests for named data; data with the direct transimissions to the sink, which are more ex-
matching the interest is then drawn down towards that node pensive but bypass the sensors lying close to the sink. Note
by selecting a single path or through multiple paths by us- that, in most protocols, these close to the sink sensors tend to
ing a low-latency tree. [17] presents an alternative approach be overused and die out early.
that constructs a greedy incremental tree that is more energy- In [1], the authors propose a new energy efficient and fault
efficient and improves path sharing. tolerant protocol for data propagation in smart dust networks,
We note that, as opposed to the work presented in this the Variable Transmission Range Protocol (VTRP). The basic
paper, the above research focuses on energy consumption idea of data propagation in VTRP is the varying range of data
without examining the time efficiency of their protocols. Fur- transmissions, i.e., they allow the transmission range to in-
thermore, these works contain basically protocol design and crease in various ways. Thus data propagation in the protocol
technical specifications, while quantitative aspects are only exhibits high fault-tolerance (by bypassing obstacles or faulty
experimentally evaluated and no theoretical analysis is given. sensors) and increases network lifetime (since critical sensors,
Note also that our protocols are quite general in the sense i.e., close to the control center are not overused). They im-
that (a) do not assume global network topology informa- plement the protocol and perform an extensive experimental
tion, (b) do not assume geolocation information (such as GPS evaluation and comparison to a representative protocol (LTP)
information) and (c) use very limited control message ex- of several important performance measures with a focus on
changes, thus having low communication overhead. energy consumption. The findings indeed demonstrate that
Finally, our third protocol is using a similar approach to the protocol achieves significant improvements in energy ef-
the recent work of [26], where a new technique called Sparse ficiency and network lifetime.
Topology and Energy Management (STEM) is proposed that In [23], Nikoletseas et al. (a) propose extended versions
aggressively puts nodes to sleep. Interestingly, the analysis of two data propagation protocols: the SleepAwake Prob-
and experiments of STEM show improvements of nearly two abilistic Forwarding Protocol (SW-PFR) and the Hierarchi-
orders of magnitude compared to sensor networks without cal Threshold sensitive Energy Efficient Network protocol
topology management. (H-TEEN). These non-trivial extensions aim at improving the
performance of the original protocols, by introducing sleep
Some recent work awake periods in the PFR protocol to save energy, and in-
troducing a hierarchy of clustering in the TEEN protocol to
In [4] the authors present a new protocol for data propaga- better cope with large networks areas; (b) they have imple-
tion that avoids flooding by probabilistically favoring certain mented the two protocols and performed an extensive exper-
(close to optimal) data transmissions. As shown by a geom- imental comparison (using simulation) of various important
etry analysis, the protocol is correct, since it always prop- measures of their performance with a focus on energy con-
agates data to the sink, under ideal network conditions (no sumption; (c) they investigate in detail the relative advantages
failures). Using stochastic processes, they show that the pro- and disadvantages of each protocol and discuss and explain
tocol is very energy efficient. Also, when part of the network their behavior; (d) in the light above they propose and discuss
is inoperative, the protocol manages to propagate data very a possible hybrid combination of the two protocols towards
close to the sink, thus in this sense it is robust. They finally optimizing certain goals. Efficient collision avoidance proto-
present and discuss large-scale experimental findings validat- cols, particularly useful for multipath data propagation, have
ing the analytical results. been proposed in [10].
In [5], the authors have implemented and experimentally A brief description of the technical specifications of state-
evaluated two variations of LTP, under new, more general and of-the-art sensor devices, a discussion of possible models
realistic modelling assumptions. They comparatively study used to abstract such networks and a presentation of some
LTP to PFR, by using extensive experiments, highlighting characteristic protocols for data propagation in sensor net-
their relative advantages and disadvantages. All protocols are works, along with an evaluation of their performance analy-
very successful. In the setting considered there, PFR seems sis, can be found in the recent book chapter of Boukerche and
to be faster while the LTP based protocols are more energy Nikoletseas [3].
efficient.
In [12], Euthimiou et al. study the problem of energy- 2. The model
balanced data propagation in wireless sensor networks. The
energy balance property guarantees that the average per sen- Smart dust is comprised of a vast number of ultra-small ho-
sor energy dissipation is the same for all sensors in the net- mogenous sensors, which we call grain particles. Each
136 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
are other processors that would meet the power and cost tar-
gets. The processor runs at 4 MHz, has 128 Kb of flash mem- 4. The problem
ory and 4 Kb of SDRAM. In a given network, thousands of
sensors could be continuously reporting data, creating heavy An adversary A selects a single particle, p, in the plane-cloud
data flow. Thus, the overall system is memory constrained, and allows it to monitor a local crucial event E. The general
but this characteristic is a common design challenge in any propagation problem P is the following:
wireless sensor network.
How can particle p, via cooperation with the rest of the
The MPR modules contain various sensor interfaces,
cloud, propagate information about event E to the receiv-
which are available through a small 51-pin connector that
ing wall W?
links the MPR and MTS modules. The interface includes: an
8-channel, 10-bit A/D converter; a serial UART port; and an
Definition 4.1. Let hopt (p, W) be the (optimal) number of
I2C serial port. This allows the MPR module to connect to a
hops (direct, vertical to W transmissions) needed to reach
variety of MTS sensor modules, including MTS modules that
the wall, in the ideal case in which particles always exist in
use analog sensors as well as digital smart sensors. The MPR
pair-wise distances R in the vertical line from p to W. Let
module has a guaranteed unique, hard-coded 64-bit address.
be a smart-dust propagation protocol, using a transmis-
The processors, radio, and a typical sensor load consumes
sion path of length L(, p, W) to send info about event E to
about 100 mW in active mode. This figure should be com-
wall W. Let h(, p, W) be the number of hops (transmis-
pared with the 30 A draw when all components are in sleep
sions) taken to reach W. The hops efficiency of protocol
mode. Figure 2 shows a synopsis of the MPR specs.
is the ratio
The MTS sensor boards currently include light/tempera-
ture, two-axis acceleration, and magnetic sensors and 420 mA h(, p, W)
Ch = .
transmitters. The wireless transmission is at 4 Kbps rate and hopt (p, W)
the transmission range may vary. Researchers are also devel-
oping a GPS board and a multisensor board that incorporates Clearly, the number of hops (transmissions) needed char-
a small speaker and light, temperature, magnetic, accelera- acterizes the energy consumption and the time needed to
tion, and acoustic (microphone) sensing devices. The MICA propagate the information E to the wall. Remark that hopt =
developers community welcomes additional sensor board de- d(p, W)/R, where d(p, W) is the (vertical) distance of p
signs. from the wall W.
In the case where protocol is randomized, or in the case
3.2. Software and the TinyOS where the distribution of the particles in the cloud is a random
distribution, the number of hops h and the efficiency ratio Ch
A considerable portion of the challenge faced by the devel- are random variables and we study here their expected values.
opers of MICA devices is in the software embedded in the The reason behind these definitions is that when p (or any
sensors. The software runs the hardware and networkmaking intermediate particle in the information propagation to W)
sensor measurements, routing measurement data, and control- looks around for a particle as near to W as possible to pass
ling power dissipation. In effect, it is the key ingredient that its information about E, it may not get any particle in the per-
makes the wireless sensor network produce useful informa- fect direction of the line vertical to W passing from p. This
tion. difficulty comes mainly from three causes: (a) due to the ini-
To this end, a lot of effort has gone into the design of a tial spreading of particles of the cloud in the area and because
software environment that supports wireless sensors. The re- particles do not move, there might not be any particle in that
sult is a very small operating system named TinyOS, or Tiny direction; (b) particles of sufficient remaining battery power
138 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
may not be available in the right direction; (c) particles may (of sufficiently high battery) in the semicircle of center p, in
temporarily sleep (i.e., not listen to transmissions) in order the direction towards W. Note that this assumption on al-
to save battery power. ways finding a particle can be relaxed in the following ways:
(a) by repetitions of the search phase until a particle is found.
Remark. Note that any given distribution of particles in the This makes sense if at least one particle exists but was sleep-
smart dust cloud may not allow the ideal optimal number of ing during the failed searches; (b) we may consider, instead
hops to be achieved at all. In fact, the least possible number of just the semicircle, a cyclic sector defined by circles of ra-
of hops depends on the input (the positions of the grain par- diuses R R, R and also take into account the density
ticles). We have chosen, however, to compare the efficiency of the smart cloud; (c) if the protocol during a search phase
of our protocols to the ideal case. A comparison with the best ultimately fails to find a particle towards the wall, it may back-
achievable number of hops in each input case will of course track.
give better efficiency ratios for our protocols. In this analysis we do not consider the energy spent in the
search phase. Note, however, that even the case where this
5. The local target protocol (LTP) is comparable to the energy spent in actual data transmission,
the number of hops accounts for both (total energy spent is
Let d(pi , pj ) the distance (along the corresponding vertical upper bounded by a multiple of actual data transmission en-
lines towards W) of particles pi , pj and d(pi , W) the (verti- ergy).
cal) distance of pi from W. Let info(E) the information about We also assume that the position of p is uniform in the
the realization of the crucial event E to be propagated. In this arc of angle 2a around the direct line from p vertical to W.
protocol, each particle p that has received info(E) from p Each data transmission (one hop) takes constant time t (so the
(via, possibly, other particles) does the following: hops and time efficiency of our protocols coincide in this
Search phase. It uses a periodic low energy broadcast of case). We also assume that each target selection is random
a beacon in order to discover a particle nearer to W than independent of the others, in the sense that it is always drawn
itself (i.e., a particle p where d(p , W) < d(p , W)). uniformly in the arc (, ).
We are aware of the fact that the above assumptions may
Direct transmission phase. Then, p sends info(E) to p
not be very realistic in practice, however, they allows us to
via a direct line (laser) time consuming transmission.
perform a first effort towards providing some concrete analyt-
Backtrack phase. If consecutive repetitions of the search ical results.
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). Lemma 5.1. The expected hops efficiency of the local tar-
get protocol in the -uniform case is E(Ch ) / sin , for
Note that one can estimate an a-priori upper bound on the large hopt . Also 1 E(Ch ) /2 1.57, for 0 /2.
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. Proof. Due to the protocol, a sequence of points is gener-
Also note that the maximum distance d(p , p ) is R, i.e., ated, p0 = p, p1 , p2 , . . . , ph1 , ph where ph1 is a particle
the beacon transmission range (for a graphical representation within Ws range and ph is part of the wall. Let i be the
see figures 4, 5). (positive or negative) angle of pi with respect to pi1 s verti-
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
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
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.
9. Implementation aspects and details In addition, for SWP we use a decreasing clock timer
CLOCK that can be explicitly activated, deactivated and set to
We now proceed by providing a more detailed description of a given value, and two constant variables PERIODw, PERI-
the protocols implementation in our simulation environment. ODs provided by the implementer that indicate the length of
We also discuss implementation aspects of our protocols in the awake and sleeping periods of the particle. For example,
current technology wireless sensor networks. in TinyOS, to create a timer that expires every PERIODw ms
We assume that the particles are equipped with TinyOS, we use the statement Timer.start(TIMER_REPEAT,
an event driven operating system suitable for smart dust [28]. PERIODw); and Timer.stop( ); terminates the timer.
The pseudo-code presented in figures 68 demonstrates how Each time the timer expires, a Timer.fired( ) event
to implement the SWP protocol. Note that the implementa- is triggered that invokes a function implemented by the
tion of LTP is very similar. user.
At every particle, we use a Boolean variable HOLDER to Initially, the Boolean variables HOLDER and EXECUTING
denote the status of the particle. It is set to true only if the are set to false, the variable PREVIOUS is set to itself and
site holds info(E) (the information about the realization of the the set OUTf is empty. For SWP, each particle is at awake
crucial event E to be propagated). A variable PREVIOUS mode with its CLOCK set to a period chosen randomly in
records the particle from which info(E) was received. A set the range (0, PERIODw + PERIODs]. In TinyOS this can
OUTf is used to store any particle that failed to propagate a be implemented using the method Random.rand( ) of the
message towards the wall (this set is used for backtracking built-in 16-bit Linear Feedback Shift Register pseudo-random
purposes). number generator.
142 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
The protocols use five types of messages: info(E), fail, and then, after waiting for a sufficient period of time (so that
success, requestBeacon, beacon. The first message type is all neighbors respond to the request by sending a beacon),
used to propagate the actual information on the crucial event, the DetectNeighbors( ) procedure processes the tem-
while the next two (fail and success) are special control mes- porary buffer and returns a set containing those particles that
sages used to signify a failure or a success in the attempt responded to the broadcast of the search beacon, ordered by
to propagate info(E) towards the receiving wall W. The the distance of the particles (i.e., d(p , p )).
requestBeacon and beacon messages are used by the search Apart from the message event handlers, the protocols use
phase. Interestingly, in TinyOS, radio communication follows two additional types of events: (i) SensorEvent created
the Active Message (AM) model, in which each packet on the by the sensors of the particle when a crucial event is real-
network specifies a handler ID that will be invoked on recip- ized (i.e., when the particle is selected by the adversary A)
ient nodes. When a message is received, the receive event and (ii) Timer.fired( ) created by the Timer when the
associated with that handler ID is signaled. Thus we only counting has finished. Figure 7 depicts the two generic event
need to define one message handler per message type. Fig- handlers implemented by our protocols. Remark that the
ure 6 depicts the five message handlers implemented by the PowerDisable( ) and PowerEnable( ) will force the
protocols. particle to enter a snooze mode where only the Timer is ac-
Remark that the beacon message handler assumes that the tive.
communication module is capable of measuring the signal Based on the above event driven functionality, particle p
strength of the message received by executing the function executes continuously the Main task, shown in figure 8.
power(msg). Similarly, the function sender(msg) is TinyOS provides a two-level scheduling hierarchy consisting
used to extract the originator of a message, assuming that of tasks and hardware event handlers. Tasks are used to per-
the message structure maintains such kind of information. form longer processing operations, such as background data
Also, the function remember(. . .) adds the information processing, and can be preempted by hardware event handler.
to a temporary buffer. Using these primitives by sending a Remark that the post operation places the task on an internal
requestBeacon message the particle initiates the search phase 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 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 angles, = {45, 60, 90} (in degrees). The reported experi-
fast to zero. This can be justified by taking into account the ments for the three different performance measures we con-
average degree of each particle for various density d. sidered are illustrated in figures 1416.
In the third set of experiments, we evaluate the perfor- Examining figure 14, that shows the probability of suc-
mance of the SWP protocol in the case when all the particles cess (P{success}), we first observe that for particle density
remain awake (i.e., en = 0). We consider this a first step to in- d < 0.05 (i.e., throwing a small number of particles) the
vestigate (a) the impact of the angle and (b) the effect of the protocol fails to propagate the critical event (i.e., the suc-
particles density d on the probability of success (P{success}), cess probability is zero). However, the probability of suc-
hops efficiency (Ch ) and average number of backtracks. The cess increases very fast exhibiting a threshold-like behaviour
particle density was 0.01 d 0.3 and used three different 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. the particle density crosses d = 0.05 (i.e., when P{success} >
As expected (due to equation (3)), setting a smaller angle 0) the hops efficiency gets close to 2.6. In fact, when d = 0.1,
reduces the probability of success in each density case Ch = 1.74 while no further gain is achieved if we throw more
since the number of particles that respond to the search phase particles (i.e., increase d). This is because of a sufficiently
gets smaller. So for = 60 the P{success} gets close to 1 large density leads to many particles found in the search, of
when d > 0.2 while for = 45, P{success} 1 when which particles already some are close to the vertical line.
d > 0.25. Similar results hold for = 60 and = 45, although at high
Regarding the average hops efficiency (Ch ) we interest- particle densities, the hops efficiency is slightly better.
ingly observe in figure 15 that even for a small particle den- Finally, in figure 16 we can see the average number of
sity, the hops efficiency is close to the optimal. Actually, as 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 The last set of experiments aims to evaluate the impact
density (i.e., d 0.1) the protocols almost always uses the of the energy saving specification en on the performance of
backtrack mechanism, while when d 1.5 the number of the protocol. Again, we measure the probability of success
backtracks performed drops very fast to zero. Furthermore, (P{success}), hops efficiency (Ch ) and average number of
we observe that the number of backtracks is initially high backtracks over energy saving specification (en), for three dif-
but decreases with a fast rate as the particle density increases. ferent particle densities (d = 0.15, 0.2, 0.25) and three differ-
This can be justified by taking into account the average degree ent angles = {45, 60, 90} (in degrees). We have set the
of each particle for various density d. More specifically, when awake period w = 2 and the sleeping period s [0, 15]
= 90 and d = 0.15 the protocol almost always succeeds in thus making en [0, 0.88] (recall that en = s/(s + w)).
propagating the crucial event without the need to backtrack. 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 ficiency when en = 0.5: the hops efficiency remains unaf-
angles . fected when en 0.5 while it decreases very fast (i.e., Ch
In figures 17 and 20, that show the probability of suc- increases) when en > 0.5. Interestingly, figures 19 and 21
cess for different particle densities d and angles , we ob- show that for the same threshold value (en = 0.5) the pro-
serve that the protocol experiences a threshold behavior when tocol almost always succeeds without the need to backtrack,
en = 0.75: when en 0.75 the probability of success while for en > 0.5 the number of backtracks increases very
is close to 1 while for en > 0.75, P{success} drops very fast with en. Thus, although for en 0.8 the probability
fast to zero. In other words, even if we set the particles to of success is close to 1, setting the energy saving specifica-
be awake only the 25% of each sleepawake cycle, it does tion to en = 0.5 seems to be more reasonable. This leads to
not affect the success of the protocol to propagate the in- the conclusion that by setting the particles to be active only
formation to W. However, in figures 18 and 20 we ob- the 50% of the overall period for which the protocol is exe-
serve a similar threshold behavior for the average hops ef- cuted, we manage to decrease the energy requirements while
148 CHATZIGIANNAKIS, NIKOLETSEAS AND SPIRAKIS
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- [4] I. Chatzigiannakis, T. Dimitriou, S. Nikoletseas and P. Spirakis, A prob-
fected. 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
11. Conclusions and future work P. Spirakis, A comparative study of protocols for efficient data propa-
gation in smart dust networks, in: Proc. International Conference on
We presented here a model for Smart Dust and three basic Parallel and Distributed Computing, EUPOPAR 2003 (2003). Also in
protocols (and their average case performance) for local de- Parallel Processing Letters 13(4) (2003) 615627.
[6] I. Chatzigiannakis and S. Nikoletseas, A sleepawake protocol for in-
tection and propagation. We plan to investigate protocols that formation propagation in smart dust networks, in: Proc. 3nd Interna-
trade-off hops efficiency and time, as well as study the fault- tional Workshop on Mobile and Ad-Hoc Networks, 17th Annual In-
tolerance of protocols as a function of smart dust parameters ternational Parallel and Distributed Processing Symposium (IPDPS
(such as density of the cloud, the energy saving character- 2003), Nice, France, April 2125 (2003) p. 225.
istics, etc.). We also intend to investigate alternative back- [7] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, An efficient com-
munication strategy for ad-hoc mobile networks, in: Proc. 15th Inter-
track mechanisms and study their effect on the efficiency and national Symposium on Distributed Computing (DISC 2001), pp. 285
fault-tolerance of the protocol. Also, we are currently work- 299. See also brief announcement in: Proc. 20th Annual Symposium on
ing towards the design of local protocols than can monitor Principles of Distributed Computing (PODC 2001), pp. 320322.
the spreading of a time-sequence of events (i.e., tracking pro- [8] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, On the average and
worst-case efficiency of some new distributed communication and con-
tocols). We plan to provide performance comparisons with
trol algorithms for ad-hoc mobile networks, in: Proc. 1st ACM Work-
other protocols mentioned in the related work section. Fi- shop on Principles of Mobile Computing (POMC 2001), pp. 119, in-
nally, we plan to also explicitly introduce sensor faults and vited paper.
study the performance (efficiency, fault-tolerance) of our pro- [9] I. Chatzigiannakis, S. Nikoletseas and P. Spirakis, Smart dust protocols
tocols in this case. 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
References propagation, in: Proc. of the ACM Workshop on Performance Evalua-
tion of Wireless Ad Hoc, Sensor, and Ubiquitous Networks, PE-WASUN
[1] T. Antoniou, A. Boukerche, I. Chatzigiannakis, G. Mylonas and (2004).
S. Nikoletseas, A new energy efficient and fault-tolerant protocol for [11] D. Estrin, R. Govindan, J. Heidemann and S. Kumar, Next century
data propagation in smart dust networks using varying transmission challenges: scalable coordination in sensor networks, in: Proc. 5th
range, in: Proc. 37th ACM/IEEE Annual Simulation Symposium ANSS, ACM/IEEE International Conference on Mobile Computing (MOBI-
Arlington, VA, USA, April 1822 (2004) pp. 167177. COM 1999) (1999) pp. 263270.
[2] Berkeley Wireless Research Center (2001), http://bwrc.eecs. [12] H. Euthimiou, S. Nikoletseas and J. Rolim, Energy balanced data prop-
berkeley.edu agation in wireless sensor networks, in: Proc. 4th International Work-
[3] A. Boukerche and S. Nikoletseas, Protocols for data propagation in shop on Algorithms for Wireless, Mobile, Ad-Hoc and Sensor Networks
wireless sensor networks: a survey, in: Wireless Communications Sys- (IPDPS/WMAN04) (Computer Society Press, 2004) p. 225.
tems and Networks, ed. M. Guizani (Kluwer Academic, 2004) chap- [13] W. Feller, An Introduction to Probability Theory and Its Applications
ter 2, pp. 2351. (Wiley, New York, 1957).
EFFICIENT AND ROBUST PROPAGATION PROTOCOLS FOR SMART DUST 149
[14] W.R. Heinzelman, A. Chandrakasan and H. Balakrishnan, Energy- in major international conferences. He has participated in several European
efficient communication protocol for wireless microsensor networks, Union funded R&D projects, and worked in the private sector.
in: Proc. 33rd Hawaii International Conference on System Sciences E-mail: ichatz@cti.gr
(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- Sotiris E. Nikoletseas is currently a Lecturer Pro-
BICOM 1999) (1999) pp. 174185. fessor at the Computer Engineering and Informatics
[16] S.E.A. Hollar, COTS dust, M.Sc. Thesis in engineeringmechanical en- Department of Patras University, Greece and also a
gineering, University of California, Berkeley, USA (2000). Senior Researcher and Director of Research Unit 1
[17] C. Intanagonwiwat, D. Estrin, R. Govindan and J. Heidemann, Impact (Foundations of Computer Science, Relevant Tech-
of network density on data aggregation in wireless sensor networks, nologies and Applications) at the Computer Tech-
Technical Report 01-750, University of Southern California Computer nology Institute (CTI), Greece. His research in-
Science Department (November, 2001). terests include probabilistic techniques and random
[18] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion: graphs, average case analysis of graph algorithms
A scalable and robust communication paradigm for sensor networks, and randomized algorithms, algorithmic applications
in: Proc. 6th ACM/IEEE International Conference on Mobile Comput- of probabilistic techniques in distributed computing (focus on ad-hoc mobile
ing (MOBICOM 2000) (2000) pp. 5667. networks and wireless sensor networks), algorithmic applications of combi-
[19] B. Karp, Geographic routing for wireless networks, Ph.D. Dissertation, natorial and probabilistic techniques in fundamental aspects of modern net-
Harvard University, Cambridge, MA (2000). works (focus on network reliability and stability), approximation algorithms
[20] -adaptive multi-domain power aware sensors (April 2001), for computationally hard problems. He has published over 80 scientific ar-
http://www-mtl.mit.edu/research/icsystems/uamps ticles in major international conferences and journals and has co-authored
[21] A. Manjeshwar and D.P. Agrawal, TEEN: A routing protocol for en- a book on probabilistic techniques, a chapter in the Handbook of Random-
hanced efficiency in wireless sensor networks, in: Proc. 2nd Interna- ized Computing (Kluwer Academic) and several chapters in books of inter-
tional Workshop on Parallel and Distributed Computing Issues in Wire- national circulation in topics related to distributed computing. He has been
less Networks and Mobile Computing (PDCIWNMC02), 16th Annual invited speaker in international scientific events and Universities and he has
International Parallel and Distributed Processing Symposium (IPDPS delivered several tutorials and keynote talks. He has been a reviewer for
2002), Fort Lauderdale, FL, USA, April 1519, 2002 (2004) p. 1956. important computer science journals and has served in the Program and Or-
[22] K. Mehlhorn and S. Nher, LEDA: A Platform for Combinatorial and ganizing Committees of International Conferences and Workshops. He has
Geometric Computing (Cambridge University Press, 1999). participated in many European Union funded R&D projects.
[23] S. Nikoletseas, I. Chatzigiannakis, H. Euthimiou, A. Kinalis, A. Anto- E-mail: nikole@cti.gr
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). Paul G. Spirakis born in 1955, got his Ph.D. from
[24] C.E. Perkins, Ad Hoc Networking (Addison-Wesley, Boston, MA, Harvard University in 1982. He became an Assistant
2001). (Reference to book spans across different chapters since the text Professor at NYU (the Courant Institute) the same
is discussing the case of porting existing mobile ad-hoc routing proto- year. He was then elected as a tenured Associate
cols to smart dust networks.) Professor at Patras University, Greece, in the De-
[25] S.M. Ross, Stochastic Processes, 2nd ed. (Wiley, 1995). partment of Computer Engineering and Informatics.
[26] C. Schurgers, V. Tsiatsis and M.B. Srivastava, STEM: Topology He became a Full Professor in the same Department
management for energy efficient sensor networks, in: Proc. IEEE in 1990. He served as the Chairman of the Depart-
Aerospace Conference 2002, Vol. 3 (2002) pp. 10991108. ment of Computer Engineering and Informatics for
[27] S. Singh and C.S. Raghavendra, PAMAS: Power aware multi-access six years. Since 1996 he is the Director of the Re-
protocol with signalling for ad hoc networks, ACM Computer Com- search and Academic Computer Technology Institute (RACTI) of Greece.
munication Review 28(3) (1998) 525. His research interests include algorithms, probabilistic techniques, distrib-
[28] TinyOS, A component-based os for the network sensor regime (Octo- uted and parallel computing and average case analysis. Recently P. Spirakis
ber, 2002), http://webs.cs.berkeley.edu/tos/ is active in the area of algorithmic aspects of game theory and also in algo-
[29] Wireless Integrated Sensor Networks (April 2001), http://www. rithmic aspects of ad-hoc and sensor networks. Paul Spirakis has won the top
janet.ucla.edu/WINS/ prize of the Greek Mathematical Society in 1973 as a student. Since then he
[30] W. Ye, J. Heidemann and D. Estrin, An energy-efficient MAC protocol won several awards and many competitive grants. He was appointed a dis-
for wireless sensor networks, in: Proc. 12th IEEE International Con- tinguished Visiting Researcher of Max Planck Informatik. He served as the
ference on Computer Networks (INFOCOM 2002) (2002) pp. 947957. 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
Ioannis Chatzigiannakis is a Researcher of Re- Vice Chairs of the European Association for Theoretical Computer Science
search Unit 1 (Foundations of Computer Science, (EATCS). He is a high level consultant of the Greek State and Industry in
Relevant Technologies and Applications) at the Informatics. Paul Spirakis has published extensively in many journals and
Computer Technology Institute (CTI), Greece. He conferences of his field (more than 150 publications currently). The journals
has received his B.Eng. degree from the University and conferences where his work appears are among the most competitive
of Kent, UK in 1997 and his Ph.D. degree from the worldwide. He coauthored two books with Cambridge University Press and
Computer Engineering and Informatics Department seven books in Greek. Paul Spirakis serves as an Editor in many scientific
of Patras University, Greece in 2003, under the su- journals of computer science and usually in the scientific commitees of some
pervision of Prof. Paul Spirakis. His research in- of the most prestigeous computer science conferences. He also serves peri-
terests include distributed computing, mobile com- odically as a high level research evaluator for the EU and Greece.
puting and algorithmic engineering. He has served as an external reviewer E-mail: spirakis@cti.gr
Mobile Networks and Applications 10, 151168, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
sensor network. In such cases the task of harvesting the infor- training protocol secure by using a parameterized variant
mation produced by the sensor network and that of providing of frequency hopping.
an interface to the outside world may be performed by aircraft Next, we show that in a trained wireless sensor network
and/or helicopters over-flying the sensor network, or by laser routing and data fusion can be performed by very simple
transmission to a satellite constellation. In this latter case, and energy-efficient protocols.
the bulk of the inter-sensor communications is by radio, since
Finally, we show how to design the coordinate system such
such communications are point to multi-point, while special-
as to minimize the power expended in collecting and rout-
ized sensors acting as local sinks communicate with the satel-
ing data.
lite constellation using laser beams, for example.
The remainder of this paper is organized as follows. Sec-
1.1. Securing wireless sensor networks tion 2 discusses the sensor node model used throughout the
work. In particular, it discusses genetic material with which
It is anticipated that in most application domains, sensor net- sensor nodes are endowed prior to deployment and which will
works will constitute an information source that is a mission- be key in securing sensor networks. Section 3 discusses wire-
critical system component and will, thus, require commensu- less sensor networks, as a conglomerate of individual sensor
rate security protection. If an adversary can thwart the work nodes that have to self-organize and self-govern. In particu-
of the network by perturbing the information produced, stop- lar, therein we discuss interfacing sensor networks with the
ping production, or pilfering information, then the usefulness outside world, a work model for sensor networks, as well as a
of sensor networks will be drastically curtailed. Thus, secu- brief preview of the training process. Next, section 4 proposes
rity is a major issue that must be resolved in order for the routing and data fusion algorithms in a trained sensor net-
potential of wireless sensor networks to be fully exploited. work. Section 5 is the backbone of the entire paper, presenting
The task of securing wireless sensor networks is compli- the theoretical underpinnings of the training process. We note
cated by the fact that the sensors are mass-produced anony- that within this section we discuss the details of a lightweight
mous devices with a severely limited energy budget, and, synchronization protocol for sensor networks. Section 6 dis-
initially no knowledge of their location in the deployment cusses the longevity of sensor networks in terms of a number
environment. Security must be provided even though sensor of system parameters. Section 7 takes a look at the problem
nodes are unattended and vulnerable to a vast array of attacks of evaluating the energy expenditure per sensor node. Sec-
[3,4,10,22,30,42]. tion 8 shows how to choose the coronas in such a way that the
Wireless sensor networks are sufficiently different from energy expended for conveying the results to the sink node is
ad-hoc networks that security solutions designed specifically minimized. Finally, section 9 offers concluding remarks and
for the former do not apply to the latter [10,32,42]. Indeed, maps out areas for future investigations.
in was recently noted that the ultra-lightweight protocols im-
posed by the stringent energy limitations may leave not much
2. The sensor node model
room for advanced encryption schemes. Consequently, pro-
tection against overhearing in military applications and pri- We assume a sensor node to be a device that possesses three
vacy protection in personal systems needs to be inherently basic capabilities: sensory, computation, and wireless com-
built into the concepts underlying sensor network models and munication as illustrated in figure 1. The sensory capability is
protocols from the beginning. Reliability is expected to be a necessary to acquire data from the environment; the computa-
result of the large number of sensors deployed for a specific tional capability is necessary for aggregating data, processing
task. However, this can only be obtained if defective sensors control information, and managing both sensory and commu-
can be excluded from the communication, and the sensors are nication activity. Finally, the wireless communication capa-
calibrated either individually or collectively, either before bility is necessary for sending (receiving) aggregated data and
deployment or continuously in their environment. control information to (from) other sensors or the sink.
We assume that individual sensor nodes operate subject to
1.2. Our contributions following fundamental constraints.
We view our main contribution at several levels: (a) Sensor nodes are anonymous they do not have fabrica-
tion-time identities.
First we propose a virtual infrastructure a dynamic co-
(b) Sensor nodes are tiny, commodity devices that are mass-
ordinate system for a massively deployed collection of
produced in an environment where testing is a luxury.
anonymous sensor nodes. This coordinate system yields,
at no extra cost, a clustering scheme: two nodes are in the (c) Each sensor has a non-renewable energy budget; when
same cluster only if they have the same coordinates. the on-board power supply is exhausted, the sensor node
We then go on to show that training the sensor nodes the is expired.
process through which nodes learn their coordinates can (d) In order to save energy, each sensor node is in sleep mode
be performed by a protocol that is at the same time light- most of the time, waking up at random points in time for
weight and secure. Indeed, we outline a way of making the short intervals under the control of an internal timer.
TRAINING A WIRELESS SENSOR NETWORK 153
(e) Each sensor has a modest transmission range, perhaps a Classical frequency hopping mechanisms have been used
few meters. This implies that out-bound messages sent as a means of combating jamming both hostile and non-
by a sensor can reach only the sensors in its proximity, hostile and of implementing frequency diversity and interfer-
typically a small fraction of the sensors deployed. ence averaging in a non-hostile context [13,45]. Typically
these mechanisms offered little cryptographic value. Cryp-
(f) Individual sensor nodes must work unattended once de-
tographic techniques such as encryption, on the other hand,
ployed it is either infeasible or impractical to devote at-
are customarily used to address security problems in all but
tention to individual sensor nodes.
the physical layer in the network. The key idea behind our
At any point in time, a sensor, will be engaged in perform- proposed security solution is that by extending classical fre-
ing one of a finite set of possible operations, or will be asleep. quency hopping techniques using symmetric key cryptogra-
Three basic operations are sensing (to collect raw measure- phy, security problems in the physical layer, as well as in
ments), data fusion and/or aggregation (to derive target data other layers in the network can be uniformly addressed in a
from raw measurements), routing (to communicate raw mea- unified framework; we call this framework randomized fre-
surements, target data, and control data). We assume each quency hopping.
operation performed by a sensor consumes a known fixed We are now in a position to show how the genetic material
amount of energy and that a sleeping sensor performs no op- is used in support of secure communications in a sensor net-
eration and consumes, essentially, no energy. work. For this purpose, it is useful to imagine three sequences
It is worth mentioning that while the energy budget can of random numbers as follows:
supply short-term applications, sensors dedicated to work
over years may need to scavenge energy from the specific en-
an infinite sequence of t1 , t2 , . . . , ti , . . . of time epoch
vironment they are placed into, employing light, temperature,
lengths;
vibration, kinetics, magnetic fields, etc.
an infinite sequence n1 , n2 , . . . , ni , . . . of frequency sets
2.1. Genetic material drawn from a large universe, e.g., the ISM band;
for every ni (i 1), an infinite permutation f1i , f2i , . . . of
The nodes genetic material plays a key role in driving the frequencies from ni .
functionality of different node protocols. To illustrate this, we
consider protocols implementing our proposed security solu- Importantly, these sequences can be generated locally by each
tion for the sensor network. sensor node using the injected genetic material and, therefore,
We assume that at pre-deployment time the sensor nodes do not need to be communicated after deployment.
are injected, in a secure environment, with the following ge- We assume that time is ruled into epochs. During the ith
netic material: time epoch, of length ti , a frequency set ni will be used sub-
a standard pseudo-random number generator (one of ject to a hopping pattern described by the hopping sequence
public-domain algorithms available); f1i , f2i , . . . . Thus, as long as a sensor node is synchronous
a set of secret seeds to be used as parameters for the ran- to the sink, it knows the current time epoch, the offset into
dom number generator; the epoch, the set of frequencies in use during the epoch, as
well as the hopping pattern in force during the epoch. To
an initial time (at this point all the sensor nodes are syn- an outside observer, however, successive epoch lengths, hop-
chronous to the sink node).
ping sets, and hopping patterns appear as the product of an
It is important to note that immediately after deployment unknown random process. Given that techniques are known
all the clocks are synchronous. In time, however, clocks will to discover a hopping sequence by monitoring transmissions,
drift and periodic re-synchronization becomes necessary. For security can only be provided if the design modifies the hop-
reasons of simplicity, we assume that synchronization is al- ping sequence in less time than is required to discover the
ways done to the master clock running at the sink. As we will sequence. The choice of frequency hopping parameters de-
show in detail later in this section, one of our main contribu- termines the time required to discover the sequence (the mag-
tions is a light-weight re-synchronization protocol. nitude of the challenge to an adversary).
154 WADAA ET AL.
2.2. Making sensor nodes tamper-resistant the sensor network. Note also that tampering attempts that in-
volve the removal of several sensor nodes simultaneously will
The most obvious tamper resistance strategies are hardware- also be defeated since some node in the set of removed nodes
based and involve special hardware circuits within the sen- will notice changes in its NSA and can alert the others.
sor node to protect sensitive data, special coatings or tamper
seals. However, hardware solutions to the tampering prob- 3. Structure and organization of a wireless sensor
lem require extra circuitry that increases the cost and hard- network
ware complexity of sensor nodes. Worse yet, the additional
We envision a massive deployment of sensor nodes, perhaps
hardware is very likely to consume valuable energy, already
in the thousands or even tens of thousands. The sensor nodes
in short supply. Also special coatings and seals may offer
are aggregated into sophisticated computational and com-
protection against some but, certainly, not all tampering at- munication infrastructures, called wireless sensor networks,
tempts. Indeed, it is assumed that a sufficiently capable adver- whose goal is to produce globally meaningful information
sary can extract confidential information, thus compromising from data collected by individual sensor nodes. However,
the sensor node. Thus, not surprisingly, tamper resistance or the massive deployment of sensors nodes in a sensor net-
tamper protection is not found in present-day sensor nodes work, combined with anonymity of individual sensors, lim-
[3,4,10]. Since wireless sensor networks must function unat- ited power budget and in many applications a hostile envi-
tended, the potential for physical tampering attacks is signif- ronment, pose daunting challenges to the design of protocols
icant. It is worth noting that while pre-deployment tamper for sensor networks. For one thing, the limited power bud-
detection may be worthwhile, post-deployment tamper de- get at the individual sensor node level mandates the design of
tection is of little use in wireless sensor networks since, in ultra-lightweight communication protocols. Likewise, issues
the vast majority of applications, inspecting individual sensor concerning how the data collected by individual sensor nodes
nodes is not an option. Also, physical tampering may com- could be queried and accessed and how concurrent sensing
promise only the node attacked (ideal), the immediate neigh- tasks could be executed internally are of particular signifi-
borhood of the node attacked, or the entire network. To cope cance. An important guideline in this direction is to perform
with these conditions, our solution subscribes to the notion of as much local data processing at the sensor level as possible,
self-guarding in that each sensor node should be able to de- avoiding the transmission of raw data through the sensor net-
tect, independently, physical tampering and should react such work. Indeed, it is known that it costs 3 J of energy to transmit
that the impact of the attack is minimal. Our solution to en- 1 Kb of data a distance of 100 meters. Using the same amount
dow individual sensor nodes with tamper resistance does not of energy, a general-purpose processor with the modest speci-
require additional or more sophisticated hardware. fication of 100 million instructions/Watt executes 300 million
In order to set the stage for discussing our solution, we note instructions [31,37].
that the tampering threat model assumes that the adversary is As a consequence, the sensor network must be multi-hop
and only a limited number of the sensor nodes count the sink
either trying to force open an individual sensor node in-
among their one-hop neighbors. For reasons of scalability,
situ; or
it is assumed that no sensor node knows the topology of the
is physically removing the sensor node from the deploy- network.
ment area. Our work focuses on the design of ultra-lightweight orga-
nization and communication protocols for a class of wireless
We guard against the first threat by blanking out the mem-
sensor networks consisting of a single sink node and a large
ory, triggered by a simple switch. We guard against the sec-
number of sensors nodes randomly deployed in the transmis-
ond threat by relying on local data that the sensor can collect,
sion range of the sink.
thus establishing a unique signature of its neighborhood that A basic management problem in wireless sensor networks
is difficult to forge. To be more specific, immediately after is to balance the utility of the activity in the network against
deployment each sensor transmits, during its wake time, on the cost incurred by the network resources to perform this
a specified sets of frequencies, using a frequency hopping activity. The scarce resource in the network that is of primary
sequence established prior to deployment. This allows in- concern is energy.
dividual sensor nodes to collect an array of signal strengths
from the sensors in their locale. It is important to recall that 3.1. Interfacing sensor networks
sensors do not have identities and that, consequently, the ar- We assume that the sensor network is connected to the outside
ray of signal strengths is the only data available to the sensor world (e.g., point of command and control, the Internet, etc.)
node. This array, establishes, in the obvious way, a signature through a gateway node. The gateway node may or may not
of the neighborhood of the node. For this reason the array be collocated with the sensor nodes in the deployment area.
will be referred to as the nodes neighborhood signature array Referring to figure 2, we note that the interface with the out-
(NSA, for short). If the node is removed from the area of de- side world may be achieved by a helicopter or aircraft over-
ployment, it will notice changes in the signals received when flying the sensor network, and collecting information from a
compared to its NSA and erase its own memory to prevent the select group of reporting nodes. In such scenarios communi-
tampering agent from gaining access to information secret to cation between individual sensor nodes is by radio, while the
TRAINING A WIRELESS SENSOR NETWORK 155
3.3. Training a wireless sensor network coordinate system onto the sensor network. An interesting
by-product of our training protocol is that it provides a par-
It was recognized that some applications require sensory data titioning into clusters and a structured topology with natural
with some location awareness, encouraging the development communication paths. The resulting topology will make it
of communication protocols that are location aware and per- simple to avoid collisions between transmissions of nodes in
haps location dependent. The practical deployment of many different clusters, between different paths and also between
sensor networks will result in sensors initially unaware of nodes on the same path. This is in contrast with the major-
their location: they must be trained in this vital information. ity of papers that assume routing along spanning trees with
Further, due to limitations in form factor, cost per unit and frequent collisions.
energy budget, individual sensor nodes are not expected to be Clustering was proposed in large-scale networks as a
GPS-enabled. Moreover, many probable application environ- means of achieving scalability through a hierarchical ap-
ments limit satellite access. proach. For example, at the medium access layer, clustering
The localization problem is for individual sensor nodes to helps increase system capacity by promoting the spatial reuse
determine, as closely, as possible their geographic coordinates of the wireless channel; at the network layer, clustering helps
in the area of deployment. Prominent solutions to the localiza- reducing the size of routing tables and striking a balance be-
tion problem are based on multilateration [79,12,16,29,33]. tween reactive and proactive routing. It is intuitively clear that
Most of these solutions assume the existence of several an- wireless sensor networks benefit a great deal from clustering;
chor nodes that are aware of their location (perhaps by endow- indeed, separating concerns about inter-cluster management
ing them with a GPS-like devices). Sensor nodes receiving and the intra-cluster management can substantially decrease,
location messages from at least three sources can approxi- and load balance the management overhead. Given the im-
mate their own locations. For a good survey of localization portance of clustering, a number of clustering protocols for
protocols for wireless sensor networks we refer to [25]. wireless sensor networks have been proposed in the recent
In some other applications, exact geographic location is literature [5,11,15]. However, virtually all clustering proto-
not necessary: all that the individual sensor node need is cols for wireless sensor networks assume tacitly or explicitly
coarse-grain location awareness. There is an obvious trade- that individual sensor nodes have identities.
off: coarse-grain location awareness is lightweight but the As it turns out, our clustering protocol has the following
resulting accuracy is only a rough approximation of the ex- desirable features:
act geographic coordinates. Figure 5 illustrates a possible
way of inducing such a coarse-grain location awareness by lightweight as a by-product of training;
an overflying aircraft or helicopter. All that the individual
organizes anonymous asynchronous nodes;
sensor nodes need is to determine their approximate distance
to three different positions of the training agent. We omit the a cluster is the locus of all nodes having the same coordi-
details. nates; and
Our approach is different: we obtain this coarse-grain individual nodes need not know the identity of other nodes
location awareness by the training protocol that imposes a in their cluster.
TRAINING A WIRELESS SENSOR NETWORK 157
is completely eliminated. Thus, our training protocol implies vidual data values are being transmitted to the sensor whose
an efficient MAC protocol as well. identity is 1 who will perform data fusion and will send the
final result to the sink node as discussed in section 7. The
4.2. Data fusion advantage of this data fusion scheme is that there is no data
loss and all the collected values will be correctly fused. There
Once sensory data was collected by a multitude of sensor are, however, many disadvantages. For one thing, the initial-
nodes, the next important task is to consolidate the data in ization algorithm of [26] requires every sensor in D to expend
order to minimize the amount of traffic to the sink node. We an amount of energy proportional with log |D|. For another,
place the presentation in the context of our work model. To be the final result of the data collection is concentrated in a sin-
more specific, we assume that the cluster identified by (i, j ) gle sensor (i.e., the sensor with temporary identity 1), who is
that is, the set of sensor nodes located in sector Ai,j , where a single point of failure.
i is the corona identifier, and j is the wedge identifier, are We now propose a much simpler data fusion scheme that
to perform a certain task T . A number of sensors in sec- involves some data loss but that is fault tolerant and does not
tors A1,j , A2,j , . . . , Ai1,j are selected to act as routers of require the sensors in D to have unique identities. The idea
the data collected by the sensors in Ai,j to the sink. Collec- is that the sensors in D transmit the data collected bit by bit
tively, these sensors are the support sensors of task T . starting, say, left to right as follows: a value of 0 is not trans-
It is, perhaps, of interest to describe the process by which mitted, while a 1 will be transmitted. The sensors in Ai1,j
the sensors associated with T are selected. To begin, dur- that have been elected as routers in support of transaction T
ing a time interval of length the sink will issue a call for pick up the values transmitted. The following disambiguation
work specifying the identity j of the wedge in which the task scheme is used:
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 No bit is received in this case a 0 is recorded;
that happen to wake up during the interval and that have A bit of 1 is received in this case a 1 is recorded;
an appropriate energy level stay awake and will participate in A collision is recorded in this case a 1 is recorded.
the task either as either data collectors or as routers depending
on their respective position within the wedge. It is intuitively It is clear that as a result of this disambiguation scheme,
clear that by knowing the number of sensors, the density of every sensor in Ai1,j that is in support of T stores the logical
deployment and the expected value of sleep periods, one can OR of the values stored by sensors in D. Note also that while
fine-tune in such a way that a suitable number of routers there was loss of information in the process of fusing data,
will be awake in wedge j in support of T . Likewise, we can no further loss can occur in traversing the path from Ai1,j
select the set D of data collecting sensors in Ai,j . Let S de- to the sink: this is because all routers in Ai1,j transmit the
note the set of support sensors for T . It is appropriate to recall same bit string.
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- 4.3. An example
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 For an example of data fusion consider a sensor network that
time epoch, and a hopping sequence to be used within each is tasked to monitor and report the temperature in cluster Ai,j .
epoch. For details we refer the reader to the description of the Referring to table 1, for the application at hand temperatures
randomized frequency hopping security framework proposed below 111 F are considered to be non-critical and if such a
in section 2. temperature is reported no specific action is to be taken. By
Assume that the results of the data collection specific to contrast, temperatures above 111 F are considered to be criti-
task T can be partitioned into 2m (m 0), disjoint groups. cal and they trigger a further monitoring action. The encoding
Thus, each sensor performing data collection will encode its featured in table 1 is specifically designed to reflects the rela-
data in a string of m bits. tive importance of various temperature ranges. For example,
Since, typically, D contains a large number of sensors, it the temperature ranges in the non-critical zone are twice as
is important to fuse individual results into a final result that large as those in the critical zone. Also, notice that the left-
will be sent to the sink node. We now outline two possible most bit differentiates critical from non-critical temperatures.
solutions to the data fusion problem. Using the algorithm of Thus, if the sink nodes receives a reported temperature whose
Nakano and Olariu [26] that does not require sensors to have leftmost bit is a 1, then further action is initiated; if, on the
identities, the sensors in D acquire temporary identities rang- other hand, the leftmost bit is 0, then no special action is nec-
ing from 1 to |D|. Using their newly acquired identities, indi- essary.
Table 1
Illustrating temperature ranges and their encoding.
Temp 5160 6170 7180 8190 91100 101110 111115 116120 121125 126130 131135 136140 141145 146150
Code 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
TRAINING A WIRELESS SENSOR NETWORK 159
2. Sensors that have a 1 in position d transmit in the first time It is natural to assume that, just prior to deployment, the sen-
slot and sleep in the second. sor nodes are synchronized. However, due to natural clock
160 WADAA ET AL.
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 161
Proof. The proof is by induction on the depth i of node u in With these technicalities out of the way, we now return to
T . To settle the basis, note that for i = 1, u must be the root the corona training protocol. In our setting, the preorder and
and p(u) = 1, as expected. inorder numbers of internal nodes in T correspond, respec-
For the inductive step, assume the statement true for all tively, to time slots in the training protocol and to the trans-
nodes in T of depth less that u. Indeed, let v be the parent of mission ranges used by the sink. More precisely, consider an
u and consider the unique path of length i 1 joining the root arbitrary integer i (2 i log k 1), and assume that at
to u. Clearly, nodes u and v share b1 , b2 , . . . , bi2 and, thus, the end of time slot s a sensor node has learned the leftmost
c1 , c2 , . . . , ci2 . By the inductive hypothesis, i 1 bits b1 , b2 , . . . , bi1 . The following important result is
implied by lemma 5.1 and lemma 5.2.
i2
p(v) = 1 + cj . (2)
Corollary 5.3. Having learned bits b1 , b2 , . . . , bi1 a sensor
j =1
node must wake up in time slot z = 1 + i1 j =1 cj to learn bit
On the other hand, since v is the parent of u, we can write bi . Moreover in time slot z the sink node uses a transmission
1 if u is the left child of v,
range of rinorder(z) .
p(u) = p(v) + k (3)
otherwise. (To illustrate corollary 5.3, refer again to figure 9 where the
2i1 internal nodes are labeled by their preorder numbers. Con-
Notice that if u is the left child of v we have bi1 = 0 and sider the node labeled 2. It is easy to verify that its inorder
ci1 = 1; otherwise bi1 = 1 and ci1 = k/2i1 . This number is 4. Thus, all the nodes in the subtree rooted at 2 will
observation, along with (2) and (3) combined, allows us to be awake in slot 2 and the sink node will transmit with a range
write of r4 . Consequently, the sensor nodes at a distance from the
sink not exceeding r4 will receive the signal, while the others
i2
i1
p(u) = 1 + cj + ci1 = 1 + cj will not.)
j =1 j =1
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-
completing the proof of the lemma. ers will sleep minimizing the power expenditure. Yet another
interesting feature of the training protocol we just described
For further reference we also need the following technical is that individual sensor nodes sleep for as many contiguous
result. slots as possible before waking up, thus avoiding repeated
wakesleep transitions that are known to waste energy.
Lemma 5.2. Let u be an arbitrary node of the tree T and let Securing the training protocol is especially important since
n(u) denote its inorder number in T . Let m be the left-to- training is a prerequisite for subsequent network operations.
right rank among the leaves of T of the rightmost leaf of the As argued in [22] our parameterized frequency hopping
left subtree of T rooted at u. Then, n(u) = m. scheme guarantees that the physical layer of wireless com-
munications is secure.
Proof. We proceed by induction on the inorder number of At the same time, in case the corona training process has
a node in T . Indeed, if n(u) = 1, then u must be the left- to be aborted before it is complete, corollary 5.3 guarantees
most leaf in T and, thus, its left subtree in T consists of the that if the training process re-starts at some later point, every
leftmost leaf of T t, settling the base case. sensor node knows the exact time slots when it has to wake
Assume that the statement true for all nodes of T with up in order to learn its missing bits.
inorder number smaller than that of u. we shall distinguish
between the following two cases.
6. Reasoning about the longevity of the sensor network
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 The main goal of this section is to explore the energy require-
right subtree of T (v). Let q be the left-to-right rank among ments of the sensor network in terms of a model of work. In-
the leaves of T of the rightmost leaf of the left subtree of deed, we adopt a transaction-based model whereby the sensor
T (v). By the inductive hypothesis n(v) = q. Since u is a network is subjected to a set T of transactions. Each transac-
leaf in T it has exactly two children in T , namely the leaves tion involves the nodes in a sector (i.e., a cluster) and involves
of ranks q +1 and q +2. Thus, in this case, n(u) = n(v)+1 = performing local sensing by the sensors, data fusion and send-
q + 1, as claimed. ing the resulting information to the sink. Recall that, as dis-
Case 2. u is an ancestor of v in T . Let T (u) be the subtree cussed in section 4.1, one of the key benefits of our training
of T rooted at u. In this case v must be the rightmost leaf in is that transmitting the result of the transaction from a sector
the left subtree of T (u). Assume that n(v) = r. Observe that to the sink node amounts to routing the information along a
v has exactly two leaf children T . By the induction hypothesis path lying within the same wedge (see also figure 7). Thus,
these children have ranks r and r + 1. Thus, in this case, we associate each transaction with such a path. We will now
n(u) = n(v) + 1 = r + 1, as claimed. analyze the energy expended by sensor nodes to fulfill their
This completes the proof of lemma 5.2. path-related duties.
162 WADAA ET AL.
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
= . (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
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
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 1p<qi 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 1p<qi
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 1p<qi
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
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 .
1p<qi p=1 p=1 p=1 E1 > E2 > > Ek .
TRAINING A WIRELESS SENSOR NETWORK 165
Table 3
sensor nodes do not know the identity of the other nodes in the
Illustrating various energy ratios.
same cluster. Our second contribution was to show that train-
Corona Energy ratio ing the sensor nodes the process of learning their coordi-
2 0.3333 . . . nates can be performed by a protocol that is at the same time
3 0.2 lightweight and secure. Being energy efficient, this training
4 0.1428 . . . can be repeated on either a scheduled or ad-hoc basis to pro-
5 0.1111 . . .
vide robustness and dynamic reorganization. We also showed
6 0.0909 . . .
7 0.0769 . . . 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.
Figure 11. Illustrating the energy ratio Ei /E1 . In spite of these encouraging results, more work has to be
done. We are currently looking at security-related problem
It is instructive to get a feel for how fast the energy expen- in wireless sensor networks. This is an extremely important
ditures drop as we move away from the sink node. For this problem as the information provided by the sensor network
purpose, we rewrite equations (10) and (15) as may be used for decision making in military or civilian en-
vironments where human life is at stake. Looking at various
2
In this work we have proposed a virtual infrastructure a dy- A.1. Call for training protocol
namic coordinate system for a massively-deployed collec-
tion of anonymous sensor nodes. This coordinate system This protocol enables the sink node to issue calls for training
provides, at no extra cost, an interesting clustering scheme to a random subset of nodes of an arbitrary size post deploy-
according to which two nodes are in the same cluster only if ment. Each call communicates the next rendezvous time for
they have the same coordinates. Notice that this clustering corona and wedge training along with the local time at the
scheme works for anonymous sensor nodes. As a corollary, 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
References [26] K. Nakano and S. Olariu, Randomized initialization protocols for radio
networks, in: Handbook of Wireless Networks and Mobile Computing,
[1] J. Agre and L. Clare, An integrated architecture for cooperative sensing ed. I. Stojmenovic (Wiley, New York, 2002) pp. 195218.
networks, IEEE Computer 33(5) (2000) 106108. [27] K. Nakano and S. Olariu, Uniform leader election for radio networks,
[2] F. Akyildiz, W. Su, Y. Sankarasubramanian and E. Cayirci, Wireless IEEE Transactions on Parallel and Distributed Systems 13 (2002) 516
sensor networks: A survey, Computer Networks 38(4) (2002) 393422. 526.
[3] R. Anderson, Security Engineering: A Guide to Building Dependable [28] National Research Council, Embedded, Everywhere: A Research
Distributed Systems (Wiley, New York, 2001). Agenda for Systems of Embedded Computers, Committee on Net-
[4] R. Anderson and M. Kuhn, Tamper resistance a cautionary note, worked Systems of Embedded Computers, for the Computer Science
in: Proc. 2nd USENIX Workshop on Electronic Commerce, Berkeley, and Telecommunications Board, Division on Engineering and Physical
CA (1996), pp. 111. IEEE Wireless Communications 9(1) (2002) 40 Sciences, Washington, DC (2001).
48. [29] D. Niculescu and B. Nath, Ad-hoc positioning system, in: Proc. IEEE
[5] S. Bandyopadhyay and E. Coyle, An efficient hierarchical clustering GLOBECOM, San Antonio, TX (November 2001) pp. 29262931.
algorithm for wireless sensor networks, in: Proc. INFOCOM2003, San [30] A. Perrig, R. Szewczyk, V. Wen, D. Culler and J.D. Tygar, SPINS:
Francisco, CA (April 2003). Security protocols for sensor networks, in: Proc. ACM SIGMOBILE,
[6] D. Braginsky and D. Estrin, Rumor routing algorithm for sensor net- Rome, Italy (July 2001).
works, in: Proc. Internat. Conf. on Distributed Computing Systems [31] G.J. Pottie and W.J. Kaiser, Wireless integrated sensor networks, Com-
(ICDCS-22) (November 2001). munications of the ACM 43(5) (2000) 5158.
[7] N. Bulusu, J. Heidemann and D. Estrin, GPS-less low cost outdoor lo- [32] P. Saffo, Sensors, the next wave of innovation, Communications of the
calization for very small devices, IEEE Personal Communications 7(5) ACM 40(2) (1997) 9397.
(2000) 2834. [33] C. Savarese, J. Rabaey and K. Langendoen, Robust positioning al-
[8] N. Bulusu, J. Heidemann and D. Estrin, Scalable coordination for gorithms for distributed ad-hoc wireless sensor networks, in: Proc.
wireless sensor networks: self-configuration localization systems, in: USENIX Technical Annual Conference, Monterey, CA (June 2002)
Proc. 6th Internat. Sympos. on Communication Theory and Applica- pp. 317328.
tions (ISCTA-2001) (July 2001). [34] R.C. Shah and J. Rabaey, Energy aware routing for low energy ad hoc
[9] S. Capkun, M. Hamdi and J.-P. Hubeaux, GPS-free positioning in mo- sensor networks, in: Proc. IEEE WCNC, Orlando, FL (March 2002).
bile ad-hoc networks, Cluster Computing 5(2) (2002) 157167. [35] C.-C. Shen, C. Srisathapornphat and C. Jaikaeo, Sensor information
[10] D.W. Carman, P.S. Kruus and B.J. Matt, Constraints and approaches networking architecture and applications, IEEE Personal Communica-
for distributed sensor network security, Technical Report 00-010, NAI tions (August 2001) 5259.
Labs (2000). [36] E. Shih, S. Cho, N. Ickes, R. Min, A. Sinha, A. Wang and A. Chan-
[11] D. Coore, R. Nagpal and R. Weiss, Paradigms for structure in an amor- drakasan, A physical layer driven protocol and algorithm design for
phous computer, MIT Artificial Intelligence Laboratory Technical Re- energy-efficient wireless sensor networks, in: Proc. MOBICOM2001,
port AI-1616 (October 1997). Rome, Italy (July 2001).
[12] L. Doherty, H.S.J. Pister and L.E. Ghaoui, Convex position estimation [37] K. Sohrabi, J. Gao, V. Ailawadhi and G. Pottie, Protocols for self-
in wireless sensor networks, in: Proc. INFOCOM2001, Anchorage, organization of a wireless sensor network, IEEE Personal Communi-
AK (April 2001). cations (October 2000) 1627.
[13] A. Ephremides, J. Wieselthier and D. Baker, A design concept for re- [38] S. Tilak, N.B. Abu-Ghazaleh and W. Heinzelman, A taxonomy of wire-
liable mobile radio networks with frequency hopping signaling, Pro- less micro-sensor network models, Mobile Computing and Communi-
ceedings of the IEEE 75(1) (1987) 5673. cations Review 6(2) 2836.
[14] D. Ganesan, R. Govindan, S. Shenker and D. Estrin, Highly resilient, [39] A. Wadaa, S. Olariu, L. Wilson, K. Jones and Q. Xu, On training wire-
energy-efficient multipath routing in wireless sensor networks, ACM less sensor networks, in: Proc. 3rd International Workshop on Wireless,
Mobile Computing and Communications Review 5(4) (2001). Mobile and Ad Hoc Networks (WMAN03), Nice, France (April 2003).
[15] S. Ghiasi, A. Srivastava, X. Yang and M. Sarrafzadeh, Optimal energy- [40] B. Warneke, M. Last, B. Leibowitz and K. Pister, SmartDust: com-
aware clustering in sensor networks, Sensors 2 (2002) 258269. municating with a cubic-millimeter computer, IEEE Computer 34(1)
[16] L. Girod, V. Bychkovskiy, J. Elson and D. Estrin, Locating tiny sensors (2001) 4451.
in time and space: A case study, in: Proc. International Conference on [41] A. Woo and D.E. Culler, A transmission control scheme for media ac-
Computer Design (ICCD 2002), Freiburg, Germany (September 2002). cess in sensor networks, in: Proc. MOBICOM01 (July 2001).
[17] R.G. Graham, D.E. Knuth and O. Patashnik, Concrete Mathematics [42] A.D. Wood and J.A. Stankovic, Denial of service in sensor networks,
(Addison-Wesley, New York, 1989). IEEE Computer 35(10) (2002) 5462.
[18] http://www.darpa.mil/mto/mems/ [43] W. Ye, J. Heidemann and D. Estrin, An energy-efficient MAC protocol
[19] http://www.stanford.edu/class/ee321/ho/MEMS-14- for wireless sensor networks, in: Proc. INFOCOM 2002, New York
sensors.pdf (June 2002).
[20] http://www.xs4all.nl/ganswijk/chipdir/m/sensor. [44] V.V. Zhirnov and D.J.C. Herr, New frontiers: self-assembly and nano-
htm electronics, IEEE Computer 34(1) (2001) 3443.
[21] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed diffusion: [45] J. Zyren, T. Godfrey and D. Eaton, Does frequency hopping
A scalable and robust communication paradigm for sensor networks, enhance security? http://www.packetnexus.com/docs/
in: Proc. MOBICOM00, Boston, MA (August 2000). 20010419frequency_Hopping.pdf
[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- Ashraf Wadaa is a Lecturer of Computer Science
curity Paradigms Workshop (NSPW2003), Ascona, Switzerland (Au- and a Ph.D. candidate at Old Dominion University.
gust 2003). Wadaas dissertation is on wireless sensor networks.
[23] J.M. Kahn, R.H. Katz and K.S.J. Pister, Mobile networking for Smart His research interests include wireless sensor, and ad
Dust, in: Proc. MOBICOM99, Seattle, WA, August 1719 (1999). hoc networks, network security, database systems,
[24] J. Kulik, W. Heinzelman and H. Balakrishnan, Negotiation-based pro- and distributed computing. He published in refer-
tocols for disseminating information in wireless sensor networks, Wire- eed journals and conference proceedings. He also
less Networks 8(3) (2002). served on a number of technical committees for con-
[25] K. Langendoen and N. Reijers, Distributed localization in wireless sen- ferences, and workshops. His funding record in-
sor networks, a quantitative comparison, Manuscript (May 2003). cludes a grant from the Commonwealth Information
168 WADAA ET AL.
Security Center (CISC) in Virginia to pursue research in sensor network se- Mohamed Eltoweissy is a (visiting) Professor of
curity. Wadaa earned a B.S. (1986 with top rank) in Computer Science and Computer Science at Virginia Tech. He is also a
Automatic Control from Alexandria University, Egypt. He was nominated by Professor of Computer Science at James Madison
Old Dominion University for an outstanding teacher award in 2002. Wadaa University. Eltoweissys research interests include
is a member of the honor society of Phi Kappa Phi. information security and privacy, wireless sensor
E-mail: wadaa@cs.odu.edu 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
Stephan Olariu received the M.Sc. and Ph.D. de- served on numerous technical committees for con-
grees in computer science from McGill University, ferences, workshops, seminars, and NSF panels. He has an aggressive record
Montreal, Canada in 1983 and 1986, respectively. In of funding (over $10 million). Eltoweissy founded the Commonwealth In-
1986 he joined the Computer Science Department formation Security Center (CISC) in Virginia and was a founding member
at Old Dominion University where he is now a full of the award-winning Virginia Alliance for Secure Computing and Network-
professor. He has published extensively in various ing (VA SCAN). Eltoweissy earned a Ph.D. in computer science (1993) from
journals, book chapters, and conference proceedings. Old Dominion University, and M.S. (1989) and B.S. (1986 with top rank) in
His research interests include wireless networks and computer science and Automatic Control from Alexandria University, Egypt.
mobile computing, parallel and distributed systems, He was nominated by JMU for the Virginia State-wide Outstanding Faculty
performance evaluation, and medical image process- Awards in 2003. Eltoweissy is a member of ACM, ACM SIGSAC, and the
ing. Prof. Olariu serves on the editorial board of several archival journals honor societies of Phi Kappa Phi and Upsilon Pi Epsilon.
including IEEE Transactions on Parallel and Distributed Systems, Journal E-mail: toweissy@vt.edu
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. Kennie Jones received the B.S. degree in Biology
E-mail: olariu@cs.odu.edu 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.
Larry Wilson is an Associate Professor of Computer degree in computer science from ODU. After teach-
Science at Old Dominion University. He received ing in secondary schools for 4 years, he joined Com-
his Ph.D and M.Sci. degrees from the University of puter Sciences Corporation in 1981 as a contractor
Texas at Austin. His B.Sci. is from Midwestern Uni- to the National Aeronautics and Space Administra-
versity. Dr. Wilson is currently working in the area tion (NASA). In 1990, he became an employee of
of wireless sensor networks. Previous areas of re- NASA. In that capacity, he has made significant contributions in the areas of
search include design and analysis of parallel algo- computer graphics and data management. He is currently researching sensor
rithms, software reliability and Mikusinski Operator networks and their application to NASA missions as a member of ODUs
Functions. Sensor Network Research Team.
E-mail: wilson@cs.odu.edu E-mail: k.h.jones@nasa.gov
Mobile Networks and Applications 10, 169181, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
take longer time, or may be even unable, to detect the tion closure property. Section 4 shows several quorum sys-
arrival and departure of other hosts in its radio covered tems that satisfy this property. Section 5 presents our adap-
range. Thus, hosts may become less sensitive to neigh- tive power-saving protocol. Simulation results are presented
borhood change. Neighbor discovery is essential for route in section 6. Conclusions are drawn in section 7.
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 2. Preliminaries
neighbors on the path.
2.1. Power-saving modes in IEEE 802.11
One possible solution to the above problems is to always time-
synchronize all hosts. This approach is adopted by IEEE IEEE 802.11 supports two power modes: active and power-
802.11 under the ad hoc mode. However, 802.11 only con- saving (PS). Under the PS mode, a host can reduce its radio
siders single-hop MANETs. Time synchronization in a large- activity by only monitoring some periodical signals (such as
scale distributed environment is generally very costly. It is beacons) in the network. Tuning a host to the PS mode can
even infeasible in a mobile environment since communica- save a lot of energy. For example, table 1 summarizes the
tion delays are typically long and, worse, the MANET may power consumption of ORiNOCO IEEE 802.11b PC Gold
be temporarily partitioned at any time, making time synchro- Card [21]. However, PS mode should be used cautiously so
nization impossible. Another solution is to develop asyn- that the network throughput and delay do not get hurt.
chronous power-saving protocols. This is first investigated Under the ad hoc mode, IEEE 802.11 divides the time axis
in [25], where three solutions are proposed. Among them, into equal-length beacon intervals, each of which starts with
the quorum-based protocol is probably the most interesting an ATIM (Ad hoc Traffic Indication Map) window. The ATIM
one. It has the merit of sending the fewest beacon signals window is relatively small compared to the beacon interval.
(and is thus very energy-efficient). The central idea in the PS hosts must remain active during the ATIM window so as to
quorum-based protocol can be related to the grid quorum sys- be notified by those intending senders, and may go to doze in
tem [15]. This leads to a more general question: Can we apply the rest of the beacon interval if no one intends to send packets
other forms of quorum systems to this asynchronous power- to it. It is assumed that the ad hoc network is fully connected,
saving problem? The result can potentially bridge the impor- so time synchronization is not an issue. In the beginning of
tant quorum system concept in traditional distributed systems a beacon interval, each mobile host will contend to send a
to the area of mobile computing, which may in turn gener- beacon frame. Any successful beacon serves the purpose of
ate more efficient asynchronous power-saving protocols. This synchronizing mobile hosts clocks as well as inhibiting other
work does confirm such possibility. hosts from sending their beacons. To avoid collisions, each
In this paper, we correlate the asynchronous power-saving beacon is led by a random backoff between 0 and 2CW min 1
problem to the concept of quorum systems, which are widely slots.
used in the design of distributed systems [4,12,14,15]. A quo- After the beacon, a host with buffered packets can send
rum system is a collection of sets such that the intersection of a direct ATIM frame to each of its intended receivers in the
any two sets is always non-empty. Not all quorum systems PS mode. ATIMs are transmitted by contention in accor-
are applicable to the power-saving problem. We identify a ro- dance with the DCF (Distributed Coordination Function) ac-
tation closure property for quorum systems. It is shown that, cess procedure. A receiver, on hearing the ATIM, should re-
through our mechanism, any quorum system satisfying this ply an ACK and remain active. After the ATIM window, hosts
property can be translated to an asynchronous power-saving having neither packets to send nor packets to receive can go
protocol for MANETs. We derive a lower bound for quorum back to the PS mode to save energy. The buffered unicast
sizes for any quorum system satisfying the rotation closure packets are then sent based on the DCF access procedure af-
property. We identify a group of quorum systems that are ter the ATIM window. If the sender does not receive an ACK,
optimal or near optimal in terms of quorum sizes (the grid it should retry in the next ATIM window. If a mobile host is
quorum system [15], the torus quorum system [12], the cyclic unable to transmit its ATIM frame in the current ATIM win-
quorum system [14], and the finite projective plane quorum dow or has extra buffered packets, it should retransmit ATIMs
system [15]), which can be translated to efficient asynchro- in the next ATIM window. To protect PS hosts, only RTS (re-
nous power-saving protocols. We also propose a new e-torus quest to send), CTS (clear to send), ACK, Beacon, and ATIM
quorum system, which can be translated to an adaptive proto- frames can be transmitted during the ATIM window. An ex-
col that allows designers to trade hosts neighbor sensibility ample is illustrated in figure 1.
for power efficiency. A host can dynamically adjust its bea- Table 1
con rate according to its mobility. Simulation experiments are Power consumption of the ORiNOCO IEEE 802.11b PC Gold Card
conducted to evaluate and compare the proposed protocols in (11 Mbps).
terms of the survival ratio, the route establishment probability, Mode PS (doze) Transmit Receive Monitor
and the power efficiency. Power
The rest of this paper is organized as follows. Prelimi- Consumed 60 mW 1400 mW 950 mW 805 mW
naries are given in section 2. Section 3 introduces the rota-
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS 171
formed by picking the second column plus three elements, Proof. Let H be a quorum in the cyclic quorum system Q
each from one of the third, fourth, and fifth columns. H is generated from the difference set D = {d1 , d2 , . . . , dk }. By
formed by picking the sixth column plus three elements, each definition, rotate(H, i) is also a quorum in Q for any i. Then
from one of the first, second, and third columns. G and H by the intersection property, the theorem holds.
intersect at element 7.
As shown in [12], if we let t = w/2, the quorum size will 4.4. The finite projective plane quorum system
be 2tw = 2n, which is near optimal. By equating n,
the torus quorum size is about 1/ 2 that of the grid quorum The finite projective plane (FPP) quorum system [15] arranges
size. Below, we prove the rotation closure property for the elements of the universal set U = {0, . . . , n 1} as ver-
torus quorum system. tices on a hypergraph called the finite projective plane, which
has n vertices and n edges, such that each edge is connected
Theorem 4. The torus quorum system satisfies the rotation to k vertices and two edges have exactly one common ver-
closure property. tex. (Note that the hypergraph is a generalization of typical
graphs, where each edge is connected to only two vertices.)
Proof. Let Q be a torus quorum system formed by a t w A quorum can be formed by the set of all vertices connected
array and H Q be a quorum containing column c. By de- by the edge, and thus has a size of k. It has been shown in [15]
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- In [14], the FPP construction is associated to the construction
trary i. It follows that for any G Q, G rotate(H, i) of Singer difference sets, and it is shown that the FPP quorum
= . system can be regarded as a special case of the cyclic quorum
system when n = k(k 1) + 1 and k 1 is a prime power. It
4.3. The cyclic quorum system follows that FPP quorum systems also own the rotation clo-
sure property, and are optimal, when existing.
The cyclic quorum systems [14] are constructed from the dif-
ference sets as defined below. 4.5. Quorum systems with one quorum
Definition 4. A subset D = {d1 , d2 , . . . , dk } of Zn is called a In this subsection, we discuss the rotation closure property for
difference set under Zn if for every e = 0 (mod n) there exist those quorum systems with only one quorum. The result has
elements di and dj D such that di dj = e (mod n). strong connection to the difference sets, and can help identify
the quorum systems that are solution to the QPS problem.
Definition 5. Given any difference set D = {d1 , d2 , . . . , dk }
under Zn , the cyclic quorum system defined by D is Q = Theorem 6. Let Q = {H } be a quorum system under U =
{G1 , G2 , . . . , Gn }, where Gi = {d1 + i, d2 + i, . . . , dk + i} {0, . . . , n 1}. Q satisfies the rotation closure property if and
(mod n), i = 0, . . . , n 1. only if H is a difference set of Zn .
For example, D = {0, 1, 2, 4} Z8 is a difference set Proof. For the if part, let H be a difference set of Zn . For
under Z8 since each e = 1..7 can be generated by taking the any i, there must exist two elements hx , hy H such that
difference of two elements in D. Given D, Q = {G0 = hx hy = i. It follows that hx = hy + i rotate(H, i) H .
{0, 1, 2, 4}, G1 = {1, 2, 3, 5}, G2 = {2, 3, 4, 6}, G3 = So rotate(H, i) H = for any i.
{3, 4, 5, 7}, G4 = {4, 5, 6, 0}, G5 = {5, 6, 7, 1}, G6 = For the only if part, suppose for contradiction that H is
{6, 7, 0, 2}, G7 = {7, 0, 1, 3}} is a cyclic quorum system un- not a difference set of Zn . Then there exists an i = 0 such
der Z8 . that hx hy = i for all possible combinations of hx and hy
Given any n, a difference set as small as k can be found in H . Since rotate(H, i) = {(hy + i) mod n | hy H }, it
when k(k 1) + 1 = n and k 1 is a prime power. Such a follows that H rotate(H, i) = , a contradiction.
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- Corollary 1. Let Q be a quorum system under U = {0, . . . ,
der Z31 are Singer difference sets. Note that in this case the n 1}. Q does not satisfy the rotation closure property if at
quorum size k meets the lower bound in theorem 2. So cyclic least one quorum in Q is not a difference set under Zn .
quorum systems defined by the Singer difference sets are opti-
mal. Reference [14] had conducted exhausted searches to find Theorem 6 says that if a quorum system has a difference
the minimal difference sets under Zn for n = 4..111. The re- set being its sole quorum, it satisfies the rotation closure prop-
sults are useful here to construct near-optimal cyclic quorum erty and is thus a solution to the QPS problem. Such a quo-
systems. rum system has the practical advantage that it is very easy to
maintain since it has only one quorum to keep. For exam-
Theorem 5. The cyclic quorum system satisfies the rotation ple, from each of the minimal difference sets found in [14]
closure property. (for n = 4..111), a solution to the QPS problem exists by
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS 175
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 JIANG ET AL.
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 (i + 1, j 1) of the array. The NS of two e-torus quo-
and an e-torus(2) quorum intersect with each other. The in- rum systems e-torus(k1) and e-torus(k2) under a t w torus
tersecting elements are guaranteed to appear in the trunks of is discussed below (refer to figure 7 for illustration). When
the Christmas trees. Note that two branches from two e- (k1 = k2 = 1), (k1 = 2 k2 = 1), or (k1 = 1 k2 = 2),
torus quorums may cross with each other, but intersection the two quorums may intersect at only one interval, so NS is
is not necessarily guaranteed (from the geometric structures BI n. For (k1 = 3 k2 = 1) or (k1 = 1 k2 = 3), the NS
of branches, it does look like that they are guaranteed to in- is BI (n 1), which happens when two quorums intersect
tersect). The reason is illustrated in the zoomed-in part in at two consecutive quorum intervals. For (k1 = 1 k2 = 4)
figure 6(b), where the two branches just miss each other on or (k1 = 4 k2 = 1), the NS is BI (n
2t/k1 ), which
the array. Also note that by our arrangement, the intersect- happens when the intersections fall in one column. Table 2
ing elements of two e-torus quorums are unlikely to concen- contains the NS of other cases. Table 2 also summarizes the
trated in certain areas of the array. Instead, they will be spread active ratio and neighbor sensibility of the proposed quorum-
evenly over the trunks. This is a desirable property because it based protocols. Figure 8 further demonstrates the active ratio
implies that the quorum intervals that two mobile hosts may of the proposed protocols for n = 5..100. The cyclic quorum
detect each other will be spread evenly over the time axis. performs the best in terms of active ratio. The FPP quorum
Based on the above features, we propose an adaptive QPS system, when available, represents the optimal solution.
protocol as follows. We can rank a hosts mobility into k-
levels, where level 1 means the lowest mobility, and level k 6.2. Simulation results
means the highest mobility. Whenever a host determines that
its mobility falls within level i (1 i k), it adjusts its quo- In this subsection, we compare the proposed power-saving
rum intervals based on any e-torus(i) quorum. Consequently, protocols through a simulator written in C. An area of size
a host can dynamically adjust its sensibility to the environ- 1000 m 1000 m is simulated. Each host has an antenna
ment change in its neighborhood. 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
6. Performance comparison and simulation results MAC part basically follows the IEEE 802.11 standard [11],
except the power management part. Routes with random
6.1. Analytical comparison sources/destinations are generated, and the AODV (ad-hoc
on-demand distance vector) routing protocol [16] is adopted.
In this subsection, we compare the proposed quorum-based Four parameters are tunable in our simulations:
protocols analytically. We evaluate the active ratio, which is
defined to be the number of quorum intervals over n (the size Mobility. Host mobility follows the random way-point
of universal set), and the neighbor sensibility (NS), which is model, with pause time of 20 sec. When moving, a hosts
the worst-case delay for a PS host to detect the existence of speed can range in 020 m/sec.
a newly approaching PS host in its neighborhood.
The NS Traffic load. Routes are generated by a Poisson distribu-
of the grid quorum system is BI (n n + 1), which tion with rates between 14 routes/sec. For each route,
happens when two quorums intersect at indices (i, j ) and 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.
Table 3
Power consumption parameters used in the simulation.
Table 4
Traffic-related parameters used in the simulation.
Figure 11. Throughput vs. mobility (beacon interval = 100 ms, 100 hosts,
traffic load = 1 route/sec).
Figure 9. Host survival ratio 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 10. Neighbor discovery time vs. mobility (beacon interval = 100 ms,
100 hosts, traffic load = 1 route/sec).
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 16. Throughput vs. traffic load (beacon interval = 100 ms, 100 hosts,
mobility = 020 m/sec with mean = 10 m/sec).
Figure 15. Survival ratio vs. traffic load (beacon interval = 100 ms, Figure 17. Survival ratio vs. host density (beacon interval = 100 ms, traffic
100 hosts, mobility = 020 m/sec with mean = 10 m/sec). load 1 route/sec, mobility = 020 m/sec with mean = 10 m/sec).
the throughput of the AA scheme is the highest, and the ad-hoc wireless local area networks, in: Proc. of IEEE International
throughputs of the E(714) and the C(98) schemes are very Workshop on Mobile Multimedia Communications (1999) pp. 114123.
[7] J.C. Haartsen, The Bluetooth radio system, IEEE Personal Communi-
close. Again, when we consider throughput lifetime, our
cations (February 2000) 2836.
quorum-based protocols outperform the AA scheme signifi- [8] L. Hu, Topology control for multihop packet radio networks, IEEE
cantly. 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.
7. Conclusions of International Conference on Computer, Communication, and Net-
works (2001).
[10] E.-S. Jung and N.H. Vaidya, An energy Efficient MAC protocol for
In this paper, we have addressed the asynchronous power wireless LANs, in: Proc. of INFOCOM 2002 (2002).
mode management problem for an IEEE 802.11-based [11] LAN MAN Standards Committee of the IEEE Computer Society, IEEE
MANET. We have correlated the problem to the concept Std 802.11-1999, Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) specifications (IEEE, 1999).
of quorum systems and identified an important rotation clo-
[12] S.D. Lang and L.J. Mao, A torus quorum protocol for distributed mu-
sure property for quorum systems. We have proved that any tual exclusion, in: Proc. of the 10th Internat. Conference on Parallel
quorum system satisfying the rotation closure property can and Distributed Computing and Systems (1998) pp. 635638.
be translated to an asynchronous power-saving protocol for [13] J.R. Lorch and A.J. Smith, Software strategies for portable computer
MANETs. The purpose of the rotation closure property is energy management, IEEE Personal Communications (June 1998) 60
73.
to deal with asynchronism among hosts clocks. Under the [14] W.S. Luk and T.T. Wong, Two new quorum based algorithms for dis-
rotation closure property, we have derived a quorum size tributed mutual exclusion, in: Proc. of International Conference on
lower bound for any quorum system. We have identified a Distributed Computing
Systems (1997) pp. 100106.
group of optimal or near optimal quorum systems. Optimal [15] M. Maekawa, A N algorithm for mutual exclusion in decentralized
systems, ACM Transactions on Computer Systems (1985) 145159.
or near optimal quorum systems are preferable because in [16] C.E. Perkins and E.M. Belding-Royer, Ad-hoc on-demand distance
a quorum-based power-saving protocol, the number of bea- vector routing, in: Proc. of IEEE Workshop on Mobile Computing Sys-
cons sent and the ratio of a host remaining active are both tems and Applications (1999) pp. 90100.
proportional to the quorum size. We have shown that the [17] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
grid quorum system [15], the torus quorum system [12], the wireless networks using transmit power adjustment, in: Proc. of IEEE
INFOCOM (2000) pp. 404413.
cyclic quorum system [14], and the finite projective plane [18] J.H. Ryu and D.H. Cho, A new routing scheme concerning power-
quorum system [15] are all optimal or near optimal quo- saving in mobile ad-hoc networks, in: Proc. of IEEE International Con-
rum systems satisfying the rotation closure property. We ference on Communications, Vol. 3 (2000) pp. 17191722.
have developed theorems to help identify good quorum sys- [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
tems satisfying the rotation closure property, such as quo- Vehicular Technology Conference, Vol. 4 (2000) pp. 19741978.
rum systems with only one member, which are very easy [20] A.K. Salkintzis and C. Chamzas, An in-band power-saving protocol
to maintain. We have further proposed a new e-torus quo- for mobile data networks, IEEE Transactions on Communications 46
rum system, which can be translated to an adaptive power- (September 1998) 11941205.
[21] E. Shih, P. Bahl and M.J. Sinclair, Wake on wireless: An event
saving protocol allowing hosts to dynamically tune to differ-
driven energy saving strategy for battery operated devices, in: Proc. of
ent quorum systems according to their mobility, so as to trade MOBICOM 2002 (2002).
neighbor sensibility for power expenditure. Extensive sim- [22] T. Simunic, H. Vikalo, P. Glynn and G.D. Micheli, Energy efficient
ulation results have been presented to evaluate these proto- design of portable wireless systems, in: Proc. of the International Sym-
cols. 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.
References [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-
[1] B. Chen, K. Jamieson, H. Balakrishnan and R. Morris, Span: An bile Computing and Networking (1998) pp. 181190.
energy-efficient coordination algorithm for topology maintenance in ad [25] Y.C. Tseng, C.S. Hsu and T.Y. Hsieh, Power-saving protocols for IEEE
hoc wireless networks, in: Proc. of the International Conference on 802.11-based multi-hop ad hoc networks, in: Proc. of IEEE INFOCOM
Mobile Computing and Networking (2001) pp. 8596. (2002).
[2] C.F. Chiasserini and R.R. Rao, A distributed power management policy [26] R. Wattenhofer, L. Li, P. Bahl and Y.M. Wang, Distributed topology
for wireless ad hoc networks, in: Proc. of IEEE Wireless Communica- control for power efficient operation in multihop wireless ad hoc net-
tion and Networking Conference (2000) pp. 12091213. works, in: Proc. of IEEE INFOCOM (2001) pp. 13881397.
[3] C.J. Colbourn and E.J.H. Dinitz, The CRC Handbook of Combinatorial [27] H. Woesner, J.P. Ebert, M. Schlager and A. Wolisz, Power-saving
Designs (CRC Press, 1996). mechanisms in emerging standards for wireless LANs: the MAC level
[4] C.J. Colbourn, J.H. Dinitz and D.R. Stinson, Quorum systems con- perspective, IEEE Personal Communications (June 1998) 4048.
structed from combinatorial designs, Information and Computation [28] S.L. Wu, Y.C. Tseng and J.P. Sheu, Intelligent medium access for mo-
(2001) 160173. bile ad hoc networks with busy tones and power control, IEEE Jour-
[5] L.M. Feeney and M. Nilsson, Investigating the energy consumption of nal on Selected Areas in Communications 18 (September 2000) 1647
wireless network interface in an ad hoc networking environment, in: 1657.
IEEE INFOCOM (2001) pp. 15481557. [29] Y. Xu, J. Heidemann and D. Estrin, Geography-informed energy con-
[6] J. Gomez, A.T. Campbell, M. Naghshineh and C. Bisdikian, A distrib- servation for ad hoc routing, in: Proc. of the International Conference
uted contention control mechanism for power saving in random-access on Mobile Computing and Networking (2001) pp. 7084.
QUORUM-BASED ASYNCHRONOUS POWER-SAVING PROTOCOLS 181
Jehn-Ruey Jiang received his Ph.D. degree in com- Sensor Networks, and as a Guest Editor for Journal of Information Science
puter science in 1995 from National Tsing-Hua Uni- and Engineering special issue on Mobile Computing.
versity, Taiwan. He joined Chung-Yuan Christian He is a two-time recipient the Outstanding Research Award, National Sci-
University as an Associate Professor in 1995. He is ence Council, ROC, in 20012002 and 20032005, and a recipient of the Best
currently an Associate Professor of the Department Paper Award in International Conference on Parallel Processing, 2003. His
of Information Management, Hsuan-Chuang Univer- research interests include mobile computing, wireless communication, net-
sity. He is a recipient of the Best Paper Award in In- work security, and parallel and distributed computing. Dr. Tseng is a Senior
ternational Conference on Parallel Processing, 2003. Member of the IEEE.
His research interests include distributed computing,
mobile computing, distributed fault-tolerance, proto-
cols for mobile ad hoc networks and wireless sensor networks. 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,
Yu-Chee Tseng received his B.S. and M.S. degrees Taiwan, in 1992. He joined the faculty of the De-
in computer science from the National Taiwan Uni- partment of Information Management, Nanya Insti-
versity and the National Tsing-Hua University in tute of Technology, Taiwan, as an instructor in 1996.
1985 and 1987, respectively. He worked for the He is currently a Ph.D. candidate of the Department
D-LINK Inc. as an engineer in 1990. He obtained of Computer Science and Information Engineering,
his Ph.D. in computer and information science from National Central University, Taiwan. His current re-
the Ohio State University in January of 1994. From search interests include wireless communications and mobile computing.
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 Ten-Hwang Lai received his Ph.D. in computer sci-
has become a Full Professor since 1999. Since August 2000, he has become ence from University of Minnesota in 1982. Since
a Full Professor at the Department of Computer Science and Information En- then he has been on the faculty of computer and in-
gineering, National Chiao-Tung University, Taiwan. formation science at the Ohio State University, where
Dr. Tseng served as a Program Chair in the Wireless Networks and Mobile he is presently a professor. His research interests
Computing Workshop, 2000 and 2001, as a Vice Program Chair in the Inter- include parallel and distributed computing, mobile
national Conference on Distributed Computing Systems (ICDCS), 2004, as computing, and wireless networking.
an Associate Editor for The Computer Journal, as a Guest Editor for ACM Lai was on the editorial board of IEEE Transac-
Wireless Networks special issue on Advances in Mobile and Wireless Sys- tions on Parallel and Distributed Systems from 1993
tems, as a Guest Editor for IEEE Transactions on Computers special on to 1996, and is currently an editor of ACM/Kluwer
Wireless Internet, as a Guest Editor for Journal of Internet Technology Wireless Networks, Journal of Information Science and Engineering, and Wi-
special issue on Wireless Internet: Applications and Systems, as a Guest ley Encyclopedia of Computer Science and Engineering. He served as pro-
Editor for Wireless Communications and Mobile Computing special issue on gram chair and general chair of the 1998 and 2000 International Conference
Research in Ad Hoc Networking, Smart Sensing, and Pervasive Comput- on Parallel Processing, respectively; and has been designated as program co-
ing, as an Editor for Journal of Information Science and Engineering, as chair and general chair of the 2004 and 2005 IEEE International Conference
a Guest Editor for Telecommunication Systems special issue on Wireless on Distributed Computing Systems, respectively.
Mobile Networks and Applications 10, 183197, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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
the frame body. The field source.ad gives the Ethernet address section 2.8. If k has reached the maximum K or if the bit t is
of the source of the packet. set to 1, no more request can be done on this slot.
Note that all packets, including data packets, have a fixed
size, and each mini-slot is just long enough to allow the trans- 2.4.4. Data packets
mission of the associated packet. For example, the time to In data packets, the field dest.ad gives the address of the des-
transmit a REQ including additional bits from the physical tination of the packet.
layer, the transmit-to-receive turn around time, as well as a As previously explained, each sender maintains a counter
small time interval to take into account the propagation de- that is incremented for each new packet. This sequence
lays equal the time of the REQ-mini-slot. Note also that it is number is put in the field sn and is used by the receiver to
preferred that the size of the control packets are short com- acknowledge the reception of the packet. Let us recall that
pared to the length of the data packets (e.g., 512 bytes). data packets have a fixed size, that results of a higher layer
segmentation or aggregation.
2.4.2. REQ control packet
In a REQ, the field dest.ad gives the ethernet address of the 2.5. Reservation
destination of the packet (the intended receiver). The field
qs is used by a requesting node to indicate to the intended re- Any communication between two nodes must be preceded by
ceiver the requested quality of service for the communication. a preliminary reservation phase. In the reservation phase, re-
This field may be used by higher layers to negotiate the QoS. questing nodes contend to get access to a receiver. This access
It will be used in future versions of the protocol. is done in a random way during the REQ-mini-slots and con-
sists of five sub-phases: listening of an entire frame, choice
2.4.3. RTR control packet of a time-slot, transmission of the REQ on the chosen slot,
A RTR has three different functions, as illustrated in sec- listening of the RTR, and retry of a new reservation phase in
tion 2.2 and in figure 2: respond to a REQ, poll the different case of failure (with or without random backoff). These five
senders on the current slot and acknowledge data packets. sub-phases are now detailed.
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 2.5.1. Frame listening
a request is correctly received and accepted, it is acknowl- The first phase of the reservation consists in listening to the
edged by putting the address of the requesting node in the RTR-mini-slots during an entire frame, and maintaining for
field req.ad and the value ACK in the field r. If a request has each slot in the frame the state of the slot. This listening
been correctly received, but the communication cannot be es- process starts at the beginning of the reservation phase and
tablished, the field r is set to NACK. This situation is possible lasts until the reservation has succeeded.
if the requested QoS is not allowed or if the number of current A slot can be in several states:
communications has reached its maximum, K. If the receiver FREE: no activity has been sensed during the RTR-mini-slot,
detects a collision of REQs, r is set to COL. If the receiver did i.e., no receiver has got the floor on this slot. A request will
not received any request, or if the request cannot be decoded be possible on this type of slot.
because of the channel conditions, r is set to NOTRECVD. OCC-NA: i.e., occupied and not available. This is the case
The values NACK, COL, and NOTRECVD are useful infor- if a RTR has a source.ad different from the address of the
mation for the requesting nodes to reschedule their requests. intended receiver or if the requesting node has detected a
The field polled.ad is used by a receiver to poll a sender collision during the RTR-mini-slot, or if it did not man-
that previously managed to establish a connection on this slot. aged to decode the field source.ad in the RTR, or if the
If a sender reads its address in the field polled.ad, it is allowed requesting node is itself a receiver on this slot. This is also
to send a data packet during the DATA-mini-slot of the same the case if the field k of byte n has reached the maximum
slot, just after receiving the RTR. number of communications on a slot or if the bit t of byte
The acknowledgement of data packets is done thanks to n is equal to 1. Note that a RTR collision detected on a slot
the field sn that stands for sequence number. Each node main- does not necessary mean that the slot is free in a multi-hop
tains a counter that is incremented for each new data packet. situation. A request will not be possible on this slot.
Receivers keep the last received sequence number. If in time- OCC-A-COL-k: i.e., occupied, available, collision, and k
slot i of frame j , a receiver has received a data packet with communications. In this case, the source.ad of the RTR
sequence number m, it sets the field sn to m in the RTR of the is the address of the intended receiver, a collision has
slot i of frame j + 1 and so, acknowledges the previous data been detected by the receiver during the REQ-mini-slot
packet. (r = COL in the RTR), and there are currently k < K
The byte n of a RTR gives information about the slot uti- communications on the slot. A request will be possible on
lization. It is decomposed into seven bits that indicate the this slot.
number k of current communications, and one bit t to inform OCC-A-NCOL-k: i.e., occupied, available, no collision, and
that the receiver will not accept requests on this slot anymore. k communications. In this case, the source.ad of the RTR
More details on the use of the bit t for fairness are given in 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.
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 COUPECHOUX ET AL.
a sent data packet is stored until the receipt of the acknow- 2.8. Fairness issue
ledgement. If the next RTR is not received or if this RTR does
not acknowledge the stored packet, a retransmission is nec- CROMA includes a mechanism to ensure a local fairness
essary. After M retransmissions the stored packet is thrown among data flows. On a given time-slot, fairness among in-
away. This loss can be treated by an upper layer. coming flows is assured by the receiver of the slot by means
Figure 4 shows an example of a transmission phase with of the RTRs. By using different polling strategy, a receiver
a receiver and three senders. It only shows slots i of suc- can easily give a fair allocation of the slot to incoming flows.
cessive frames. On the upper part of the figure, the RTRs of However, if the number of slots in the frame is small
the receiver are represented with the fields polled.ad and sn. compared to the number of potential receivers, situations of
A cyclic polling is pictured for the scheduling of the senders unfairness can arise and flows can be completely starved. The
bit t included in the RTRs is used in order to avoid such situ-
and data packets are shown with their field sn.
ations.
It is clear that each receiver acts on a given slot as a local
A receiver having the floor on a given slot counts the num-
base-station with respect to its associated senders. Thus, the
ber of consecutive full frames. A frame is full from the point
polling mechanism allows a high flexibility for the schedul-
of view of a receiver, if it senses activity at each slot of the
ing of different flows by higher layers and is a base for the
frame. In this case, it detects a potential blocking situation
implementation of QoS algorithms. Moreover, several paral-
for pair of nodes that cannot communicate because there are
lel communications are possible on a given time-slot.
no free slots any more. If the number of monitored full frames
reaches MAX_FULLFRAMES, the receiver sets the bit t to 1
2.7. Release indicating that it will not accept new requests and that the cur-
rent communications have to be released.
An established communication can be interrupted in three A sender detecting a bit t set to 1, sets the field sn of its
cases. next packet header to EOT and stops sending packets to the
receiver. This release is done even if the sender have still
(i) The sender informs the receiver that it sends the last packets to transmit. A requesting node detecting a bit t set to
packet of the communication by setting the field sn of 1 in a RTR update the slot state to OCC-NA.
the packets header to the value EOT (end of transmis- This strategy aims at avoiding blocking situations that can
sion). If the last packet is correctly received, the receiver lead to unfairness. Indeed, these cases are detected by the
does not re-schedule the sender any more. However, it receivers that have to free their slot if the situation lasts.
acknowledges the last packet with its next RTR, and this,
even if it has no sender to poll.
3. Correctness
(ii) If a receiver has polled a sender and does not receive In this section, we will show that CROMA is correct, i.e., that
any packet from the sender, a counter set to W is it is collision-free in both fixed and mobile environment. The
decremented. When this counter reaches 0, the commu- capture effect is not considered here, so this section shows
nication is released, and the receiver does not poll the that CROMA is collision-free in the common case provided
sender any more. If after a poll, a packet is received, the that a sender releases its communication as soon as it detects
counter is set again to W . After each polling, a sender a collision of RTRs.
starts a timer. If it does not receive any polling from the Let us first consider a fixed and multi-hop topology. We
receiver when the timer expires, the connection is con- now prove that two data packets cannot collide.
sidered to be broken. We suppose that a collision of two data packets occurs at
a receiver R1 . These packets have been sent by two differ-
(iii) During a communication, a sender may receive several ent senders, namely S1 and S2 . During the RTR-mini-slot, R1
RTRs, i.e., there is a collision of RTRs. In this case, specified the MAC address of the sender, say S1 , that was al-
the sender considers that the current communication on lowed to send its data in the current slot. As the MAC address
this slot is released. Indeed, sending a data packet could is unique, a single colliding data packet is destined to R1 .
imply a collision during the DATA-mini-slot. More pre- Therefore, we know that the data packet of S2 was destined to
cisions about this specific aspect are given in section 3. another receiver, R2 .
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 189
Now, as R1 has received a data packet from S2 and links The righthand side of figure 5 shows topologies, where
are bi-directional, S2 has received the RTR of R1 . More- communications are released because both senders detected
over, S2 has also received a RTR from R2 , since it sent a data a collision of RTRs. Case (j) shows a configuration where
packet destined to R2 . Thus, S2 has detected a collision of the network of nodes is fully connected after mobility. Here,
RTRs in the current slot without interrupting its communica- RTRs of nodes 2 and 4 collide at nodes 1 and 3. On detecting
tion with R2 . This is impossible. As a conclusion, no data the collision, they decide to interrupt their communication.
collision can occur in a fixed topology. So, in the common case, in both fixed and mobile environ-
Let us now consider the case of a dynamic topology. Two ment, CROMA is collision-free. As in all protocols that rely
concurrent communications on a slot are shown on the top of on the exchange of short control packets, the capture effect
may, however, affect this conclusion.
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 4. Analytical study
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 In this section we calculate the approximate throughput, i.e.,
range of 1, 2, or both 1 and 2. Same alternatives can occur for the slot utilization of the protocol CROMA in a fully con-
node 4. Thus, after mobility, a total of 16 relative new posi- nected network. Following [27], we claim that this topology
tions are possible. Because of the symmetry of the problem, is the worst case in terms of interference, contention, and spa-
only 10 cases are shown in figure 5. tial reuse because CROMA guarantees a collision-free trans-
mission of data after reservation in a multi-hop environment.
The left-hand side of figure 5 shows situations, where
a single communication is interrupted because the sender
4.1. Model for the slot utilization analysis
detected a collision of RTRs on the considered slot. For ex-
ample, in case (b), node 4 moved in the transmission range of First of all, we describe our analytical model for the slotted
nodes 1 and 2. In frame j + 1, nodes 2 and 4 send simulta- MAC protocol CROMA. From this model will be derived the
neously an RTR. Node 3 receives correctly the polling of 4, slot utilization of CROMA as a function of the probability
whereas node 1 senses a collision during the RTR-mini-slot. p to send a REQ for a given sourcedestination pair. Lets
Node 1 decides to interrupt the communication with node 2 enumerate the assumptions of our model.
and does not send any data packet on this slot. If node 1 has 1. We consider a fully-connected network of N synchro-
still packet in its buffer, it has to enter a new reservation phase. nized nodes.
The central part of figure 5 shows exposed-terminal
2. All packets are of constant length and are transmitted
topologies, where both communications can still share the
over an assumed noiseless channel.
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 3. There are L slots per frame.
of node 2 (resp. 4) because it is out of the transmission range 4. The maximum number of connections on a slot is K, i.e.,
of node 4 (resp. 2). Both nodes 1 and 3 can send data packet when a receiver is already polling K different senders on
during the DATA-mini-slot. a slot, no new REQ is allowed.
190 COUPECHOUX ET AL.
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.
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
P (K K) = i (1 q) + q(1 i ), (22)
P (K K 1) = 1 P (K K). (23)
Table 2
5.1. Methodology
Main parameter values for simulations.
Figure 14. End-to-end delay jitter vs. input load, squares topology.
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
Figure 22. End-to-end delay jitter vs. input load, random topology.
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,
Figure 23. Fairness index vs. input load, random topology. 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
random topologies because of the simulation time. However, (September 1997) pp. 3949.
ten different random connex networks (not shown here) have [15] J.J. Garcia-Luna-Aceves and C.L. Fullmer, Floor Acquisition Multiple
been simulated and provide similar results. 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,
6. Conclusion 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
In this paper, a new MAC protocol, called CROMA has been 6(3) (June 1998) 298306.
proposed for mobile ad hoc networks. CROMA operates in a [18] P. Karn, MACA a new channel access method for packet radio, in:
slotted environment, it is collision-free and receiver-oriented. Proc. of ARRL/CRRL (April 1990) pp. 134140.
The reservation of resources is made through a random access [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)
phase on each slot of the frame. The transmission is done pp. 940945.
thanks to a polling by the receivers. Thus, receivers of a con- [20] H. Luo, S. Lu and V. Bharghavan, A new model for packet scheduling
nection act as local base-stations and sophisticated functions in multihop wireless networks, in: Proc. of ACM/IEEE MOBICOM00
at higher layers can be easily implemented. (August 2000) pp. 7686.
[21] ns2, http://www.isi.edu/nsnam/ns
The correctness of CROMA has been proven. Even with a
[22] L.C. Pond and V.O.K. Li, A distributed time-slot assignment proto-
dynamic topology, CROMA handles both the hidden and the col for mobile multi-hop broadcast packet radio networks, in: Proc. of
exposed terminal problems. IEEE MILCOM89, Vol. 1 (October 1989) pp. 7074.
Theoretical performance analysis and extensive simula- [23] S. Ramanathan, A unified framework and algorithm for (T/F/C)DMA
tions show that CROMA can reach very high throughputs in channel assignment in wireless networks, in: Proc. of IEEE INFO-
COM97, Vol. 2 (April 1997) pp. 900907.
a fully connected network provided that the average message [24] R. Ramaswami and K.K. Parhi, Distributed scheduling of broadcasts in
length is large. Moreover, CROMA outperforms IEEE 802.11 a radio network, in: Proc. of IEEE INFOCOM89, Vol. 2 (April 1989)
at high input loads thanks to a better channel utilization. pp. 497504.
CROMA AN ENHANCED SLOTTED MAC PROTOCOL FOR MANETS 197
[25] W.J. Stewart, An Introduction to the Numerical Solution of Markov Christian Bonnet joined Institut EURECOM as an
Chains (Princeton University Press, NJ, 1994). Associate Professor in 1992. Since 1998 he has
[26] F. Talucci, M. Gerla and L. Fratta, MACA-BI (MACA by invitation) been at the head of the Mobile Communications De-
a receiver oriented access protocol for wireless multihop networks, in: partment of EURECOM. He teaches distributed and
Proc. of IEEE PIMRC97, Vol. 2 (September 1997) pp. 435439. real-time systems, mobile communication systems,
[27] Z. Tang and J.J. Garcia-Luna-Aceves, A protocol for topology- wireless LANs and protocols for mobility manage-
dependent transmission scheduling in wireless networks, in: Proc. of ment. His main areas of research are Mobility
IEEE WCNC99, Vol. 3 (September 1999) pp. 13331337. Management protocols, wireless access to IP Net-
[28] N.H. Vaidya, P. Bahl and S. Gupta, Distributed fair scheduling in a works and data communications in mobile networks
wireless LAN, in: Proc. of ACM/IEEE MOBICOM00 (August 2000) including Mobile Ad Hoc networks. He is currently
pp. 167178. participating in research projects related to UMTS in the field of QoS and
[29] S. Xu and T. Saadawi, Does the IEEE 802.11 MAC protocol work well Ipv6.
in multihop wireless ad hoc networks?, IEEE Comm. Magazine 39(6) E-mail: christian.bonnet@eurecom.fr
(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 Vinod Kumar received M.Sc., and Ph.D. from
(March 1998) pp. 322331. 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.
Marceau Coupechoux received M.Sc. in mathe-
He has more than 20 years of experience in digital
matics from the University Pierre et Marie Curie
mobile communications, signal processing and per-
(Paris) in 1998 and a double Eng. Degree from the
formance evaluation of higher protocol layers, and
Ecole Nationale Suprieure des Tlcommunications
he has more than 20 patents, and more than 30 pub-
(ENST), Paris, in 1999, jointly with the University
lications in technical journals, books and conference
of Stuttgart in 2000. He joined Alcatel Research &
proceedings in above mentioned areas. He is Associate Professor in Univer-
Innovation in 2000 and is Ph.D. student since 2001
sity of Marne La Valle, and Visiting Professor in several other institutes in
at the Institut Eurecom, Sophia-Antipolis. He is the
France. He is Officiating Secretary of the Wireless World Research Forum
author or co-author of 10 conference papers and 5
(WWRF).
patents.
E-mail: vinod.kumar@alcatel.fr
E-mail: marceau.coupechoux@alcatel.fr
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 relay their data to their respective destinations. They should
be able to directly communicate with their destinations. If a
In recent times, much effort has gone into solving the prob- base-station is used as an intermediary, direct one-hop trans-
lem of transmitting multimedia data over wireless networks. missions are needlessly made two-hop. (The base-station
Three mutually orthogonal factors make this problem chal- must only serve as an access point to the wired Internet, not as
lenging: (a) stringent QoS requirements of multimedia ap- a relay for peer-to-peer transmissions between mobile nodes
plications, (b) bursty nature of some multimedia traffic, and within the wireless network.) Furthermore, in military and
(c) unreliable and dynamic nature of the wireless network. disaster rescue environments, a group of people carrying mo-
Network-specific QoS requirements of multimedia applica- bile handheld devices should be able to communicate with
tions include minimum throughput, maximum delay and each other, with no time for planning and building a support
maximum jitter. infrastructure such as a base-station. The single-hop ad hoc
In a wireless network, the minimum throughput require- wireless network, without a base-station, thus accurately rep-
ment is more difficult to achieve than in a wireline net- resents the network used in smart-rooms, hot-spot networks,
work, because (a) this requires distributed co-operation be- emergency environments, and in-home networking.
tween nodes sharing a wireless channel, and (b) the flows in IEEE 802.11 has recently become the de facto Medium
the wireless network are exposed to various physical chan- Access Control (MAC) standard in connecting mobile hosts
nel errors. In smart-rooms and hot-spot networks, wireless in an ad hoc network environment. It relies on the Distrib-
access-enabled nodes in a small area share limited channel uted Co-ordination Function (DCF) to resolve channel access
bandwidth. Since the area is small, the wireless hosts per- contention in a distributed way. However, the IEEE 802.11
vade through the entire network and are all within each others DCF does not currently have any provision to guarantee QoS,
transmission range. There are a large number of hosts and such as minimum throughput, to flows accessing the channel.
hence connections. So, channeling all data through a sin- Without any co-ordination, if the sum of transmission rates
gle intermediate node, such as a base-station, is inefficient. of all the hosts (or flows) is greater than the channel capac-
Communication is pervasive, i.e., there are many source ity, heavy channel contention will occur and thus QoS cannot
destination pairs distributed throughout the network. The be guaranteed for any flow. Much research has been done
sources must not all rely on a single entity, the base-station, to in the area of distributed weighted fair scheduling (DWFS)
[3,13,18,19,24] for IEEE 802.11 networks operating in the
This work was supported by the DoD Multi-disciplinary University Re-
DCF mode. In DWFS, each flow is assumed to have a weight
search Initiative (MURI) program administered by the Office of Naval Re-
search under Grant NAVY CU 37515-6281, and the NSF EIA 99-72884
which defines its importance relative to other flows. A sched-
grant. Any opinions, findings and conclusions are those of the authors and uler combined with the MAC-layer IEEE 802.11 protocol
do not necessarily reflect the views of the above agencies. then schedules the flows packets on the channel such that
200 SHAH ET AL.
the throughput they receive is proportional to their weights. ment. Similarly, when a flow ends, its CTP must be suitably
However, DWFS is not yet a standard part of the IEEE 802.11 redistributed among the still existing flows so they can hope
MAC protocol. to achieve a better QoS.
In the absence of distributed MAC-layer weighted fair The BM must therefore not just perform one-time admis-
scheduling in the current IEEE 802.11 standard, we propose sion control and teardown, but also perform dynamic band-
a scheme at the higher layers of the OSI protocol stack to co- width management. The BM must re-negotiate with each flow
ordinate individual flows channel access in the single-hop its CTP as its channel characteristics change, and as the num-
ad hoc network scenario, in order to promote co-operation ber of active flows in the network varies. The detection of
between flows and provide minimum throughput guarantee change in channel characteristics, and adaptation of the flow
for each of them. To this end, we first determine the flows to this change, happen continuously through the course of the
weights based on their relative channel access requirements. session. Bandwidth re-negotiation must also occur before a
The flow weights, in turn, determine the transmission rate flow changes its packet transmission rate, as in the case of
of each flow. The flows transmission rates are controlled at bursty VBR traffic.
the application or middleware layers, without any MAC-layer The rest of the paper is structured as follows. The next
scheduling support. Therefore, our scheme can be used over section describes the overall network topology, the architec-
the standard IEEE 802.11 protocol and is easily deployable ture of the bandwidth management system and the bandwidth
using todays off-the-shelf 802.11 products. In case DWFS management protocol. Section 3 presents our experimental
becomes available at the MAC-layer in the future, our scheme results. Section 4 discusses some related work in the field.
is still required in order to provide the MAC-layer scheduler Finally, section 5 concludes the paper.
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- 2. Bandwidth management system design and
pends on its requirements relative to the requirements of other implementation
flows. Each flow maps its minimum and maximum bandwidth In the previous section, we motivated the need for admission
requirements to its minimum and maximum channel time pro- control coupled with dynamic bandwidth management in a
portion (CTP) requirements, respectively. We propose the use single-hop ad hoc wireless network. In this section, we de-
of a centralized Bandwidth Manager (BM), which obtains scribe the characteristics of the network we are concerned
from each flow its CTP requirements, at the start of its ses- with, the architecture of the bandwidth management system
sion. It uses this information to gauge what proportion of unit and the communication protocol.
channel time (CTP) each flow should be allotted. The CTP
allotted by the BM to each flow (i.e., its flow weight) lies 2.1. Network model
somewhere between the flows minimum and maximum re-
quirements. The term channel time proportion is defined as We design and implement our bandwidth management scheme
the fraction of unit time for which a flow can have the channel for a wireless network consisting of heterogeneous computers
to itself for its transmissions. Since our network model allows and devices connected together over the IEEE 802.11 MAC
only one node to transmit on the channel at a time, there is a layer. The network in our prototype testbed implementation
direct correspondence between the channel time a flow uses consists of handheld PCs and laptop computers with their
and the share of the network bandwidth it receives. The BM 802.11 interfaces configured in peer-to-peer ad hoc mode. We
may also refuse to admit a flow, i.e., allot 0% channel time. assume that each node in the network is within the transmis-
This can happen if the flows minimum CTP requirement is sion range of every other node. Hence, only one node can
so large that the network cannot support it, without violating transmit at a time over the channel. Since every node is within
some other flows minimum CTP requirement. the transmission radius of every other node, routing is single-
The problem with the admission control solution described hop.
above, however, is that it is a one-time procedure performed Unlike in [5], where a base-station determines the sched-
before the flow starts. It does not take into account the ule of transmission for the entire network and all communi-
changes in the wireless network over the duration of the flows cation is via the base-station, in our network, transmission is
operation. Not only can the perceived channel capacity vary distributed and peer-to-peer. The IEEE 802.11 MAC proto-
over time due to varying contention [6] as flows arrive and cols DCF, which is the one relevant to our network model,
depart, but the channel capacity as perceived by different net- does not have a provision for a fixed transmission schedule.
work nodes at the same time can also be different. The latter A node can send when it senses that the channel is not busy.
phenomenon is due to location-dependent fading errors and A binary exponential backoff mechanism resolves collisions
location-dependent interference from external objects. that might occur as a result of nodes transmitting at random
When a new flow arrives and demands a share of the chan- times. Moreover, any node in the network can transmit to
nel, the respective CTPs allotted to already existing flows may any other node directly without using the base-station as an
have to be reduced in order to accommodate it. This revo- intermediary hop. Figure 1 illustrates our network model as
cation of channel time should not, however, result in these compared to the base-station model. The distributed, peer-to-
existing flows ceasing to meet their minimum CTP require- peer and ad hoc nature of our wireless network model makes
DYNAMIC BANDWIDTH MANAGEMENT 201
from the BM. It then controls the transmission rate of each will also decrease. This may cause it to fall substantially
flow depending on its allotted CTP. For the sake of simplicity, below its minimum bandwidth requirements. Hence the re-
in our UDP simulation experiments and testbed experiments, negotiation. We do not wish to re-negotiate for small changes
the RA is built into the UDP application itself, to adapt its data in Bp (f ), however, in order to keep re-negotiation overhead
generation rate. Ideally, however, to avoid changing the ap- small. The RAs not reacting to small changes in Bp (f ) may
plication, we recommend that the RA be implemented sepa- thus cause small violations of the minimum bandwidth re-
rately as a module and be linked to the application at run-time. quirements. (But not minimum CTP requirements.) The mo-
It would thus function as middleware, just below the applica- ment a large violation occurs, the RA immediately reacts and
tion layer, and shape the applications traffic. Various queue- re-negotiates. The parameter that defines small and large
based rate controllers are available for this purpose [16]. Our is tunable. It trades off the hardness of the bandwidth guaran-
interest is in the design of the overall bandwidth management tee with re-negotiation overhead.
architecture, rather than the implementation of individual rate
control mechanisms. For our TCP simulation experiments, Example. Assume a flow f in a 2 Mbps wireless network has
we simulate queue-based rate control by having an RA per- minimum bandwidth requirement 300 Kbps and perceives to-
node at the network interface queue, rather than within the tal network bandwidth of 1.5 Mbps. (That is, the flow f per-
application. ceives this to be the total capacity of the 2 Mbps channel.)
Note that, in case DWFS is present at the MAC layer, shap- Assume further that the CTP allotted to it is 20%, thus en-
ing the traffic and enforcing flow rates can be left to it. The suring it just meets its minimum bandwidth requirement. If
RAs function is deprecated to merely communicating with the total network bandwidth, as perceived by f , decreases to
the BM and determining the flow rate. 1.2 Mbps due to an increase in physical channel errors or con-
tention, then the 20% channel time is no longer sufficient for
Total Bandwidth Estimator (TBE). The per-node Total the flow to meet its minimum bandwidth requirement. Its RA
Bandwidth Estimator is co-located with the IEEE 802.11 pro- must then re-negotiate for at least a 25% of the channel time.
tocol at the MAC layer. It estimates the total network band- Similarly, if a flow perceives the total network bandwidth to
width Bp (f ) for each flow f sourced at the node it resides have increased, it must release any excess share of the channel
on.1 Bp (f ) is what flow f perceives to be the total band- it has been allotted, so that some other flow can use it.
width of the network at a particular time. In other words, at
a particular instant in time, Bp (f ) is equal to the theoretical Bandwidth Manager (BM). The Bandwidth Manager per-
maximum capacity of the channel (1, 2, 5.5 or 11 Mbps for forms admission control at the time of flow establishment
IEEE 802.11) minus the bandwidth lost due to channel errors, and bandwidth redistribution at the time of flow teardown.
caused by fading, interference and contention experienced by Admission control involves revocation of some channel time
flow f s packets, at that instant. The physical channel errors from existing flows and re-allocation of this portion to the new
and contention at a particular instant in time is estimated from flow. The BM also performs re-negotiation either when some
the errors and contention experienced in recent history. De- flow detects a change in its perceived bandwidth or when its
tails of the estimation method of Bp (f ) are in section 2.4. traffic characteristics change.2
Note that the TBE is per-node whereas it performs total band- The BM admits a flow only if it can allot at least its mini-
width estimation per-flow sourced at the node it resides on. mum CTP requirement. Otherwise, the flow is rejected. The
The TBE continuously measures the total perceived band- remaining channel time as yet unallotted after all the admit-
width for each flow. It periodically passes this up to the RA ted flows minimum channel time requirements are satisfied,
of the flow at the higher layers. The RA of a flow f uses is allotted on a maxmin fair basis. We therefore deem our
it in the translation of flow f s bandwidth requirements to channel time allocation scheme at the BM maxmin fair with
its CTP requirements. When the total bandwidth Bp (f ) per- minimum guarantee. Each flow receives whatever CTP is
ceived by flow f changes, the channel time requirements cal- allotted to it by the maxmin fair algorithm, in addition to
culated using Bp (f ) also change. The TBE informs the RA its minimum CTP request, which is automatically guaranteed
of the new Bp (f ). The RA may now need to re-negotiate on when it is admitted. A detailed description of the maxmin
behalf of flow f with the BM, using flow f s new CTP re- fairness algorithm can be found in section 2.5 of the paper.
quirements that are calculated with the new Bp (f ) estimate.
Since CTP allotted to flow f is directly related to its share of
2.3. Bandwidth management protocol
total network bandwidth, if a flow perceives the total network
bandwidth as having decreased, its share of the bandwidth
This section describes the protocol used in the interactions be-
1 In a single-hop peer-to-peer wireless network, we perform bandwidth man- tween the various components of the bandwidth management
agement per-flow, since each flow can have a different destination. In a architecture and the details of the BMs operation. The BM is
base-station environment, we can perform bandwidth management per-
node since every node only communicates with the base-station. In the 2 The centralized BM does not take on the onus of channel bandwidth esti-
base-station scenario, each node, rather than application, specifies its band- mation, and leaves this to the individual per-node TBEs, because the avail-
width requirements to its RA, and bandwidth estimation is done only for able channel capacity is different for different peer-to-peer flows, due to
links between mobile nodes and the base-station. location-dependent physical errors.
DYNAMIC BANDWIDTH MANAGEMENT 203
curate estimate, and re-negotiation done with these modified effort flows, since pmin (g) = 0, pa (g) = pmm (g). In other
values. words, channel time is allotted to best-effort flows only after
Alternatively, in the case of a connection-oriented flow, the all the higher priority real-time flows are all allotted at least
first few flow-establishing packets can be used in the total their minimum share.
bandwidth estimation instead of the hardcoded estimate. For After the new flow f is admitted, the BM registers an entry
example, the physical channel errors and contention faced by pertaining to it in its flow table. This entry consists of: (a) the
TCPs three-way handshake messages can be used in the ini- new flow f s flow-id, (b) the socket descriptor of the socket
tial measurement. If the application involves some other con- used by the BM for communication with f s RA, (c) pmin (f ),
trol messages (e.g., client asking server if file exists or not), (d) pmax (f ) and (e) pa (f ). The socket descriptor is stored in
then these can be used. A current estimate being used by other the table so that if any re-negotiation needs to be done later
flows between the same end-points can also be used initially. with flow f s RA (for example, when newer flows arrive in
A fourth option is to have the BM maintain a list of current future or existing flows depart), this socket can be used. In
total bandwidth estimates for all flows. Then, a new flow can addition, a timestamp indicating the freshness of the latest
query the BM for an initial estimate. The BM simply returns request message is also maintained for each flow. This
the average of the list of total bandwidth estimates. timestamp is used for two purposes: (a) timing out stale reser-
Let the initial total bandwidth estimate, how ever it vations, and (b) proper ordering of multiple outstanding re-
may be obtained, for a new flow f be Bp (f ). The CTP negotiation requests from the same flow. Since reservations
pmin (f ), required to satisfy the new flow f s minimum band- can time-out, the entries in the flow table are soft-state en-
width requirement Bmin (f ), is pmin (f ) = Bmin (f )/Bp (f ). tries. If, for some reason, a flows reservation has timed-out
pmin (f ) = 0 for best-effort flows. Similarly, the CTP but the flow is still transmitting, this can be detected using a
pmax (f ), required to satisfy flow f s maximum bandwidth policing mechanism. (See section 2.7.)
requirement, is pmax (f ) = Bmax (f )/Bp (f ). The RA of the Finally, for every flow g F , the allotted CTP pa (g) is
new flow f sends the BM a request message containing then sent to flow gs RA using a reply message. (Note that
the flow-id of f , pmin (f ), pmax (f ) and a timestamp for or- the name of the message is a misnomer in the case of all flows
dering. g F except the new flow f because, in their case, the re-
The BM checks whether, for all flows g in the set F of ply is gratuitous, not a response to any message they sent.)
previously registered flows, 1 gF pmin (g) pmin (f ). It may be the case that all flows g F do not need to be sent
If this is true, the new flow f is admitted (F = F {f }), else a reply message. No reply message needs to be sent to a
it is rejected and a reply message offering it zero CTP is flow in F whose allotted CTP has not changed due to the ar-
returned to its Rate Adaptor. Note that a best-effort flow with rival of the new flow f . Although we implement the reply
pmin (f ) = 0 is always admitted. A rejected flow may attempt message as multiple unicast messages to individual RAs for
again later to gain access to the channel. Flows are admitted reliability, it can also be implemented for efficiency as a sub-
strictly in the order they arrive, to alleviate starvation of pre- net broadcast message, containing flow-id and pa (g), g F .
viously rejected real-time flows. The problem of starvation of A flow f is rejected using a unicast reply with pa (f ) = 0.
a best-effort flow after admission is dealt with in section 2.6. Other existing flows allotted CTPs are not affected.
Once the new flow f is admitted, the BM must redistribute The RA of every flow that receives a reply message, gra-
channel time within the new set of existing flows F . Since the tuitous or otherwise, from the BM sets its transmission rate
original admission test was passed by flow f , accommodat- respectively to pa (g) Bp (g) bits per second (bps). The new
ing it will not cause the CTP allotted to any flow g F to flow f can now begin operation whereas the older flows sim-
fall below its minimum CTP request. Hence, the BM initially ply resume operation with their respective new rates.
sets allotted CTP pa (g) = pmin (g), g F . The remain- The timestamp in the reply to flow g indicates the last
ing channel time, prem = 1 gF pmin (g), is distributed request received from g by the BM. The value of Bp (g)
among the flows g F in maxmin fair fashion. Our chan- used to compute pmin (g) and pmax (g) for this request must
nel time allocation policy is thus called maxmin fair with then be used in the transmission rate formula above, since it
minimum guarantee. The maximum CTP requirement for is based on this value of Bp (g) that pa (g) was calculated by
each flow g F in the maxmin fair computation is set to the BM. As a new Bp (g) is returned by the TBE periodically,
pnewmax (g) = pmax (g) pmin (g). This is because pmin (g) a new rate is also used periodically. If the Bp (g) change is
has already been allotted to it and it only needs pnewmax (g) large since the last period, re-negotiation must occur, as ex-
more to fulfill its maximum CTP requirement. Thus, know- plained below.
ing prem and pnewmax (g) g F , the maxmin algorithm can
now proceed. Details of the maxmin fairness algorithm can Flow teardown. When a flow f terminates, its RA sends
be found in section 2.5. a teardown message to the BM. The BM removes flow f
Suppose that out of the remaining channel time prem , the from the set of existing flows F i.e., F = F {f }. It then
amount allotted to any flow g F by the maxmin algorithm redistributes flow f s allotted CTP pa (f ) among the other
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 antees. The RA of each flow g F (the new set F ) is
flow g F is pa (g) = pmin (g) + pmm (g). Note that for best- told of its newly allotted CTP by the BM. The socket de-
DYNAMIC BANDWIDTH MANAGEMENT 205
scriptors in the flow table are used to send gratuitous re- imum requested CTP pmin (f ), or it will receive no channel
ply messages for this purpose. The entry for the terminating time at all. The guarantee in terms of bandwidth is that the
flow f in the BMs flow table is expunged. A teardown- allotted bandwidth never falls more than a factor of below
acknowledgement message is sent to f s RA. the minimum requested bandwidth Bmin (f ), since if Bp (f )
changes by a factor of , re-negotiation occurs.
Change in a flows perception of total network bandwidth. Currently, we do not use any priority scheme to cut-off
The RA of every flow periodically obtains from the TBE the particular flows. If perceived bandwidth decreases for all
flows current perceived total bandwidth. The TBE updates flows, the first flow initiating re-negotiation is cut-off. Al-
the RA with the mean of the perceived total network band- ternate strategies to pick flows to cut-off in mid-operation are
width measured for each packet successfully transmitted by discussed briefly in section 3.1.1.
the flow in recent history. The inter-update period could be
in terms of number of packets transmitted or in terms of time. Change in a flows traffic characteristics. When a VBR
We recommend using a hybrid scheme for determining update UDP flow f (e.g., MPEG video stream) needs to send a burst
period: it should be based on time when the transmission rate of traffic at a rate different from its normal rate, it must inform
of the flow is low and based on number of packets transmitted its RA. The RA will re-negotiate for a larger CTP for flow f
when it is high. In our experiments, we use high transmission depending on the bandwidth of the burst. The re-negotiation
rates in order to determine the performance of our scheme un- procedure is the same as in the case of change in perceived
der high network loads. Therefore, we use a perceived band- bandwidth. At the end of the burst duration, the RA will
width update interval based on number of packets. We use again re-negotiate to release the excess CTP. This solution is
a default interval of 100 transmitted packets in our experi- equivalent to splitting up a VBR stream in the time domain
ments, but we also measure how various other intervals affect into multiple CBR streams. There exists previous literature
the performance of the system. in the context of ATM networks [11] in which VBR streams
In case a newly obtained perceived bandwidth value are split into multiple CBR streams in the time domain. Since
NEWBp (f ) differs significantly from Bp (f ), the RA must this scheme only involves re-organizing the traffic rather than
re-negotiate its flows CTP with the BM, as indicated in the the network, it can be directly applied from ATM networks to
example in the previous section. It must also set the value wireless networks.
of perceived bandwidth Bp (f ) to the newly obtained value Figure 5 is an MPEG-4 trace of an hour-long, 25 frames
NEWBp (f ). Note that the RA only sets Bp (f ) to NEWBp (f ) per second, medium-quality, clip of the movie Silence of the
and re-negotiates with the BM using this new value when Lambs. The trace was taken from [10] and the references
there is a significant change, not with every update. A new therein. On the x-axis is a running count of the frame number.
rate using the previously allotted CTP is, however, calculated On the y-axis is the frame size averaged over non-overlapping
with every update. In our experiments, we assume a deviation blocks of 50 frames. One possible way to split up this VBR
= 15% of NEWBp (f ) from Bp (f ) as significant enough to flow into multiple CBR components is shown in figure 5 as
warrant re-negotiation. We also measure how other perceived the contour of the plot. The CBR bandwidth component thus
bandwidth deviation tolerance () percentages affect system obtained is then used as the minimum bandwidth requirement
performance. Bmin (f ) in negotiating with the BM.
If re-negotiation has to be done, the RA of flow f sends Frequent bursts could result in an explosion in re-negotia-
a request message to the BM with flow-id, pmin (f ) and tion overhead. We deal with the problem of frequent bursts
pmax (f ). The values of pmin (f ) and pmax (f ) sent in the in one of two ways: (a) setting Bmin (f ), at the time of burst-
request message are re-calculated using the new value of induced re-negotiation, large enough to engulf multiple bursts
Bp (f ). The rest of the re-negotiation procedure is almost and (b) having large buffering at the receiver to deal with the
identical to the one used for flow establishment, both at the burst.
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- Figure 5. MPEG-4 trace of Silence of the Lambs clip with corresponding
ceive (for nearly 100% of its active duration) at least its min- CBR components.
206 SHAH ET AL.
case. Here we use TCP only to generate bursty cross-traffic Input. Channel time: p_rem; set of requests: p_newmax[f ]
to the UDP flow. We measure and normalize the throughput Output. Set of allocations: p_mm[f ]
of the CBR flow every 2 seconds using the average of packet proc Maxmin(p_rem, p_newmax[f ])
R := {}; //set of satisfied flows
throughputs in the past time window. Our results show that
N := size_of (p_newmax[f ]);
under light channel contention, over 97% of the estimates are
p_mm[f ] := 0;
within 20% of error; under heavy contention, still over 80% of while (true) do
the estimates are within 20% of error. We thus conclude that total_satisfied = 0;
using average throughput of past packets to estimate current foreach f R do
bandwidth is feasible and robust. total_satisfied+ = p_mm[f ];
It should be noted that the bandwidth estimation mecha- od
nism in no way alters the IEEE 802.11 protocol. Our band- CA := (p_rem total_satisfied)/(N size_of (R));
width estimation mechanism, with the normalization exten- stop := true;
sion, was satisfactorily accurate for the scenarios in our foreach f / R do
simulation and testbed experiments. However, the theory if (p_newmax[f ] < CA) then
R := R + {f };
behind the normalization may not be applicable for arbitrar-
p_mm[f ] := p_newmax[f ];
ily large packet sizes or arbitrarily high bit-error rates. In stop := false;
such cases, the TBE could keep an indexed table of sepa- fi
rate estimates for different packet size ranges per active link, od
rather than maintaining a single normalized estimate per ac- if (stop) then
tive link and scaling it to various packet sizes at the time foreach f / R do
of flow establishment/re-negotiation. If the indexed table p_mm[f ] := CA;
method is used, the source and destination must both perform od
total bandwidth estimation, for data and acknowledgements, break;
respectively. The destination must periodically communicate fi
its bandwidth estimate for acknowledgement packets with the od
source using an in-band signaling mechanism. (The signal- Figure 8. Maxmin fair resource allocation algorithm.
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 The computation of the maxmin allocation is as follows.
data and acknowledgement packet sizes. Thus, although the Initially, the set of flows f , whose new maximum channel
indexed table estimation method improves accuracy of the es- time requirement pnewmax (f ) has been satisfied, is empty:
timate in certain special cases, it also incurs a small storage R = . Then, we compute the first-level allotment as
space and in-band signaling overhead. 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
2.5. Maxmin fairness allot each of them p mm (f ) = pnewmax (f ). Next, we com-
pute CA1 = (prem f R pnewmax (f ))/(N R). If for
Fairness is an important issue in designing our Bandwidth all flows g / R, pnewmax (g) CA1 , then we allot each of
Manager. In this paper, we adopt a maxmin fairness algo- them pmm (g) = CA1 and stop. Otherwise, we include those
rithm with minimum guarantee in allotting channel time to flows g with pnewmax (g) < CA1 in set R, allot each of them
the flows. This section describes the maxmin algorithm to pmm (g) = pnewmax (g), and re-compute the next level CA2 .
calculate how much channel time each flow gets beyond its When the algorithm terminates, the allocation pmm (f ) for
guaranteed minimum requested channel time, after the flow all the flows is maxmin fair. The pseudo-code for the al-
is admitted. gorithm is shown in figure 8. It is clear that the computational
In maxmin fairness [4], flows with small channel time complexity of this algorithm is O(N 2 ). As mentioned earlier,
requests are granted their requests first; the remaining chan- after every flow f s pmm (f ) has been determined using the
nel capacity is then evenly divided among the more demand- maxmin algorithm, the BM sets pa (f ) = pmin (f )+pmm (f )
ing flows. As described in section 2.3, pa (f ) is first set to and returns this value to flow f s RA.
pmin (f ) for all the flows. The channel time that remains,
prem , after satisfying the flows minimum requirements, is 2.6. Alternate channel time allocation strategies
allotted to the flows in maxmin fashion. The new maxi-
mum requirement for each flow in the maxmin algorithm Although we use the maxmin fairness with minimum guar-
is pnewmax (f ) = pmax (f ) pmin (f ), because pmin (f ) has antee policy for bandwidth allocation in our implementation,
already been allotted to it and must be subtracted from the a different fairness policy or even a biased, priority-based
original maximum requirement. We denote the channel time scheme could also be used.
allotted to flow f by the maxmin algorithm as pmm (f ). This In our policy, as mentioned earlier, best-effort flows are
is in addition to pmin (f ) allotted before the maxmin algo- only given access to the channel after all the real-time
rithm is even invoked. flows minimum requirements are satisfied. This could lead
208 SHAH ET AL.
to
starvation of the best-effort flows, in the rare case that that abnormally high channel contention has occurred. Subse-
gF pmin (g) 100%. One way to eliminate this prob- quently, the bandwidth manager switches into policing mode
lem would be to partition channel time into a large minimum- to monitor the activity of the network. It may be that the chan-
guarantee portion and a small maxmin fair portion, similar nel contention is due to a sudden increase in physical errors
to the bandwidth partitioning in [1]. The minimum require- or it may be that it is due to a malicious or unmanaged flow.
ments of the real-time flows, i.e., all pmin (g) > 0, are al- The policing scheme can identify which of the above is the
lotted only from the minimum-guarantee portion. The max cause. It could also happen that the unreliable subnet broad-
min fair portion, along with any left over minimum-guarantee cast reply message did not reach a particular RA, so a flow
portion, is used to allot the flows extra CTP pmm (g), using is continuing to transmit packets faster than its re-allotted rate.
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- 3. Experimental results
sures that, however large the minimum requirements of the
real-time flows, some channel time is always available for We evaluate the performance of our Admission Control and
best-effort flows to vie for, so they are never starved. The dis- Dynamic Bandwidth Management system using both a proto-
advantage of having a separate maxmin fair portion is that type testbed as well as simulations using the ns-2 simulator.
the channel time available to satisfy minimum guarantees of We used our testbed when evaluating the performance of a
real-time flows (the minimum-guarantee portion) is reduced, flow in the presence of both physical channel errors caused by
which could lead to more real-time flows being dropped. fading and interference effects as well as medium contention
Another alternate scheme involves pricing of channel time from two other active stations, because there is no way to set
and enforcing priorities based on flow budgets. The max up physical obstacles such as walls, ceilings and doors that
min fair policy with minimum guarantee lends itself to an ele- cause signal weakening in ns-2. We used ns-2 simulations to
gant two-tier pricing scheme. The guaranteed minimum CTP evaluate the performance of the system when there is heavy
pmin (g) is valued at a substantial price, whereas any chan- medium contention due to the presence of a large number of
nel time pmm (g) in excess of this is relatively very cheap. active stations.
Under this two-tier pricing scheme, users would be inclined
to request as little minimum guaranteed bandwidth as possi- 3.1. Simulation experiments
ble, in order to save cost. High minimum requirements are
thus punished while high maximum requirements carry no For experiments with large numbers of nodes (5 nodes)
penalty. The BM adjusts the price so as to trade-off blocking and flows, we used the ns-2 simulator. We compared
probability of the flows with its revenue. If the price is too the performance of an Admission Control and Bandwidth
high, too few flows can afford it and hence blocking probabil- Management-enhanced IEEE 802.11 network (henceforth
ity is high. If the price is low, blocking probability is low, but called enhanced IEEE 802.11 scheme) with an IEEE 802.11
revenue may suffer. Pricing for wireless networks has been network without bandwidth management (henceforth called
studied previously [17,20,22,26], but our two-tier approach is base IEEE 802.11 scheme). We used a 170 m 170 m net-
especially suitable for our bandwidth allocation policy. work area and the transmission range of each node was 250 m.
Hence, the entire network area falls within every nodes trans-
2.7. Policing mission range. The maximum theoretical channel capacity
was 2 Mbps. We used the random waypoint mobility model
In our bandwidth management scheme, policing refers to the with moderate node speeds in our simulations.
task of monitoring the users, to make sure that they conform
to their allocated bandwidth. The bandwidth manager oper- 3.1.1. UDP throughput performance
ates in two modes: normal and policing. When operating in Our first simulation scenario consisted of a 20-node network
policing mode, the bandwidth manager listens promiscuously with 10 flows. Each flow had a minimum bandwidth require-
to the network traffic, and checks whether a flow, identified ment of 100 Kbps and a maximum bandwidth requirement
by the source and destination addresses and port numbers in of 200 Kbps, which are typical of an audio streaming appli-
its packet headers, is sending out packets faster than its al- cation. All the 10 flows used 512 byte packets. The sim-
lotted rate. Additionally, it can also catch those flows who ulation ran for 600 seconds. The transmission rate used by
have not registered with the bandwidth manager. This can be our scheme at any instant was determined using the method
some type of denial of service attack by a malicious users, described in section 2.3. The transmission rate used in the
or caused by some unmanaged applications. base IEEE 802.11 scheme was a constant set to the maximum
Operating in policing mode is expensive. Therefore, the requested rate of the CBR flow, as would be the case in an
bandwidth manager should operate in this mode only when unmanaged application. The RAs inter-update interval was
necessary. To this end, the bandwidth manager relies on the 100 packets and its perceived bandwidth variation-tolerance
sudden, sharp decrease of channel bandwidth as an indica- threshold = 15%, by default.
tion, in the re-negotiation process. If there is a sudden flock of Figure 9(a) is a plot of number of packets successfully
re-negotiation requests due to reduction in Bp (g), it is likely transmitted over every 1 second interval for each of the 10
DYNAMIC BANDWIDTH MANAGEMENT 209
Figure 10. Comparative behavior of a single flow over base 802.11 versus
enhanced 802.11.
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, (b) With smoothing.
the flow that notices an unacceptably poor channel qual- Figure 11. Perceived bandwidth and re-negotiations corresponding to its vari-
ity and complains first is dropped. Alternate flow dropping ation.
strategies can also be employed, such as dropping the flow
last admitted. Pricing could also pay a role here: the flow Figure 11(a) shows the variation of perceived bandwidth
paying the least can be dropped. for one of the flows in the above experiment as measured by
It is clearly evident from the plots that our protocol dra- its TBE at the MAC layer. The superimposed stepwise curve
matically improves throughput fairness. In the base IEEE shows the bandwidth last used for re-negotiation in the above
802.11 scheme, flows often fall far below their minimum experiment. Recall that = 15%. We also experimented with
bandwidth requirement over the 1 second measurement in- smoothed perceived bandwidth estimates, which reduced the
terval, resulting in a chaotic plot. Using our scheme, flows overhead of re-negotiation frequency. Figure 11(b) is a plot
almost never fall below their minimum bandwidth require- of a running average of the measured perceived bandwidth
ment shown with the horizontal line at 24 packets per second. with exponential decay, which is used for smoothing of the
(100 Kbps/4096 bit packets is approximately 24 packets per estimate. The smoothed estimate falls as contention increases
second.) Even when they do, it is only by a small amount. and rises when the two flows are dropped and contention de-
Our scheme thus ensures that the minimum bandwidth re- creases. Other methods to reduce re-negotiation overhead are
quirements of the flows are met with a far higher probability described in the next subsection.
than the base IEEE 802.11 scheme. Figure 10 is a 100-second In section 1, we mentioned that improving fairness is es-
snapshot from the combined plot of figures 9(a) and 9(b) that sential for providing minimum throughput guarantees to wire-
shows the comparative behavior of a single flow (flow 1). less multimedia applications. The key factor enabling our
210 SHAH ET AL.
Table 2 Table 3
Effect of Bp (f ) inter-update period on performance and overhead. Effect of various Bp (f ) variation tolerance levels on performance and over-
head.
Inter-update period FM JM Overhead
(pkts.) (pkts.) (pkts.) (requests/flow/sec.) Tolerance level FM JM Overhead
(%) (pkts.) (pkts.) (requests/flow/sec.)
50 3.62 4.37 0.5
100 4.06 4.66 0.143 10 3.22 4.36 0.333
150 4.15 4.93 0.059 15 4.06 4.66 0.143
200 4.18 5.10 0.019 20 4.89 5.19 0.056
25 5.77 5.37 0.026
ing. This is lower than the 14 seconds for the scenario in the change as, say, a = 25% change, then we can reduce
previous section because the transmission rate is higher and re-negotiation overhead because the RA now waits longer
hence the 100-packet inter-update interval is reached faster. and tolerates more Bp (f ) fluctuation before initiating re-
As in the case of the scenario in the previous subsection, negotiation. Again, this worsens the performance of the sys-
there is a 28% packet drop rate in the case of base IEEE tem because fidelity to bandwidth variations is reduced. Ta-
802.11, but negligible drop rate using our scheme. Also as ble 3 shows how overhead and performance vary with differ-
in the previous scenario, the mean throughput of base IEEE ent levels of tolerance to Bp (f ) variation. The inter-update
802.11 is 15% higher during the period under measurement interval was set to 100 packets for this experiment. Tables 2
(all 3 flows are active) than our scheme. This is because of and 3 both show that for a small price in terms of perfor-
three reasons: (a) the flows are pumping data into the net- mance, we can obtain large gains in overhead reduction.
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 3.1.3. Additional UDP performance results
rate control, (b) our TBE is configured to return a conserva- In this section, we present results for two additional scenar-
tive estimate for Bp (f ), and (c) in the Dynamic Bandwidth ios: (a) when the flows have different minimum bandwidth
Management scheme, the re-negotiation messages between requirements and (b) when the arrival time of the flows is
the various RAs and the BM consume some network band- staggered. We use the 6-node, 3-flow scenario used in the
width. previous section, with the default perceived bandwidth toler-
The conservative Bp (f ) estimate was used to minimize ance of = 15% and the default inter-update interval of 100
packet drop rate. The cost of using such a conservative esti- packets.
mate is that our enhanced IEEE 802.11 scheme under-utilizes Figure 13 shows the comparative base IEEE 802.11 and
the network. Mean throughput is less than it would be under enhanced IEEE 802.11 throughput performance when the 3
full network utilization. However, the TBEs estimate can be flows each have different minimum bandwidth requirements.
suitably tuned so that throughput of our scheme approaches The minimum requirements of the 3 flows are 100 Kbps,
that of the base IEEE 802.11 scheme and network utilization 200 Kbps and 400 Kbps, respectively. The maximum band-
increases. On the other hand, this will also increase the packet width requirement, 600 Kbps, is the same for all 3 flows.
drop rate of our scheme and thereby degrade performance as The plots show that while no guarantee can be made with
packets are dropped randomly from flows. So, there exists a base IEEE 802.11, we can make coarse guarantees with our
trade-off between throughput and packet drop rate. scheme.
In addition to the perceived bandwidth smoothing de- While in all our previous scenarios, all participating flows
scribed in the previous section, we now discuss two other started at around the same time, figure 14 shows the through-
methods to minimize re-negotiation overhead and hence the put performance of the enhanced IEEE 802.11 scheme when
network bandwidth re-negotiation consumes. One method is the start times are staggered. All simulation parameters are
to increase the inter-update period between successive per- identical to those in section 3.1.2, except the staggered start
ceived bandwidth updates from the TBE to the RA. Recall times and the length of the simulation run, which is set to
that we use 100 packets as the default inter-update interval 200 seconds. The bandwidth requirements are identical for
in our experiments. Table 2 shows how overhead and perfor- all 3 flows, as in section 3.1.2. This plot is similar to figure 1
mance vary with different inter-update intervals. As the inter- from [5] and figure 11 from [2], which were for a base-station
update interval increases, some changes in perceived band- network with centralized scheduling. We have produced a
width go undetected and cannot be responded to. Hence, the similar effect for a single-hop ad hoc network that uses the
fairness and throughput jitter worsen while the overhead im- IEEE 802.11 protocols DCF.
proves. The overhead is measured as the frequency of re-
negotiation requests per flow. The threshold tolerance to per- 3.1.4. TCP experiments
ceived bandwidth changes was set at the default of = 15% So far our simulation experiments have focused on multime-
for this experiment. dia applications and UDP flows. In this section we investi-
The other method to reduce re-negotiation overhead is gate the behavior of TCP flows and their interactions with the
to increase the tolerance to changes in perceived bandwidth BM scheme. To this end, we simulate three TCP flows, each
Bp (f ). Recall that we define significant change as a = 15% running between different nodes, in a single-hop ad hoc net-
change in perceived bandwidth. If we define significant work managed by a BM, i.e., using enhanced IEEE 802.11.
212 SHAH ET AL.
Table 4 Table 5
Performance and throughput loss comparison using TCP with interface queue Performance and throughput loss comparison using TCP with interface queue
size smaller than congestion window limit. size larger than congestion window limit.
Base IEEE 802.11 6.70 9.40 565 45065 Base IEEE 802.11 6.53 8.50 0 33804
Enhanced IEEE 802.11 2.12 2.39 33 35698 Enhanced IEEE 802.11 2.51 2.72 0 26577
Figure 16. TCP congestion window behavior when interface queue size is
larger than congestion window limit.
3.2.1. Throughput performance Figure 18 shows the throughput performance for the in-
We conducted two throughput experiments, one indoors and doors scenario. On the x-axis is time in 45 second units. The
one outdoors. In each case, we started some unmanaged ping y-axis shows the adaptation of the audio streaming applica-
sessions, as shown in figure 17, to bring about contention. tion, between 32 and 256 Kbps, to the change in available
The ping ICMP packet transmission on the channel also ar- channel capacity. The channel bit-rate was fixed at 2 Mbps
tificially reduced its bandwidth so that the bandwidth per- at the network cards. The perceived bandwidth variation-
ceived by the audio streaming application actually fluctuated tolerance was set at = 15% and the inter-update interval
between 32 and 256 Kbps depending on the physical errors. was 100 packets. The BM was located on the same machine
In the absence of the pings, the reduction in perceived band- as the sender, 12.0.0.11.
width brought about by the physical errors alone was not suf- The flurry of re-negotiations with the BM on the left-
ficient to cause the audio streaming application to adapt its hand side of the plot corresponds to our moving the sender
quality. The physical errors, at their worst, reduced the per- (12.0.0.11) down to a secluded portion of the basement of the
ceived bandwidth by a few hundreds of Kbps. Given a 2 Mbps building while the receiver (12.0.0.12) and the third laptop
channel and an application with a peak rate of 256 Kbps, (12.0.0.10) remained in the lab on the second floor. While
these errors thus had no effect on the application. Its qual- in the basement, the sender moved around, down narrow cor-
ity level did not fluctuate. To bring about adaptation on the ridors, over staircases and through fire doors. As the level
part of the application, the physical errors had to vary the of fading and interference changed drastically, the perceived
available channel capacity for the application between 32 and channel capacity also changed drastically and hence the flurry
256 Kbps. Hence, we used the pings to contend with the ap- of channel time re-negotiations. The contending pings also
plication for the channel and thus artificially reduce the avail- were affected by the physical errors and produced variable
able channel capacity it perceives to the necessary range. The contention, thus inducing even greater instability in the appli-
pings brought down the available channel capacity to around cation QoS.
500 Kbps so that fading and interference errors could then re- We then brought the sender back to the second floor, the
duce it further below the 256 Kbps threshold needed for the perceived bandwidth returned to around 500 Kbps, and the
application to adapt. quality of the audio returned to its maximum. We then placed
DYNAMIC BANDWIDTH MANAGEMENT 215
the sender and receiver next to each other so that physical er- the start of the experiment, the sender 12.0.0.11 and the re-
rors were rare. The 3 dips in the graph on the right-hand side ceiver 12.0.0.12 were next to each other on the sidewalk of
correspond to experiments with no physical errors, but 3 dif- a street. Then, keeping the receiver 12.0.0.12 stationary on
ferent levels of contention due to 3 different ping rates. All the sidewalk, the sender 12.0.0.11 was moved away by a per-
3 of these ping rates were greater than those used for the first son walking at a normal pace down the street on the sidewalk.
part of this experiment. In the first part, the pings reduced the When the sender was around 150 meters away, the available
available channel capacity to around 500 Kbps and the phys- channel capacity perceived by the audio flow began fluctu-
ical errors dragged it further down. In this part there were ating due to signal fading effects. This resulted in a flurry of
no physical errors, but the larger ping rates themselves took re-negotiations shown in figure 19. The sender then wandered
the available channel capacity below 256 Kbps, causing re- for a while around the point 150 meters away before return-
negotiation from the application. The purpose of this experi- ing to the starting position. As the sender moved closer to the
ment with no physical errors was to demonstrate the effect of receiver, at one point, the available channel capacity returned
the contending ping sessions: they produce a reduction in the to its ping-induced constant level and the application returned
perceived available channel capacity of the managed audio to its highest quality level.
streaming application, in a controlled fashion, and the reduc- We repeated our experiments using ARS (auto rate selec-
tion is a constant one. tion) feature of the wireless card, instead of using constant
Next, we performed another set of experiments outdoors. rates 2 Mbps and 5.5 Mbps mentioned above. Our results
The channel bit-rate was set at 5.5 Mbps for this experiment. were very similar when using ARS as compared to when
As before, we had pings produce contention to artificially re- using fixed rates. We also experimented with the BM at
duce available channel capacity for the audio streaming flow. the destination node, with no change in performance. The
Other parameters such as the value of and the inter-update request-reply delay overhead for re-negotiation requests does
interval were the same as in the indoor experiment. In the not affect performance much because the application paral-
outdoor scenario, we used only two of the laptops. The BM lelly continues transmitting at its previously allotted CTP un-
was once again co-located with the sender, 12.0.0.11. At til the re-negotiation reply arrives, a few milliseconds later.
216 SHAH ET AL.
3.2.2. Request-reply delay In [25], the authors propose an admission control scheme
The request-reply delay is the time delay between the send- for a peer-to-peer, single-hop, ad hoc wireless network model
ing of a request message and the receipt of a reply mes- similar to the one we have used. Their scheme requires the use
sage. All our control messages had a 32-byte payload. This of special probe packets to obtain the service curve, which is
exchange of messages occurs both during flow establishment an estimate of network load. Using the service curve, one-
as well as when perceived bandwidth changes significantly. time admission control is performed. In contrast, our scheme
We set the bandwidth of the network to be 5.5 Mbps, as in estimates network load using the data packets of the connec-
the case of the outdoor experiment. We used all 3 laptops for tion itself. Moreover, we perform dynamic bandwidth re-
the request-reply delay experiments, with 12.0.0.11 being the negotiation over the course of the connection, in addition to
sender, 12.0.0.12 being the receiver and the BM being located admission control at flow startup.
on 12.0.0.10. We found that, if there is no contention, each Another area of related work is the distributed weighted
request-reply round-trip took 23 ms on average. In the pres- fair scheduling (DWFS) schemes in single-hop and multi-
ence of ping-induced contention, each request-reply round- hop wireless networks [3,13,18,19,24]. As mentioned before,
trip took 61 ms on average. Flow establishment occurs only our bandwidth management scheme is required to assist the
once per flow, obviously, and if the perceived bandwidth does DWFS scheme when it is available. At the same time, as
not change much, then the 2060 ms request-reply delay is a shown in our experiments, our scheme also works well with-
small one. 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
4. Related work any DWFS extensions. Therefore, our bandwidth manage-
ment scheme, of which we already have a working prototype,
In this section we discuss two areas of related work: (a) cen- is highly deployable in todays smart-rooms.
tralized channel allocation, and (b) distributed fair scheduling
in single-hop and multi-hop wireless networks.
In wireless network environment, past research has fo- 5. Conclusion
cused on flow scheduling at the access-point to achieve cer-
In this paper, we presented an Admission Control scheme
tain fairness criteria between flows competing for the wireless to determine what fraction of channel time each flow in
channel [5,7,15]. Bianchi et al. [5] proposed the utility fair a single-hop ad hoc wireless network receives. To this
criteria in bandwidth allocation, where each users bandwidth end, we mapped the bandwidth requirement at the applica-
is allocated in such a way that their individual utility is equal- tion/middleware layer to a channel time proportion (CTP)
ized. It assumes that the central manager at the base station requirement at the MAC layer. We presented an applica-
has exact knowledge of the asymptotic utility curves of all the tion/middleware layer rate control mechanism to ensure that
applications, which might be difficult to obtain. The flows in flows conform to their respective CTPs. Since one-time ad-
our scheme can specify a simple linear utility curve using just mission control is not sufficient to handle the changes in net-
two points. In our scheme, the BM guarantees a minimum work and flow characteristics, we also presented a Dynamic
bandwidth for each flow, and allots the rest of the channel Bandwidth Management system that adapts the flows respec-
capacity in a maxmin fashion to each flow up to its maxi- tive CTPs during the course of their operation. The adaptation
mum request. We believe our approach is simple yet effective can be a response to change in the network environment or
in a smart-room where random users walk up to the room and change in a particular flows traffic characteristics. The sim-
share the wireless channel. Another difference is that we use a plicity and robustness of our system enables the incorporation
distributed peer-to-peer transmission (details in section 2.1), of elegant pricing and security features into it. We have devel-
rather than an access-point model, in allocating the channel oped a prototype implementation of the system and we have
resources. used this implementation in a testbed, in addition to extensive
Another wireless network channel allocation scheme is the simulations, to demonstrate the feasibility and utility of our
effort-limited fair scheduling by Eckhardt and Steenkiste [9]. scheme.
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- References
ing other best-effort flows. The usage of air time to measure
[1] G. Ahn, A. Campbell, A. Veres and L. Sun, Swan: Service differen-
the bandwidth requirement of a flow is similar to the chan- tiation in stateless wireless ad hoc networks, in: Proceedings of IEEE
nel time in our scheme. However, it is unclear how the power InfoCom, New York (June 2002).
factor can be chosen for different flows because this will give [2] O. Angin, A. Campbell, M. Kounavis and R. Liao, The mobiware
preferential treatment to certain users. In our scheme, when toolkit: Programmable support for adaptive mobile networking, IEEE
a flows minimum requirement cannot be satisfied, the flow Personal Communications, Special Issue on Adapting to Network and
Client Variability 5(4) (1998) 3244.
is simply rejected. This creates incentive for the users to set [3] B. Bensaou, Y. Wang and C. Ko, Fair medium access in 802.11 based
a minimum channel time requirement as small as possible to wireless ad hoc networks, in: Proceedings of IEEE MobiHoc, Boston,
reduce the possibility of being denied access to the channel. MA (August 2000).
DYNAMIC BANDWIDTH MANAGEMENT 217
[4] D. Bertsekas and R. Gallager, Data Networks, 2nd ed., chapter 6 [24] N. Vaidya, P. Bahl and S. Gupta, Distributed fair scheduling in a wire-
(Prentice-Hall, 1992). less LAN, in: Proceedings of ACM MobiCom, Boston, MA (August
[5] G. Bianchi, A. Campbell and R. Liao, On utility-fair adaptive services 2000).
in wireless networks, in: Proceedings of IEEE/IFIP IWQoS, Napa, CA [25] S. Valaee and B. Li, Distributed call admission control in wireless ad
(May 1998). hoc networks, in: Proceedings of IEEE VTC, Vancouver, Canada (Sep-
[6] F. Cali, M. Conti and E. Gregori, Dynamic tuning of IEEE 802.11 pro- tember 2002).
tocol to achieve a theoretical throughput limit, IEEE/ACM Transactions [26] Y. Xue, B. Li and K. Nahrstedt, Price-based resource allocation in wire-
on Networking 8(6) (2000) 785799. less ad hoc networks, in: Proceedings of IEEE IWQoS, Monterey, CA
[7] T. Chen, P. Krzyzanowski, M. Lyu, C. Sreenan and J. Trotter, A sum- (June 2003).
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 Samarth Shah received his B.E. degree in com-
MobiCom, Seattle, WA (August 1999). puter science and engineering from the University of
[9] D. Eckhardt and P. Steenkiste, Effort-limited fair (ELF) scheduling for Madras, India, in 1998. He is currently a Ph.D. can-
wireless networks, in: Proceedings of IEEE InfoCom, Tel Aviv, Israel didate in the Department of Computer Science at the
(March 2000). University of Illinois at Urbana-Champaign. His in-
[10] F. Fitzek and M. Reisslen, Mpeg-4 and h.263 video traces terests include quality of service (QoS) in wireless
for network performance evaluation, IEEE Network Magazine networks.
15(6) (2001) 4054, http://www-tkn.ee.tu-berlin.de/ E-mail: shshah@cs.uiuc.edu
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). Kai Chen received his B.Eng. degree in computer
[12] E. Guttman, C. Perkins, J. Veizades and M. Day, Service location pro- science from Tsinghua University, Beijing, China, in
tocol, version 2, RFC 2608 (June 1999). 1995, and M.S. degree in computer science from the
[13] V. Kanodia, C. Li, A. Sabharwal, B. Sadeghi and E. Knightly, Distrib- University of Delaware, Newark, Delaware, USA, in
uted multi-hop scheduling and medium access with delay and through- 1998. Currently he is a Ph.D. candidate in the De-
put constraints, in: Proceedings of ACM MobiCom, Rome, Italy (July partment of Computer Science at the University of
2001). Illinois at Urbana-Champaign. From 1998 to 2000,
[14] M. Kazantzidis, M. Gerla and S. Lee, Permissible throughput network he worked as a research programmer at the National
feedback for adaptive multimedia in AODV Manets, in: Proceedings of Center for Supercomputing Applications (NCSA) at
IEEE ICC, Helsinki, Finland (June 2001). Urbana, IL, USA. His research interests include mo-
[15] J. Kim and M. Krunz, Bandwidth allocation in wireless networks with bile ad hoc networks, transport layer issues in mobile networks, quality of
guaranteed packet-loss performance, IEEE/ACM Transactions on Net- service, incentive engineering, and pervasive computing.
working 8(3) (2000) 337349. E-mail: kaichen@cs.uiuc.edu
[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- Klara Nahrstedt is an Associate Professor at the
less LAN based access networks, in: Proceedings of IEEE ICNP, Paris, University of Illinois at Urbana-Champaign, Com-
France (November 2002). puter Science Department. Her research interests are
[18] H. Luo, S. Lu and V. Bharghavan, A new model for packet schedul- directed towards multimedia middleware systems,
ing in multihop wireless networks, in: Proceedings of IEEE MobiCom, Quality of Service (QoS), QoS routing, QoS-aware
Boston, MA (August 2000). resource management in distributed multimedia sys-
[19] H. Luo, P. Medvedev, J. Cheng and S. Lu, A self-coordinating approach tems, and multimedia security. She is the coauthor of
to distributed fair queuing in ad hoc wireless networks, in: Proceedings the widely used multimedia book Multimedia: Com-
of IEEE InfoCom, Anchorage, AK (April 2001). puting, Communications and Applications published
[20] P. Marbach and R. Berry, Downlink resource allocation and pricing for by Prentice Hall, the recipient of the Early NSF Ca-
wireless networks, in: Proceedings of IEEE InfoCom, New York (June reer Award, the Junior Xerox Award, and the IEEE Communication Soci-
2002). ety Leonard Abraham Award for Research Achievements. She is the editor-
[21] M. Mirhakkak, N. Schult and D. Thomson, Dynamic bandwidth man- in-chief of ACM/Springer Multimedia Systems Journal, and the Ralph and
agement and adaptive applications for a variable bandwidth wireless Catherine Fisher Associate Professor. Klara Nahrstedt received her B.A. in
environment, IEEE JSAC 19(10) (2001) 19841997. mathematics from Humboldt University, Berlin, in 1984, and M.Sc. degree
[22] Y. Qiu and P. Marbach, Bandwidth allocation in ad-hoc networks: in numerical analysis from the same university in 1985. She was a research
A price-based approach, in: Proceedings of IEEE InfoCom, San Fran- scientist in the Institute for Informatik in Berlin until 1990. In 1995 she re-
cisco, CA (MarchApril 2003). ceived her Ph.D. from the University of Pennsylvania in the Department of
[23] The network simulator ns-2, http://www.isi.edu/nsnam/ns/, Computer and Information Science.
updated October 2001. E-mail: klara@cs.uiuc.edu
Mobile Networks and Applications 10, 219231, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.
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 keys. This RAM requirement does not exist in the implemen-
tations of encryption algorithms, which compute their keys
Security is a primary requirement of any wired and wireless during the encryption/decryption operation. Cellular phones
communication. Encryption algorithms are meant to provide technology demands specific characteristics of the cryptog-
secure communications applications. However, if the sys- raphy science. Ciphers have to be compatible with wireless
tem is not designed property, it may fail. New encryption devices restricted standards in hardware resources.
algorithms have to perform efficiently in a variety of cur- Data-Dependent (DD) Permutations (DDP) performed
rent and future applications, doing different encryption tasks. with so called Controlled Permutation (CP) boxes [6,14,16]
All hardware implementations have to be efficient, with the appears to be very efficient cryptographic primitive for fast
minimum allocated number of logic gates. This means sim- hardware encryption. Security estimation of the DDP-based
plicity in ciphers architectures with enough clever data ciphers CIKS-1 [11] and SPECTR-H64 [10] against linear
transformation components. The implementation of a com- cryptanalysis has shown that DDP are efficient, provided they
munication protocol, demands low power devices and fast are combined with other non-linear operations. The DDP-
computation components which imply that the number and based ciphers are proposed for hardware implementation with
complexity of the encryption operations should be kept as low cost.
simply as possible. A basic transformation in the operation of In this paper we present two new DDP-based ciphers
todays ciphers is needed, including transformations in both Cobra-H64 and Cobra-H128 and the results of their hard-
data and key blocks size. The ciphers of the near future have ware implementations. The design of the presented ciphers
to be key agile. Many applications need a small amount of takes into account some recommendations arising from the
text to be encrypted with keys that are frequently changed. linear and differential analysis of other DDP-based ciphers
Many well-known applications, like IPsec, use this way of ci- [5,10,11]. Both proposed ciphers Cobra-H64 and Cobra-
phers operation. Although the most widely used mode of op- H128 have been implemented in ASIC and FPGA hardware
eration is encryption with the same key for all the amount of modules. Two different VLSI architectures are examined for
transport data, the previous mode is also very useful for future each one of the proposed ciphers. The synthesis results of all
applications. Ciphers that requiring subkeys precomputation hardware integrations are presented in detail.
have a lower key agility due to the precomputation time, and The paper is organized in the following way. In section 2
they also require extra RAM to hold the precomputed sub- we consider construction of the controlled operational boxes
220 SKLAVOS ET AL.
Figure 2. Structure of boxes: (a) P8/12 and (b) P1 Figure 4. Structure of the CP-boxes: (a) P64/192 and (b) P1
64/192 .
8/12 .
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
(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 223
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)).
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 and Cobra-H128 are fast in the case of key refreshing, since
as concatenation of eight bytes Xi (Yi ), i {1, . . . , 8}, and they are free of external key scheduling.
Y1 = X64 , Y2 = X54 , Y3 = X44 , Y4 = X34 , The ciphers Cobra-H64 and Cobra-H128, in comparison
Y5 = X24 , Y6 = X14 , Y7 = X84 , Y8 = X74 . with the prototype SPECTR-H64, have the following fea-
Use of this permutational involution provides that in one tures:
(V )
round each input bit of the box P64/192 influences all output 1. They use the initial secret key in each round.
bits of the box (P1
64/192)
(V ) improving statistical properties
2. The round transformation includes fixed permutational in-
of the round function. Permutation I introduces no time delay
volution, performed on one of the data subblocks.
since it is implemented in hardware as simple connections of
wiring. 3. The round transformation includes one switchable opera-
tion that prevents the weak keys with the structure K =
4.3. Non-linear operation G (X, X, X, X).
The role of the fixed permutation in procedure Crypt(e) The ciphers Cobra-H64 and Cobra-H128 have very sim-
is to make each bit of the left data subblock influences more ilar structure of the procedure Crypt(e). The differences
different digits of R while combining outputs of two opera- are based on the use of (i) the operations having different
tions G with R. The permutation contains four cycles of size of input (32-bit for Cobra-H64 and 64-bit for Cobra-
the length 16 and is described as follows: H128), (ii) different switchable operations: switchable per-
mutation (e) in Cobra-H64 and e-dependent swapping in
(1, 50, 9, 42, 17, 34, 25, 26, 33, 18, 41, 10, 49, 2, 57, 58) Cobra-H128, (iii) the use of the subkeys to calculate control-
(3, 64, 43, 24, 19, 48, 59, 8, 35, 32, 11, 56, 51, 16, 27, 40) ling vectors in Cobra-H128, while in Cobra-H64 the control-
(4, 7, 28, 47, 52, 23, 12, 63, 36, 39, 60, 15, 20, 55, 44, 31) ling vectors depend only on the current value of the left data
(5, 14, 13, 6, 21, 62, 29, 54, 37, 46, 45, 38, 53, 30, 61, 22). subblock.
The transformation W = GA A (L) defining the operation
5.2. Security estimations
GA A is described as follows:
W = L0 A0 L1 A0 L2 L5 L6 A1 A2 A1 We have considered different variants of the differential crypt-
analysis. We have obtained that the fewer active bits in the
L3 L4 L1 L4 L6 L2 L6 A1 L1 L2 L4 A1 ,
difference the higher the probability of the differential charac-
where j {0, 1, . . . , 6} we have Lj = L64j and A0 = teristic. This corresponds to the results of the analysis of other
A, A1 = (1, a1 , . . . , a63 ), A2 = (1, 1, a1 , . . . , a62 ), (A = A DDP-based ciphers. Our best attack against Cobra-H64 and
or A = A ). 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
5. Discussion Cobra-H128. Probability of the best two-round characteristic
of SPECTR-H64 is p(2) = 1.15 213 . Minimal number of
5.1. Peculiarities of Cobra-H64 and Cobra-H128 rounds required to thwart differential attacks is 8 for Cobra-
H64 and 10 for Cobra-H128 and SPECTR-H64.
In ciphers Cobra-H64 and Cobra-H128 extensive use of the Our preliminary linear analysis of Cobra-H64 and Cobra-
CP-box operations is applied. They are used in three differ- H128 has shown that they are secure against linear attacks for
ent ways: (i) as DDP that are one of two basic cryptographic number of rounds r 5. Accordingly to [10] SPECTR-H64
primitives, (ii) to swap subkeys when changing encryption is secure against linear attack for values r 6. High degree of
mode for decryption one, and (iii) to switch permutation (e) the algebraic normal form and the complexity of the Boolean
when changing ciphering mode. In addition to DDP, these function describing round transformation of the developed ci-
ciphers use the non-linear operation G. The round transfor- phers prevent the interpolation and high order differential at-
mation of both ciphers is characterized by high parallelism tacks. Our statistic experiments has shown that 4 round of
that provides pre-requisites of their high performance. Due to Cobra-H64 and 5 round of Cobra-H128 are sufficient to sat-
the use of very simple key scheduling the ciphers Cobra-H64 isfy test criteria proposed for the AES finalists [18].
226 SKLAVOS ET AL.
Because of the very simple key scheduling used in the pro- pipelining data transformation, or parallel processing. On
posed ciphers it appears to be important to study how a single the other hand, the internal structure of the microprocessors
bit of key statistically influences ciphertext (keys propagation functional units limits the parallel processing and pipelining
property). For this purpose we have used the criteria of [18] transformation. In addition, the instruction parallelism level
considering the secret key as input vector and fixing different is a factor of great importance that must be taken under con-
plaintexts. Such statistic testing has shown that five rounds sideration for microprocessor performance. Furthermore, the
of Cobra-H64 and six rounds of Cobra-H128 are sufficient to hardware devices of these types can operate on arbitrary size
satisfy test criteria. of words, in contrast with processors, that operate only on
The used key scheduling is secure against basic related- fixed-sized words.
key attacks. In spite of the simplicity of the key schedule ASICs are in general far more expensive devices due to
the keys K = (X, Y, X, Y ) or K = (X, X, X, X), where time consuming and high cost fabrication procedure, which
X, Y {0, 1}32 for Cobra-H64 or X, Y {0, 1}64 for Cobra- is done by expertise industry departments. FPGAs can be
H128, are not weak, since encryption and decryption require reached from anyone, since they are enough cheaper and can
changing the parameter e. It seems to be difficult to calculate be programmed or reconfigured by the designers/researchers.
a semi-weak key-pair for presented ciphers, if it is possible FPGAs have the major advantage that can perform a com-
at all. Thus, the role of the switchable permutation (e) is pletely different task/function after simple designers recon-
preventing the weak keys. For example, for SPECTR-H64 figuration. ASICs performance is tight and can not be mod-
which uses no switchable operations for all X the 256-bit key ified after the chips fabrication. The offered reconfiguration
K = (X, X, X, X, X, X, X, X) is weak. 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-
6. Hardware implementation duced by the dedicated circuits parts needed to reconfigura-
tion. In general, any implemented system of digital logic in
6.1. ASIC and FPGA devices an FPGA, is slower than ASIC implementation of the same
system.
Hardware implementations of both proposed ciphers are de-
signed and coded in VHDL hardware description language. 6.2. Implementation architectures for Cobra-H64 and
Both Cobra-H64 and Cobra-H128 were implemented using Cobra-H128
two complete different implementation hardware modules:
Application Specific Integrated Circuit (ASIC) and Field Pro- Both Cobra-H64 and Cobra-H128 are examined in hardware
grammable Gate Array (FPGA). The performance character- implementation by using two different architectures: Full
istics of both ASICs and FPGAs are substantially different Rolling and Pipeline for both ASIC and FPGA devices. The
compared with a general-purpose microprocessor. ASICs and used Full Rolling architecture is shown in figure 11(a). It is
FPGAs have the advantage that can use all the resources for a typical architecture for secret key block cipher implementa-
tion. This architecture operates efficiently for both encryption 6.3. VLSI implementations synthesis results
and decryption process. According to this architecture only
one block of plaintext/ciphertext is transformed at a time. The The synthesis results are shown in table 4, for Cobra-H64
necessary number of clock cycles to encrypt/decrypt a data and Cobra-H128, for both hardware implementations (ASIC
block is equal to the specified number of cipher rounds (10 and FPGA), where: D Flip-Flops (DFFs), Configurable Logic
for Cobra-H64 and 12 for Cobra-H128). The key expansion Blocks (CLBs), Function Generators (FGs).
The above synthesis results for both implementations
unit produces the appropriate round keys which are stored and
(ASIC and FPGA) prove that the Pipeline architectures of
loaded in the used RAM blocks. One round of the encryption
Cobra-H64 and Cobra-H128 have very high speed perfor-
algorithm is performed by the Data Transformation Round
mance. Especially Cobra-H64 throughput is up to 5.5
Core. This core is a flexible combinational logic circuit and it
and 7.1 Gbps for FPGA and ASIC implementation, respec-
is supported by a n-bit register and n-bit multiplexer (64-bit tively. Cobra-H128 throughput reaches the values of 11 and
for Cobra-H64 and 128-bit for Cobra-H128). In the first clock 12.1 Mbps for the same implementation devices. On the other
cycle, the n-bit plaintext/ciphertext is forced into the Data hand, Full Rolling architectures for both proposed ciphers
Transformation Round Core. Then in each clock cycle, one allocate minimized area resources with good data rate. Of
round of the cipher is performed and the transformed data are course, for the Full Rolling architectures, of both Cobra-H64
stored into the n-bit register. According to Full Rolling archi- and Cobra-H128, the main goal is the minimized allocated
tecture a 64-bit data block is completely transformed every 10 area resources, with good achieved throughput. The opera-
clock cycles for Cobra-H64 (10 transformation rounds). The tion frequency for both ciphers is very high, for both proposed
operation of Cobra-H128 (12 transformation rounds) needs architectures and for all the examined hardware modules. Es-
12 clock cycles in order a 128-bit plaintext/ciphertext to be pecially, Cobra-H64 frequency ranges between 82 MHz to
generated 110 MHz. Cobra-H128 operates up to 90 MHz for FPGA
The second proposed architecture, figure 11(b), is a devices and up to 95 MHz for ASIC approaches. It is obvi-
N-stage pipeline architecture. The main characteristics of this ous that according to the applications major demands, area
are: (i) the pipelining used technique, and (ii) the usage of a or performance, the designer can use the Full Rolling or the
RAM for the round keys storage and loading, which are pre- Pipeline architecture respectively, for each one of the two pro-
computed. Pipelining is not possible to be applied in many posed ciphers. The kind of application and the characteristics
cryptographic applications. However, Cobra-H64 and Cobra- of the application itself, will determine the use of the FPGA
or the ASIC as the integration device, for the implementa-
H128 block ciphers structures provide the availability to be
tion either Cobra-H64 or Cobra-H128. Both Cobra-H64 and
implemented with pipelining technique. The pipelining ar-
Cobra-H128 architecture simplicity in addition to the offered
chitecture offers the benefit of the high-speed performance.
high-level security strength make their hardware integration
The implementation can be applied in applications with hard
very useful in wireless communications networks.
throughput needs. This goal is achieved by using a number In order to evaluate the very good performance, in hard-
of operating blocks with a final cost to the covered area. The ware terms, of both Cobra-H64 and Cobra-H128 we com-
proposed architecture uses 10 basic round blocks for Cobra- pare the proposed ciphers implementations with the encryp-
H64 and 12 basic round blocks for Cobra-H128, which are tion algorithms that are used in todays wireless protocols.
cascaded by using equal number of pipeline registers. Based Especially in figures 12 and 13, the proposed ciphers perfor-
on this design approach, 10 and 12 different n-bit data blocks mance is compared with the best hardware implementations
can be processed at the same time, for Cobra-H64 and Cobra- of ciphers used in wireless communications protocols. In
H128, respectively. Pipeline proposed architecture produces IEEE 802.11 first versions (ad), the Wired Equivalent Pri-
a new plaintext/ciphertext block every clock cycle. vacy (WEP) is widely used to ensure privacy in the trans-
Table 4
Implementation synthesis results.
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
and RC5 are the alternative ciphers that can be used for bulk References
encryption in WTLS.
For the FPGA implementation approach as it is shown [1] V.E. Benes, Mathematical Theory of Connecting Networks and Tele-
phone Traffic (Academic Press, New York, 1965).
in figure 12, Full Rolling (FR) architectures of both Cobra-
[2] O.Y.H. Cheung, K.H. Tsoi, P.H.W. Leong and M.P. Leong, Tradeoffs in
H64 and Cobra-H128 has minimized area resources com- parallel and serial implementations of the international data encryption
pared with FPGA approaches of AES [13], RC4 [7], SAFER+ algorithm, in: Proceedings of CHES 2001, Lecture Notes in Computer
[9], IDEA [2], and DES [8] (for the last ciphers we present the Science, Vol. 2162 (Springer, 2001) pp. 333337.
best results corresponding to different implementation archi- [3] C. Clos, A study of nonblocking switching networks, Bell System
tectures). The pipeline architectures (P)s of Cobra-H64 and Technical J. 32 (1953) 406424.
[4] A.P. Fournaris, N. Sklavos and O. Koufopavlou, VLSI architecture and
Cobra-H128, as it was expected needs more resources than FPGA implementation of ICE encryption algorithm, in: Proceedings of
then (FR)s. The main advantage of pipeline architectures, as 10th IEEE International Conference on Electronics, Circuits and Sys-
it is proven from figure 12, is the very high speed perfor- tems (ICECS03), United Arab Emirates, December 1417 (2003).
mance, compared with the achieved throughput of the other [5] N.D. Goots, B.V. Izotov, A.A. Moldovyan and N.A. Moldovyan, Mod-
conventional ciphers [2,79,13]. ern Cryptography: Protect Your Data with Fast Block Ciphers (A-LIST
Publishing, Wayne, 2003) 400 p.,
In addition, the proposed architectures are compared with http://www.alistpublishing.com
other ASIC implementations [19,2123] in figure 13. For the [6] N.D. Goots, A.A. Moldovyan and N.A. Moldovyan, Fast encryption al-
ASIC approaches, a covered area comparison is not efficient gorithm SPECTR-H64, in: Proceedings of the International Workshop,
because of the variety in units (mm2, sqmil, gates, transistors) Methods, Models, and Architectures for Network Security 2001, Lec-
of the reported area of the conventional architectures [19,21 ture Notes in Computer Science, Vol. 2052 (Springer, 2001) pp. 275
286.
23]. Both Cobra-H64 and Cobra-H128 higher performance
[7] P. Hamalainen, M. Hannikainen, T. Hamalainen and J. Saarinen, Hard-
compared with AES [21], IDEA [23], RC5 [19], and DES ware implementation of the improved WEP and RC4 encryption al-
[22], in all of the cases. gorithms for wireless terminals, in: Proceedings of the European Sig-
Furthermore, for the FPGA implementations, all the nal Processing Conference (EUSIPCO2000), September 58, Finland
proposed Cobra ciphers architectures are compared in the (2000).
[8] J. Kaps and C. Paar, Fast DES implementations for FPGAs and its ap-
Area/Performance model. This model can easily be calcu-
plication to a universal key-search machine, in: Proceedings of 5th An-
lated, by the equivalent coved area resources (in CLBs) and nual Workshop on Selected Areas in Cryptography (SAC 98), August
the throughput (Mbps), by using the formula: Area/Perfor- 1718, Canada (1998).
mance model = Area/Performance. The Area/Performance [9] P. Kitsos, N. Sklavos, K. Papadomanolakis and O. Koufopavlou, Hard-
model comparison is illustrated in figure 14. It is obvious that ware implementation of the Bluetooth security, IEEE Pervasive Com-
the FRs architectures of both Cobra-H64 and Cobra-H128 are puting, Mobile and Ubiquitous Systems 2(1) (2003) 2129.
[10] Y. Ko, D. Hong, S. Hong, S. Lee and J. Lim, Linear cryptanalysis on
much better compared with RC4 [7], SAFER+ [9], IDEA [2], SPECTR-H64 with higher order differential property, in: Proceedings
and DES [8], while they are too close to AES [13] value. of International Workshop, Methods, Models, and Architectures for
Cobra-H64 (P) architecture as well as Cobra-H128 (P) Area- Network Security 2003, Lecture Notes in Computer Science, Vol. 2776
Delay model have higher value than FRs architectures. This (Springer, Berlin, 2003) pp. 298307.
is the paid penalty of the very good performance (figure 12) [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
compared with the other conventional ciphers. 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.
7. Conclusions No. 30 (1999).
[13] M. McLoone and J.V. McCanny, High performance single-chip FPGA
In this work, we propose two new fast ciphers: Cobra-H64 Rijndael algorithm implementation, in: Proceedings of CHES 2001,
and Cobra-H128. These ciphers are based on DDP transfor- Lecture Notes in Computer Science, Vol. 2162 (Springer, 2001) pp. 65
mations. Security analysis has shown that the both ciphers 76.
are secure against known attacks. Due to high parallelism of [14] A.A. Moldovyan, Fast block ciphers based on controlled permutations,
Computer Science Journal of Moldova 8(3) (2000) 270283.
computations in one round and the use of the switchable op- [15] A.A. Moldovyan and N.A. Moldovyan, A method of the cryptograph-
erations one can use very simple key scheduling that makes ical transformation of binary data blocks, Russian patent # 2141729,
hardware implementation cheaper and faster in the case of Bull. No. 32 (1999).
frequent change of keys. Both ciphers achieve high-speed [16] A.A. Moldovyan and N.A. Moldovyan, A cipher based on data-
performance in FPGA devices and especially for ASIC ap- dependent permutations, Journal of Cryptology 15(1) (2002) 6172.
[17] M. Portz, A generalized description of DES-based and benes-based
proaches. The implementation cost and the performance of
permutation generators, in: Advances in Criptology, Lecture Notes in
the proposed ciphers are compared with the security layers of Computer Science, Vol. 718 (Springer, 1992) pp. 397409.
the most widely used wireless protocols: IEEE 802.11, Blue- [18] B. Preneel, A. Bosselaers, V. Rijmen, B. Van Rompay, L. Granboulan,
tooth, WAP, OMA, and UMTS. These comparisons present J. Stern, S. Murphy, M. Dichtl, P. Serf, E. Biham, O. Dunkelman,
the advantages of the proposed ciphers in terms of area re- V. Furman, F. Koeune, G. Piret, J.-J. Quisquater, L. Knudsen and
H. Raddum, Comments by the NESSIE project on the AES finalists,
sources, operating frequency, and throughput. These advan-
May 24 (2000), http://www.nist.gav/aes
tages prove the suitability of the proposed ciphers in wireless [19] A. Schubert and W. Anheier, Efficient VLSI implementation of modern
communications, which set hard specifications in security im- symmetric block ciphers, in: Proceedings of ICECS99, Cyprus (1999).
plementations. [20] B. Van Rompay, L.R. Knudsen and V. Rijmen, Differential cryptanaly-
HIGH SPEED NETWORKING SECURITY 231
sis of the ICE encryption algorithm, in: Proceedings of the 6th Interna- Nikolay A. Moldovyan is an honoured inventor
tional Workshop Fast Software Encryption FSE98, Lecture Notes in of Russian Federation (2002), a chief researcher
Computer Science, Vol. 1372 (Springer, 1998) pp. 270283. with the Specialized Center of Program Systems
[21] B. Weeks, M. Bean, T. Rozylowicz and C. Ficke, Hardware perfor- SPECTR, and a Professor with the Saint Pe-
mance simulations of round 2 advanced encryption standard algo- tersburg Electrical Engineering University. His
rithms, in: Proceedings of 3rd Advanced Encryption Standard (AES) research interests include computer security,
Candidate Conference, New York, USA, April 1314 (2000). cryptography, and currently developed concept of the
[22] D.C. Wilcox, L.G. Pierson, P.J. Roberston, E.L. Witzke and K. Gass, variable transformations as a new direction in ap-
A DES ASIC suitable for network encryption at 10 Gbps and bey- plied cryptography. He received his Diploma and
oned, in: Proceedings of CHES99, Lecture Notes in Computer Sci- Ph.D. in Academy of Sciences of Moldova (1981).
ence, Vol. 1717 (Springer, 1999) pp. 3748. He is a member of the IACR.
[23] R. Zimmermann, A. Curiger, H. Bonnenberg, H. Kaeslin, N. Felber E-mail: nmold@cobra.ru
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.
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)
chronization. We investigate four packet scheduling algo- synchronization [24], network-based schemes [28], buffering-
rithms: First-In-First-Out (FIFO), Priority-Queuing (PQ), and based schemes [4], and reactive control schemes have been
Round-Robin (RR) and Weighted Fair-Queuing (WFQ) queu- studied over the past few years. Network synchronization
ing. In todays wireless and wired communication systems techniques have been used to improve the quality of service,
FIFO is used because of the simplicity of the algorithm [14]. in terms of multimedia synchronization for smooth play-out
PQ is an priority-based scheduling algorithm useful for the of audio and video, at the mobile client. In our earlier work,
transfer of real-time traffic but usually produces unfairness we have discussed MoSync, an algorithm [8] is the only algo-
among traffic classes [14]. RR scheduling creates fairness by rithm that deals with continuous media in high layers, partic-
given each packet a certain amount of play-out time. WFQ ularly audio and video at the application layer. The primary
scheduling, at different network layers, gives fair sharing of difference between packet-level performance and continuous-
bandwidth among various network traffic. 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
2. Related work (wired/wireless), extensive research has been done on the
effect of packet scheduling synchronization in the network.
Media synchronization control is paramount for multimedia
Earlier network designs (wireless and wired) were meant for
applications such as video and audio stream over wired/wire-
non-delay affected applications such as text and graphics.
less networks. Media synchronization control is necessary
The new paradigm for, both wired and wireless networks,
for preserving the temporal relationships among plural media
are resource intensive. Applications like animated graphics,
streams by compensating for network delay jitters [13]. Me-
on-demand video and voice-over IP require bounded delays
dia synchronization can be classified as intra-stream synchro-
nization and inter-stream synchronization [26]. The intra- and guarantee throughput, but tolerate some errors [19]. The
stream synchronization refers to the temporal constraints growth of smart phones, PDAs and laptop computers have
within a single stream such as the time intervals between two increase the need for unification between wired and wire-
successive packets of a video or audio stream. Inter-stream less networks. However, like all technological advancement,
synchronization pertains to synchronization among multiple new technology invents new problems, particularly synchro-
streams of concurrent packets such as voice and audio stream nization with multimedia packets, being played out on the
(lip sync). To ensure quality of service of multimedia services mobile client. The combination of wired and wireless net-
and to minimize end-to-end multimedia packet delay for the works differ greatly from the typical wired networks. First, all
mobile client, intra-stream synchronization and inter-stream packets must pass through a base station before reaching its
synchronization must be guaranteed as illustrated in figure 1. mobile terminal destination [8]. Second, a mobile clients re-
Research has been done in the area of QoS packet schedul- sources (battery, memory, screen, bandwidth, etc.) are very
ing. Most of the research and publications deal with network limited. Due to resource limitations, conventional synchro-
level performance like packet overhead, throughput and de- nization strategies for delivery of multimedia packets to the
lay. For overall QoS of the mobile client, several areas such mobile client, cannot be applied in a mixed environment.
as scheduling for synchronization, feedback techniques for 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
Figure 1. Intra-stream and inter-stream synchronization. multimedia systems that uses a Quasi-sink to control synchro-
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 235
nization. The proposed solution copes with network jitters, client and last, as a Quasi-receiver, it only receives the first
end-system jitters, clock drift and changing network condi- packets from the servers. When a mobile client request mul-
tions [4,20]. MoSync can be employed for both intra-stream timedia packets from a server the multimedia packets are sent
during playback video and resynchronization streams in a via the base station. The servers can control the supply rate
wireless network. Since MoSync uses delay time to allow of multimedia packets to mobile clients by using the mobile
intra-synchronization, inter-synchronization is guaranteed by clients feedback messages. MoSync focuses on three areas to
the MoSync algorithm [8]. A set of experiments was done solve the combination of wired and wireless communication
earlier assuming FIFO play-out of the multimedia packets on synchronization problems:
the mobile client from the base station. In this paper, we study
1. No jitters and constant delay case (extending Biersack and
the hand-off problem and propose tow hand-off schemes to
Geyer [4]).
enhance the performance of the media synchronization for
wireless systems, as well as investigate the multimedia packet 2. Intra-stream and inter-stream synchronization problem
play-out onto the mobile client in RR, PQ, and WFQ from the (using network load).
base station.
3. Resynchronization (using exponential smoothing forecast-
ing).
3. MoSync algorithm MoSync uses three simple protocols. In case one, to avoid
buffering of early arrival multimedia packets sub-stream the
The MoSync algorithm [8] assumes that the network can pro- start-up protocol is used. This allows evaluation of the round
vide sufficient resources to deliver multimedia packets to the trip delay for different sub-streams. In case two, end-system
mobile client. The multimedia data consists of many mul- jitters and network jitters is counteracted. Both end-system
timedia packets that may be progressively transmitted over and network jitters must be countered to avoid underflow or
the network. MoSync does not use a global synchronization overflow at the mobile client. For multimedia packets to play-
clock. In the MoSync algorithm, the servers time stamp each out smoothly the servers data transmission rate must match
multimedia packet, with current local time to allow the base the available service rate at the mobile client. The overall
station to calculate round trip delay, jitters and inter-arrival goal when dealing with network and end-system jitters, is to
time. The multimedia packets carries the servers number and keep the average delay, buffer underflow and buffer overflow
packet sequence number but these additions are negligible so at an acceptable minimal level. In case three, to solve the
multimedia synchronization is the primary focus of the this average delay changes, clock drift and server dropouts expo-
paper. nential smoothing forecasting is used [8]. When a mobile
The MoSync model consists of K scalable server nodes client receives a multimedia packet, the mobile client calcu-
and L mobile clients. A base station can communicate with lates the expected arrive time of the next multimedia packet
multiple mobile clients because the communication between and sends the calculation to the base station. The base station
the base station and mobile client is wireless. Figure 2 illus- then forwards the information to the servers. In both case two
trates the servers, base station and mobile client connectivity. and three the MoSync algorithm uses either the pessimistic or
The system can contain multiple servers (data, voice, graph- optimistic synchronization protocol when requesting and re-
ics, video) depending upon the need of the mobile client. All ceiving multimedia packets from the servers. With MoSyncs
mobile clients receives the multimedia services from the same pessimistic protocol, the base station updates the servers on
servers. The communicate is done from the the servers to the each multimedia packet request made by the mobile client.
mobile client via the base station. The BS communicates with With MoSyncs optimistic protocol, the server sends the mul-
many servers and mobile client at any given time. The BS in timedia packets directly to the mobile client without contin-
the MoSync algorithm have special roles when servicing mo- uous updates being made by the base station once the first
bile client in their area. The base station has three roles: as a multimedia packet from the server arrives to the mobile client.
messenger it passes multimedia units to the mobile client, as The MoSync algorithm operates at the networks application
a filter it sends request to the servers per request of the mobile layer.
information about the arrival time of each packet, it calcu- We have investigated several hand-off schemes for Mo-
lates the synchronization time for the next packet. A sched- Sync [10]. In our experiments, we settle with the two-hand
uler, in the server, manages the on-time transfer of the of scheme, which consists of two parts: setup hand-off and
subframes, as a part of the frames to the mobile client. The end hand-off. In phase one, setup hand-off updates new ar-
mobile client requests multimedia service from the multi- rival base stations and maintains synchronization for newly
media servers, through the base station. When a base sta- arrived mobile clients. If a mobile client can communicate
tion requests the first group of multimedia packets, it sends with another BS, the mobile client will send a new BS ar-
the synchronization point information to all servers [8]. The rived message to its primary BS. For the mobile client to
base station can receive messages from the mobile client and request service from the new BS, the new BS must have delay
from neighboring base stations. The mobile client messages time information about each multimedia server. If the new
are for requesting multimedia packets from servers and done base station does not have delay time information about the
messages to let base stations know all packets have been re- multimedia servers, then the new BS sends a Request mes-
ceived by the mobile client. The neighboring base station sage to all multimedia servers. When a mobile client receives
messages are used to inform the base station if its neighbor multimedia packets, the mobile client calculates the latest ar-
is on or off, to set-up hand-offs to neighboring base stations. rival time and the differences between multimedia packets us-
It is the mobile clients job to calculate the latest arrival time ing the MoSync algorithm. In phase two, the end hand-off
and the differences between every multimedia packets arrival deals with the ordering of multimedia packets and the flow
time, the buffer usage and update its serving BS with the new of the multimedia packets to the mobile client. Any BS can
data. serve as the newly selected primary BS. After informing the
mobile client, BS and multimedia servers, the end hand-off
3.2. MoSync algorithm and hand-off management schemes phase selects the closest common node from the primary BS
and the newly selected BS. The common node must be within
A wireless network consist of a mobile client, a BS that ser- the wireless network. If there is no common node, then the
vices the mobile client and MSC that performs the hand-off MSC of the current primary BS will be the common node un-
between base stations. The BS is connected to a wired server til a common node can be found. The common node reroutes
network that provides the mobile client with services, such as the multimedia packets in time-stamped order using the newly
data, graphics, audio and video, via the BS, through a wire- selected primary BS. Once the mobile client moves within
less link. Each BS has a coverage area known as a cell. When the coverage area of the selected BS, the end hand-off phase
a mobile client, receiving services from a BS, moves out of is terminated. The MoSync algorithm works with both hard
the coverage area of the servicing BS and into the cell of an- hand-off and soft hand-off to synchronize multimedia packet
other BS, the mobile clients services must be transfer to the data flows [9].
new BS to ensure continuous service. Hand-offs in a wire-
less network are classified as Soft Hand-offs and Hard Hand-
offs. 4. Qos packet scheduling algorithms
In a hard hand-off, the connection between the mobile
client and servicing BS, is broken before new services are es- In a wireless network, the mobile client requests multimedia
tablished with the new cells BS. The mobile client can only services through a base station. The base station forwards
be connected to one BS at any time, with a hard hand-off, the multimedia packet request to the distributed multimedia
because frequencies between base stations must be different server. The multimedia server sends the multimedia packet to
to avoid interference between mobile client signals. Because the base station where the packets are buffered and sched-
a mobile client can only be connected to one BS at a time, uled for play-out onto the mobile client that requested the
during hand-off, delays between services are introduce into multimedia service. With todays wireless technology, the
the network. Hard hand-off is mostly used in FDMA and mobile client can request video, audio, text and graphics ser-
TDMA [1]. vices from distributed multimedia servers via the base station.
In a soft hand-off, the connection between the servicing The base station cannot play-out the multimedia packets onto
BS and the mobile client is maintained while another con- the mobile client until the latest packet has arrived. If an on-
nection between the mobile client and the new cells BS is demand-video movie is requested by the mobile client, the
established. The mobile client can be connected to one or base station must receive both audio and video packets for a
more base stations at any time, in a soft hand-off and con- particular frame from the distributed multimedia servers be-
current connections between mobile clients and base stations fore the frame can be played-out onto the mobile client. The
can exist because the frequency band at the base stations dif- multi-data stream dependence between multimedia packets,
fers. Soft hand-off is mostly used in CDMA [1]. In a wireless such as video and audio, is known as inter-stream synchro-
network, soft hand-off is the preferred method of hand-off be- nization or lip-synchronization. Each audio and video mul-
cause the mobile clients services are not interrupted when a timedia packet must be kept in temporal order for smooth
hand-off is performed. Because there is no interruption in ser- play-out audio and video at mobile client. The time de-
vice, using soft-hand off, audio and video is played-out on the pendency of multimedia packets of the same data stream is
mobile client smoothly. known as intra-stream synchronization. The mobile client
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 237
calculates the arrive of the multimedia packets and informs to the highest priority. PQ serves the highest priority pack-
the base station of the arrival time difference. The updated ets first and then moves to the lowest after all higher packets
messages from the mobile client to the base station allows have been processed. Priority-based algorithms reduce delays
the base station to synchronize its request to the multime- for high priority classes that are bandwidth sensitive, such as
dia server and synchronizes the multimedia packets play-out video, audio and interactive traffic. PQ reduce delays for high
onto the mobile client. In the MoSync algorithm the base sta- priority classes with out reducing the overall throughput of
tion schedules play-out of multimedia packets on the mobile the lower classes. PQ can lead to unfairness among classes
client in a FIFO order. The resources at the mobile client if not monitored closely because higher priority classes that
are limited in terms of power, bandwidth and memory. The arrives on a continuous basis will always be processed before
object of multimedia synchronization is to keep the buffer the lower priority classes are processed. To reduce starvation
at the mobile client full so smooth play-out of multimedia among lower priority classes, PQ monitors the traffic and will
services can be presented. If the multimedia packets from update the lower priority classes priority to enable process-
the servers are scheduled too early, buffer overflow occurs ing of lower priority classes during continuous high priority
at the mobile client. If multimedia packets from the servers classes traffic flows.
are scheduled after the deadline to maintain temporal order,
buffer underflow occurs at the mobile client. Mobile client 4.3. Round Robin (RR)
buffer underflow and overflow is a network problem that has
to be dealt with through synchronization of packet delivery, Lately, there has been an increase in multimedia traffic such
to the mobile client from the servers, via the base station. as audio and video. The heavy demand for multimedia traffic
The scheduler in the base station effects network synchro- increases QoS needs to enhance the multimedia communica-
nization in terms of buffer usage, overflow, underflow, mes- tion. Packet scheduling in the networks play an important part
sage complexity and multimedia packet delay at the mobile in the QoS of each client in the network. These multimedia
client. services are bursty and bandwidth hungry. To ensure fairness
In this paper, we wish to investigate how packet schedul- between processing of packets in the network, Round Robin
ing algorithms may affect the media synchronzation in wire- (RR) packet scheduling can be used. With RR packet schedul-
less and mobile communication systems. We have investigate ing, arriving packets are store in the same buffer on a First-In-
several packet scheduling schemes, In this paper, we settle First-Out (FIFO) strategy. The packets are removed from the
down with the following QoS packet scheduling algorithms: buffer in a FIFO basis. Each packet is guaranteed a certain
FIFO, PQ, WFQ and RR. quantum of processing time. If the packet finish playing-out
during the allowable processing time, the packet is remove
4.1. First-In-First-Out (FIFO) from the queue. If the packet does not finish processing dur-
ing the allowable time, the packet is stamp with information
In todays wired and wireless systems, First-In-First-Out
such as arrival time and burst time and is place back into the
(FIFO) is the most common scheduling algorithm. There is
buffer as a new arriving packet to be processed. The Round
only one queue and all packets are treated equally. Packets
Robin algorithm creates fairness among packets by guaran-
are stored at a buffering location and processed on a first come
teeing play-out time and by isolating each traffic flow from
first serve basis. FIFO does not recognize priority or classes
misbehaving traffic. Bursty packet like audio and video that
of traffic. If the buffering area does not have space to store
arriving packets, the buffer area discards the packets [14]. In arrives for a particular flow are spread out and transmitted one
terms of data flow, since large flows may arrive first and cap- at a time according to the RR packet scheduling algorithm.
ture large bandwidth, FIFO is a since of unfairness among RR does not easily adjust to the network load due in part to
flows. On the average, a flow burst received is usually trans- the quantum slice that is static.
mitted in a similar burst using FIFO scheduling. FIFO is the
fastest queuing method. It is effective for large links that have 4.4. Weighted fair queuing (WFQ)
little delay and minimal congestion.
Multiple traffic flows can be identical in priority and their net-
4.2. Priority Queuing (PQ) work resource usage. With the multimedia traffic of today,
assigning a priority or guaranteeing play-out time of multi-
Some multimedia requests, like on-demand-video, requires media packets is not enough to ensure high QoS for a client.
large bandwidth and guaranteed processing order of multime- The Weighted Fair Queuing (WFQ) algorithm combines the
dia packets. In this special case, Priority Queuing (PQ) can PQ and RR into a dynamic fair queuing that divides the avail-
be used to guarantee the bandwidth, memory and process- able bandwidth among traffic queues base on weight or pri-
ing time for a particular high priority data flow. PQ sets a ority. In todays networks, video packets are given a higher
certain number of priority classes according to some parame- priority than audio packets, audio packets are given a higher
ter. In our experiment, the packet with the lowest play-out priority than graphic packets and graphic packets are given a
time receives higher priority. In a wired and wireless com- higher priority than text during play-out of multimedia pack-
munication system the priority can be an IP address, proto- ets. WFQ ensures that each packet is treated fairly according
col or interface [14]. The packets are played out according to its weight. Packets of the same priority is stored in the same
238 BOUKERCHE AND OWENS
buffer. Multimedia packets of the same priority are played- are exponentially distributed, with mean values of 20, 40 and
out in a round robin order. Hence, satisfactory response time 60, respectively. The upper bound delay is set in the nonuni-
is a result of WFQ fairness. For delay sensitive applications form distribution to prevent failure, caused by large delays,
such as video-on-demand, animated graphics and interactive in the stimulation. The mean communication session is set to
applications, WFQ is the most appropriate scheduling algo- 20 multimedia packets and the mean buffer size of a mobile
rithm. Furthermore, WFQ dynamically adjusts to traffic flow. client is set to three times the size of the multimedia pack-
Low volume and high volume traffic both get fair allocation ets [8]. Our results will that each range of delay jitters cause
of available bandwidth. WFQ may not perform well in a wire- the network to behave differently.
less environment where the resources such as memory, power
and bandwidth are limited.
5.1. Method of experiment
5. Simulation experiments The BS plays-out the multimedia packets, sent by the mul-
timedia servers, onto the mobile client that requested the
We have developed a discrete-event model to simulate a cellu- packets. To avoid mobile client buffer overflow and buffer
lar wireless multimedia system, on a wireless and wired net- underflow and to ensure intra-stream and inter-stream syn-
work. Our assumption is that the communication between chronization of the multimedia packets, multimedia packets
servers and base stations are wired and connections between request and delivery must be synchronized with the mobile
mobile clients and base stations are wireless. In our model, client that requests the multimedia packets.
there are 300 channels available for 60 cells, with the system The multimedia servers, which send the multimedia pack-
load equally distributed over all cells. Each mobile client has ets request, have a built-in scheduler that delivers the pack-
a buffer size thats at least 3 times that of the largest multi- ets on time. The base station receives and buffers multime-
media packets and at most six time as large as the multimedia dia packets from the multimedia servers and schedules the
packets received from the server. In this simulation, we em- packets for play-out onto the mobile client that requested the
ploy a two-phase hand-off scheme using soft-hand hand-off multimedia packets. In earlier studies, using the MoSync
in CDMA [5]. algorithm, the base station plays-out the multimedia pack-
We have performed several tests, using many jitters and re- ets onto the mobile clients in a FIFO order. We will make
synchronization. We evaluate MoSyncs performance in two a correlation between the scheduling of multimedia pack-
environments: uniform and nonuniform jitters using FIFO, ets by the BS, to be played-out onto the mobile client and
PQ, WFQ and RR scheduling. Table 1 describes the parame- the effect the multimedia packets scheduling, by the BS, has
ters we use in our simulation experiments. In the uniform on intra-stream synchronization, inter-stream synchronization
case, for the four packet-scheduling algorithms, we evalu- and mobile clients buffer usage. We assume that the wired
ate each algorithm, assuming uniform multimedia unit de- network can provide efficient data flow from the multimedia
lay, under jitters. All cells have the same multimedia packets servers, to the mobile client, per request of multimedia pack-
demand and the requests from the mobile client have inter- ets.
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. 5.2. Performance metrics
Due to the distribution of the delays, the jitters will cause dif-
ferent effects on the synchronization algorithm. To assess the performance of each of the scheduling algo-
In the nonuniform case, we have assumed exponential dis- rithms, we choose the same performance metrics as in earlier
tribution of network delay jitters. This distribution was used work [8,9]. Our areas of focus are as follows:
in earlier studies of the MoSync algorithm [8]. For our exper-
Message complexity: measures the overhead in terms of
iments, there are multimedia media delay ranges as follows:
the number of messages needed to satisfy the users mul-
(1) 0200 msec; (2) 0400 msec; (3) 0600 msec. The ranges
timedia requests.
Table 1
Simulation parameters.
Buffer usage: measures the synchronization behavior for
each mobile client.
Packet scheduling algorithms FIFO, PQ, WFQ, RR
Number of cells 60 Overflow rate: measures the average number of multime-
Number of multimedia servers 4 dia units that overflow a mobile client buffer.
Mobile client buffer size 30 times of a multimedia packets Underflow rate: measures the average deficiency in the
Play-out time/multimedia packets 100 msec
Mean service time/session
clients buffer of the number of multimedia units currently
Arrival rate in normal cell needed for smooth play-out.
Round-Trip-Time (RTT) to Multimedia unit arrival rate: measures the reaction
request/deliver a multimedia packets 50 msec
Jitters (uniform) 020, 2040, 4060, 6080 msec
of the network and the network condition by moni-
Jitters (non-uniform) 0200, 0400, 0600 msec toring multimedia packets arrival time to client (intra-
synchronization).
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 239
6. Simulation results
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
inter-stream synchronization during mobile client request and [14] K. Ito, S. Tasaka and Y. Ishibashi, Media synchronization quality of
play-out of multimedia packets. The different packet schedul- packet scheduling algorithms, IEICE Trans. Comm. E85-B(1) (January
2002).
ing algorithms affect the QoS of the mobile client, in terms
[15] J.P. Jarmasz and N.D. Georganas, Design a distributed multimedia
of buffer overflow, buffer underflows, buffer usage and intra- synchronization scheduler, in: Proc. IEEE Internat. Conf. Multimedia
stream synchronization and inter-stream synchronization. We Computing and Systems (June 1997) pp. 3442.
conclude that the RR packet scheduling is the most efficient in [16] T.V. Johnson and A. Zhang, Dynamic playput scheduling algorithms
the wireless multimedia in terms of synchronization between for continuous multimedia streams, Multimedia Systems 7 (1999) 313
315.
multimedia servers, BS and mobile client. The RR packet
[17] R. Krishnan and T.D.C. Little, Service aggregation through rate adapta-
scheduling algorithm allows for 50 percent less delay between tion using a single storage format, in: Proc. 7th Internat. Workshop on
concurrent data streams. The decrease in delay of concurrent Network and Operating System Support for Digital Audio and Video,
data streams enhances the QoS at the mobile client and al- St. Louis, MO (1997) pp. 121232.
lowed for smooth multimedia packet play-out at the mobile [18] L. Lamont, L. Li, R. Brimont and N.D. Georganas, Synchronization of
multimedia data for a multimedia news-on-demand application, IEEE
client. RR packet scheduling, with its fair packet schedul-
J. Selected Areas Comm. 14(1) (1996) 264278.
ing and flow isolation technique, is the best choice in terms [19] K.S. Lee and M. El Zarki, Comparison of different scheduling al-
of buffer usage and inter-stream synchronization, at the ap- gorithms for packetized real-time traffic flows, in: Internat. Sympos.
plication layer, for multimedia packet play-out. Through our on Wireless Personal Multimedia Communications, Denmark (2001)
research, we discover that the play-out of multimedia pack- pp. 2028.
[20] T.D.C. Little and F. Kao, An intermediate skew control system for mul-
ets, by the BS onto the mobile client affects the overall net-
timedia data presentation, in: Proceedings 3rd Internat. Workshop Net-
work synchronization and can be used to enhance the QoS at work and Operating System Support for Digital Audio and Video (No-
the mobile client by allowing on-time delivery of multimedia vember 1992) pp. 121132.
packets to the mobile client. [21] D.M.H. Ly, User to network QoS parameter transformation in net-
We see several directions for future research. We plan to worked multimedia systems, in: Workshop on Resource Allocation
Problems in Multimedia Systems (with RTSS96) (December 1996)
extend our MoSync scheme and make them power aware [6]. pp. 1119.
[22] D.L. Mills, Internet time synchronization: the network protocol, IEEE
Trans. Comm. 39(10) (1991) 14821493.
References [23] M. Mowbray, G. Karlsson and K.T. Kohler, Capacity reservation for
multimedia traffic, Hewlett Packard Laboratories Report (2000).
[1] S. Abhyankar, Handoff in mobile wireless network (November 2000). [24] S. Pamanathan and V. Rangan, Adaptive feedback techniques
[2] M. Adler, Y. Bartal, J.W. Byers, M. Luby and D. Raz, A modular analy- for synchronization multimedia retrieval over integrated networks,
sis of network transmission protocols, in: Proc. 5th Israeli Sympos. on IEEE/ACM Trans. Networking 1(2) (1993) 246260.
Theory of Computing and Systems (June 1997) pp. 1018. [25] N.U. Qazi, M. Woo and T.D.C. Little, A temporal reference framework
[3] Z. Ali, M. Woo and A. Ghafoor, Distributed synchronization proto- for multimedia synchronization, IEEE J. Selected Areas Comm. 14(1)
cols for multimedia services on Internet, IEEE/ACM Trans. Network- (1996).
ing 2(2) (1994) 2532. [26] S. Ramanathan and P.V. Rangan, Adaptive feedback techniques for syn-
[4] E. Biersack and W. Geyer, Synchronization delivery and play-out of chronized multimedia retrieval over intergrated networks, IEEE/ACM
distributed stored multimedia streams, Multimedia Systems 7 (1999) Trans. Networking 1(2) (1993) 246259.
7090. [27] S. Shirani, F. Kossentini and R. Ward, Error concealment methods, a
[5] U. Black, Second Generation Mobile and Wireless Networks (Prentice comparative study, in: Proceedings of CCECE, Edmonton, Alberta,
Hall, NJ, 1999). Canada (May 1999) pp. 835840.
[6] A. Boukerche, Media synchronization protocols for mobile ad-hoc [28] T. Znati, R. Simon and B. Field, A network-based scheme for synchro-
wireless networks. In preparation. nization of multimedia streams, in: Proc. Internat. Workshop Multime-
[7] A. Boukerche and T. Huang, Real-time transport protocols of view mor- dia Synchronization, Tysons Corner, VA, May 19 (1995) pp. 1220.
phing over wireless communication. [29] T. Znati, R. Simon and B. Field, A network-based scheme for synchro-
[8] A. Boukerche, S. Hong and T. Jacob, An efficient synchronization nization of multimedia streams, Technical Report TR95-02, Depart-
scheme of multimedia streams in wireless and mobile systems, IEEE ment of Computer Science, University of Pittsburgh, PA (1995).
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. Azzedine Boukerche is a Canada Research Chair
[10] A. Boukerche, H. Sungbum and T. Jacob, A two-phase handoff man- and an Associate Professor of Computer Sciences at
agement scheme for synchronizing multimedia units over wireless net- the School of Information Technology and Engineer-
works, in: IEEE Internat. Sympos. on Computer and Communications ing (SITE) at the University of Ottawa, Canada. He
(ISCC) (July 2003) pp. 10781084. is also the Founding and Director of PARADISE Re-
[11] H.-Y. Chen and J.-L. Wu, MultiSynch: A synchronization model for search Laboratory (PARAllel, Distributed and Inter-
multimedia systems, IEEE J. Selected Areas Comm. 14(1) (1996) 238 active Simulation of LargE scale Systems and Wire-
248. less&Mobile Networking). Prior to this he was an
[12] S. Floyed, B. Denny, V. Jacobson, S. Shenker and L. Zhang, Packet Assistant Professor of Computer Sciences at the Uni-
scheduling research, in: DARTnet II Meeting (March 1995), Technical versity of North Texas, and Director of the Parallel
Report (1995). Simulations, Distributed and Mobile Systems Research Laboratory) at UNT.
[13] Y. Ishibashi, S. Tasaka and H. Ogawa, A comparison of media syn- He also worked as a Senior Scientist at the Simulation Sciences Division,
chronization quality among reactive control scheme, Department of Metron Corporation located in San Diego. He was employed as a Faculty at
Electrical and Computer Engineering, Nagoya Institute of Technology, the School of Computer Science (McGill University, Canada) from 1993 to
Nagoya 466-8555, Japan (2002). 1995. He spent the 19911992 academic year at the JPL-California Institute
MEDIA SYNCHRONIZATION AND QOS PACKET SCHEDULING 249
of Technology where he contributed to a project centered about the specifica- puting (JPDC), ACM Wireless Networks (WINET), ACM Mobile Networks
tion and verification of the software used to control interplanetary spacecraft and Applications (MONET), and the International Journal of Wireless and
operated by JPL/NASA Laboratory. Mobile Computing. He was the main organizer of a Special Session on Per-
His current research interests include wireless and mobile networks, dis- formance Analysis of Mobile and Wireless Communication systems at the
tributed and mobile computing, distributed systems, parallel simulation, dis- 7th IEEE HiPC Conference. He has been a member of the Program Commit-
tributed interactive simulation, and performance modeling. Dr. Boukerche tee of several international conferences such as ICC, VTC, ICPP, MASCOTS,
has published several research papers in these areas. He was the recip- BioSP3, ICON, ICCI, MSWiM, PADS and WoWMoM, LWN, Networking
ient of the best research paper award at IEE/ACM PADS97, and the re- conferences.
cipient of the 3rd National Award for Telecommunication Software in 1999 Dr. A. Boukerche serves as a General Chair for the 1st International Con-
for his work on a distributed security systems on mobile phone operations, ference on Quality of Service on Heterogenous Wired/Wireless Networks
and has been nominated for the best paper award at IEEE/ACM PADS99 (QShine 2004), a Program Chair for the 5th IEEE Workshop on Wireless,
and at the ACM Modeling, Analysis and Simulation of mobile and wire- Mobile Ad-Hoc and Sensors Networks, as an Associate Editor for the In-
less systems Conference 2001. He served the General co-Chair of the prin- ternational Journal of Parallel and Distributed Computing (Mobile Comput-
ciple Symposium on Modeling Analysis, and Simulation of Computer and ing Area), ACM/Kluwer Wireless Networks, Wiley International Journal on
Telecommunication Systems (MASCOTS), in 1998, General Chair of the Wireless Communications and Mobile Computing, and SCS Transactions on
Fourth IEEE International Conference on Distributed Interactive Simula- Simulation, a Founding and a Steering Committee Chair of ACM MSWiM
tion and Real Time Applications (DS-RT2000), General Chair for the 3rd Symposium, ACM Performance Evaluation of Wireless Ad hoc and Sensors
ACM Conference on Modeling Analysis, and Simulation of Wireless and Networks (PE-WASUN), and IEEE DS-RT Symposium, and on the IEEE
Mobile Systems (MSWiM2000), as Program Co-Chair for the 5th IEEE In- Task Force on Cluster Computing (TFCC) Executive Committee. He is a
ternational Conference on Mobile and Wireless Computing and Communica- member of the IEEE and ACM.
tion (MWCN03), ACM/IFIP Europar 2003, IEEE Wireless Local Networks E-mail: boukerch@site.uottawa.edu
(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 Harold Owens II received his B.Sc. degree and M.Sc. degree from the De-
Time Applications (DS-RT99), and ACM MSWiM2000, and as a Deputy partment of Computer Science at the University of North Texas. His main
Vice Chair of Wireless and Mobilty Access Track for ACM WWW 2002, as areas of research interests include mobile ad hoc networks, and wireless mul-
a Guest Editor for VLSI Design, the Journal of Parallel and Distributed Com- timedia systems.