10 1 1 81

CAPACITY AND CROSS-LAYER DESIGN OF WIRELESS AD HOC NETWORKS
A DISSERTATION SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
Stavros Toumpis July 2003
c Copyright by Stavros Toumpis 2003 All Rights Reserved
ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.
Andrea J. Goldsmith (Principal Adviser)
Nicholas Bambos
Fouad A. Tobagi
Approved for the University Committee on Graduate Studies.
iii
iv
Abstract
HE FOCUS
of this thesis is on the establishment of theoretical lower and upper
bounds on the trafc-carrying capabilities of wireless ad hoc networks. This theo-
retical investigation brings forward good design principles, therefore a part of the thesis is devoted to the design of wireless ad hoc networks based on these principles. We rst dene and study capacity regions for networks with an arbitrary number of nodes and topology. These regions describe the set of achievable rate combinations between all source-destination pairs in the network under various transmission strategies, such as variable-rate transmission, single hop or multihop routing, power control and successive interference cancellation. Multihop cellular networks and networks with energy constraints are studied as special cases. With slight modications, the developed formulation can handle node mobility and time-varying at fading channels. Numerical results indicate that multihop routing, the ability for concurrent transmissions, and successive interference cancellation signicantly increase the capacity of ad hoc and multihop cellular networks. On the other hand, gains from power control are signicant only when variable-rate transmission is not used. Also, time-varying at fading and node mobility actually improve the capacity. Finally, multihop routing greatly improves the performance of energy-constrained networks. We next investigate the performance and interaction of various protocols lying in different layers of the OSI protocol stack. Our approach is to compare the actual performance of wireless networks with theoretical upper bounds, as determined by the capacity region formulation of the rst part. In particular, we study the routing protocol, the queuing discipline, the power control algorithm and the medium access control protocol. Our study culminates in the introduction of two medium access control protocols: the Progressive Back Off Algorithm (PBOA), which performs medium access jointly with power control, v
and the Progressive Ramp Up Algorithm (PRUA), which sacrices energy efciency in order to achieve higher throughputs. Both protocols slot time, and are integrated with queuing disciplines that are more relaxed than the First In First Out (FIFO) rule. They are totally distributed, they couple well with a variety of routing protocols, and the overhead they require does not increase with the size and node density of the network. We conclude by studying wireless ad hoc networks with a large number of nodes. We rst concentrate on a network of n immobile nodes, with n randomly created sourcedestination node pairs. We develop a scheme under which, in the absence of fading, the network can provide each pair with a trafc rate 1 (n) = K1 (n log n) 2 . This result has been shown recently under a similar setting, however the proof presented here is much shorter. We then proceed to show that, under a general model of fading, each source can send data to its destination with a rate 2 (n) = K2 n 2 (log n) 2 . Next, we extend our formulation to study the effects of mobility. We rst develop a simple scheme under which each of the n nodes can send data to a randomly chosen destination with a rate 3 (n) = K3 n 2 (log n) 2 , and with a xed upper bound on the packet delay dmax that does not depend on n. We subsequently develop a scheme under which each of the nodes can send data to its destination with a rate 4 (n) = K4 n
d1 2 1 3 1 3 1
(log n) 2 , provided that nodes are
willing to tolerate packet delays smaller than a dmax (n) < K5 nd , where 0 < d < 1. With both schemes, a general model of fading is assumed. In addition, nodes require no global topology or routing information, and only need to coordinate locally.
vi
Acknowledgments
URING MY STAY AT
S TANFORD I had the opportunity to associate with a number
of truly remarkable people.
I would like to deeply thank my advisor, Prof. Andrea Goldsmith, for giving me the opportunity to work as part of her research group. Over the years that I had the privilege of working with her, I grew to respect her not only for her great intuition and intellect, but also, equally importantly, for her unbounded enthusiasm for her work, and for her dedication to her students. I also wish to thank Prof. Leonard Tyler, for giving me the life-altering opportunity to come to Stanford. When I rst met him I was instantly impressed by his scientic background, but also by the personal interest he took for his students. My respect for him has only grown with the years. Dr. Richard Simpson has been a constant source of support, and also inspiration, in particular during my rst two years at Stanford. He is one of the most fascinating individuals to work with, although I am sure he would be quick to disagree, using a witty remark. I wish to thank Prof. Nicholas Bambos for serving as my associate advisor and as a reader of this dissertation. I am grateful to Prof. Fouad Tobagi for serving as a member of my orals committee and as a reader. I would also like to thank Prof. Balaji Prabhakar for serving as the chairman of my orals committee. My interaction with members and visitors of my research group has been extremely benecial. Many thanks to Hrishikesh, Kevin, Neelesh, Nihar, Sriram, Shuguang, Syed, Taesang, Taek, Tim, Xiangheng, and Zek. In particular, I wish to express my appreciation to Kevin for his computer administration work. To Taesang I wish all the success with his new, glamorous, and critical position as the group caterer. I am sure everyone hopes that he will not turn out as repetitive as I was. I am also grateful to Joice DeBolt and Pat Oshiro vii
for all their administration work. I wish to thank my several friends at Stanford that helped make my stay so enjoyable. Particularly, I would like to mention Dimitris, George, and Savvas. While at Stanford, I had the incredible luck of meeting Yanna. I would like to thank her for her patience and support, particularly during the less exciting periods of my studies. Last, I would like to thank my parents and sister, for the unconditional love and support that they have showed over the years. Stavros Toumpis July 2003
viii
For Ivan Tsikouras, 1975-1993
ix
Contents
Abstract Acknowledgments 1 Introduction 1.1 1.2 1.3 Wireless Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . Current Research on Wireless Ad Hoc Networks . . . . . . . . . . . . . . . Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v vii 1 1 3 8 10 14
2 System Model 3 Capacity Regions for Wireless Ad Hoc Networks 3.1 3.2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 Rate Matrices and Capacity Regions . . . . . . . . . . . . . . . . . . . . . 15 3.2.1 3.2.2 3.2.3 Transmission Schemes and Time Division Schedules . . . . . . . . 15 Rate Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Ad Hoc Network Capacity Regions and Uniform Capacity . . . . . 18 Single Hop Routing, No Spatial Reuse . . . . . . . . . . . . . . . . 22 Multihop Routing, No Spatial Reuse . . . . . . . . . . . . . . . . . 23 Multihop Routing with Spatial Reuse . . . . . . . . . . . . . . . . 24 Power Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Successive Interference Cancellation (SIC) . . . . . . . . . . . . . 26
3.3
Capacity Regions for Various Transmission Protocols . . . . . . . . . . . . 21 3.3.1 3.3.2 3.3.3 3.3.4 3.3.5
3.4 3.5
Uniform Capacity of Canonical Topologies . . . . . . . . . . . . . . . . . 27 Discrete-Rate Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . 28 xi
3.6 3.7 3.8
Computational Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Multihop Cellular Networks . . . . . . . . . . . . . . . . . . . . . . . . . 32 Fading and Mobility Increase Capacity . . . . . . . . . . . . . . . . . . . . 35 3.8.1 3.8.2 Time-Varying Flat Fading Channels . . . . . . . . . . . . . . . . . 35 Node Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 Capacity Region Formulation . . . . . . . . . . . . . . . . . . . . 41 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . 42
3.9
Energy Constrained Networks . . . . . . . . . . . . . . . . . . . . . . . . 40 3.9.1 3.9.2 3.9.3
3.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 4 Cross-Layer Design of Wireless Ad Hoc Networks 4.1 4.2 4.3 4.4 4.5 4.6 4.7 45
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Effects of Power Control on the Capacity . . . . . . . . . . . . . . . . . . 48 Effects of the Routing Protocol on Capacity . . . . . . . . . . . . . . . . . 50 When not to FIFO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 Carrier Sense Multiple Access with Collision Avoidance . . . . . . . . . . 53 Progressive Back Off Algorithm (PBOA) . . . . . . . . . . . . . . . . . . 58 4.7.1 4.7.2 4.7.3 4.7.4 Frame Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 Protocol Operation . . . . . . . . . . . . . . . . . . . . . . . . . . 60 Parameter Selection . . . . . . . . . . . . . . . . . . . . . . . . . 62 Performance of PBOA . . . . . . . . . . . . . . . . . . . . . . . . 64 Protocol Operation . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Parameter Selection . . . . . . . . . . . . . . . . . . . . . . . . . 66 Performance of PRUA . . . . . . . . . . . . . . . . . . . . . . . . 67
4.8
Progressive Ramp Up Algorithm (PRUA) . . . . . . . . . . . . . . . . . . 65 4.8.1 4.8.2 4.8.3
4.9
Fundamental Limits of Distributed Protocols . . . . . . . . . . . . . . . . . 68
4.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5 Large Wireless Networks 5.1 72
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 xii
5.2
Network Model and Overview of Results . . . . . . . . . . . . . . . . . . 73 5.2.1 5.2.2 5.2.3 Immobile nodes with no fading . . . . . . . . . . . . . . . . . . . 73 Immobile nodes with fading . . . . . . . . . . . . . . . . . . . . . 75 Mobile nodes with fading and delay constraints . . . . . . . . . . . 76 The Cell Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 The routing rules . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 The time division scheme . . . . . . . . . . . . . . . . . . . . . . 90 Performance of the basic scheme . . . . . . . . . . . . . . . . . . . 91
5.3
Immobile Nodes with no Fading . . . . . . . . . . . . . . . . . . . . . . . 80 5.3.1 5.3.2 5.3.3 5.3.4
5.4 5.5
Immobile Nodes with Fading . . . . . . . . . . . . . . . . . . . . . . . . . 92 Node Mobility with a Fixed Delay Constraint . . . . . . . . . . . . . . . . 95 5.5.1 5.5.2 5.5.3 The xed-delay scheme . . . . . . . . . . . . . . . . . . . . . . . 95 No packets are dropped . . . . . . . . . . . . . . . . . . . . . . . . 96 Performance of the xed-delay scheme . . . . . . . . . . . . . . . 99 The polynomial-delay scheme . . . . . . . . . . . . . . . . . . . . 100 No packets are dropped . . . . . . . . . . . . . . . . . . . . . . . . 101 Performance of the polynomial-delay scheme . . . . . . . . . . . . 105
5.6
A Polynomially Increasing Delay Constraint . . . . . . . . . . . . . . . . . 100 5.6.1 5.6.2 5.6.3
5.7
Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 105 109 113
6 Conclusions and Future Work Bibliography
xiii
List of Figures
1.1 1.2 1.3 3.1
The ad hoc topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The cellular topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The OSI protocol stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (i) A small wireless ad hoc network of 4 nodes. (ii) Transmission scheme S1 . (iii) Transmission scheme S2 . Transmit-receive node pairs in operation are connected by arrows. The node whose information is being transmitted and the link transmission rate are shown next to the link arrows. The rates are dictated by the SINR of each link and the receiver function fR (). . . . . . . . . . . . . . . . . . . . 15
2 3 5
3.2
End-to-end information ows for two time division schedules that use the schemes of Fig. 3.1: (i) T1 = 0.5S1 + 0.5S2 . (ii) T2 = 0.75S1 + 0.25S2 . Arrows signify the end-to-end information ows, and numbers denote the overall communication rates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 3.4
An example wireless ad hoc network with 5 nodes. . . . . . . . . . . . . . . . . 21 Capacity region slices of the example ad hoc network along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c). . 23
3.5
Uniform capacity of two canonical topologies as a function of the number of nodes: (i) Linear networks. (ii) Ring networks. (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
xiv
3.6
(i) A class of stepwise receiver functions with 1, 3, 7, and 15 rate levels (plots (a), (b), (c), and (d) respectively). The Shannon bound of (2.3) is also plotted (plot (e)). (ii) Uniform capacity versus the number of available rate levels for the example ad hoc network. Each curve corresponds to a different transmission protocol: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d,d,d) Power control with 2, 3, and 4 levels added to (c). (e) Successive interference cancellation added to (c). . . 29
3.7
The time division schedule that maximizes r12 = r34 , if the rest of the node pairs have no communication requirements, for the example ad hoc network of Fig. 3.3. Arrows signify actual transmissions. Numbers denote link rates (in Mbps). . . . . 31
3.8
The total number of rate matrices that exist versus the number of nodes n and various protocols: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse and 1 4 distinct power levels. (d) Successive interference cancellation added to (c). (e) The maximum number of rate matrices that our current software and hardware can create. Although the number of total rate matrices increases very fast with the number of nodes, it is not necessary to explicitly create all of them. A small subset of them will in reality contribute to the capacity region, and only those must be explicitly created. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.9
(i) An example multihop cellular network with 6 nodes and a base station. (ii) Capacity region slices of the example multihop cellular network along the plane ri = 0, i = 3, 5: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c). . . . . . . . . 34
3.10 A slice of the capacity region C c for the example ad hoc network of Section 3.3
along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j for different combinations of fading states: (a) Fading state F 1. (b) Fading state F 2. (c) Fading states F 1, F 2. (d) Fading states F 1, . . . , F 10. (e) Fading states F 1, . . . , F 15.
. . . . . . . . . 36
xv
3.11 Estimates of the uniform capacity of various protocols, for a ring ad hoc network
with 5 nodes, as a function of the standard deviation of the shadowing component. Estimates are accurate within 0.2 Mbps. (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c).
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 . . . . . . . . . . . 39
c 3.12 Convergence of the estimate for the uniform capacity Cu with node mobility, for
the example ad hoc network of Section 3.3, with = 0 dB.
3.13 (i) An example energy constrained network with 15 nodes. (ii)Capacity region
slices of the network of Fig. 3.13 along the plane bi = 0, i = 5, 10: (a) Single hop routing. (b) Multihop routing. . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1
A wireless ad hoc network of 12 nodes. A signal transmitted by a node reaches all others. However, signal strength decays with distance, and only nodes connected by a straight line can communicate directly with each other, in the absence of interference from competing transmissions. . . . . . . . . . . . . . . . . . . . 47
4.2
Capacity region slices of the example ad hoc network along the plane Rij = 0, (i, j) = (3, 12), (4, 8), i = j: (a) Power control, optimal routing. (b) No power control, optimal routing. (c) Power control, RP-120. (d) No power control, RP120. (e) RP-120, CSMA/CA. (f) RP-120, PBOA, with m = 15, p = 0.8. (g) RP-120, PRUA, with m = 8, p = 0.3, PT = 1011 W. . . . . . . . . . . . . . . 48
4.3
A small network of 6 nodes. Only nodes connected by a dashed line can communicate directly. The rest do not interfere at all. The strategy that maximizes the end-to-end trafc from node A to node B divides the time equally in three transmission schemes, and employs two different routes. . . . . . . . . . . . . . 51
4.4
Uniform capacities versus the routing protocol for the example network: (a) Power control, optimal routing. (b) No power control, optimal routing. (c) Power control, minimum-hop routing. (d) No power control, minimum-hop routing. (e) Minimum-hop routing, CSMA/CA. (f) Minimum-hop routing, PBOA, m = 15, p = 0.8. (g) Minimum-hop routing, PRUA, m = 8, p = 0.3, PT = 1011 W. . . . 52
4.5 4.6
A simple example where employing a non-FIFO queuing discipline is advantageous. 53 The 4-way handshake employed in CSMA/CA. . . . . . . . . . . . . . . . . . 55
xvi
4.7
CSMA/CA imposes articial restrictions on how close transmissions are placed. (i) If B is transmitting a packet to A, C will not be able to initiate a data packet transmission to D. (ii) If A is transmitting a packet to B, D will not be able to initiate a data packet transmission to C. However, if time is slotted and all RT S packets are transmitted simultaneously, the articial restriction on how close transmission can be is lifted. . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.8
Throughput-delay curves for the example network, under RP-120. (a) CSMA/CA. (b) PBOA with m = 15, p = 0.8. (c) PRUA with m = 8, p = 0.3, PT = 1011 W. (d) Uniform capacity with no power control. (e) Uniform capacity with power control. (f) Lower bound on delay, caused by data packetization and the transmission of each packet over multiple hops. . . . . . . . . . . . . . . . . . 57
4.9
Average energy consumed for the transportation of a packet to its nal destination as a function of the achieved throughput, for the example network: (a) CSMA/CA, RP-10. (b) PBOA, RP-10. (c) PRUA, RP-10. (d) CSMA/CA, RP-30. (e) PBOA, RP-30. (f) PRUA, RP-30. (g) CSMA/CA, RP-120. (h) PBOA, RP-120. (i) PRUA, RP-120. For PBOA, m = 15, p = 0.8. For PRUA, m = 8, p = 0.3, PT = 1011 W. 59
4.10 Frame format of PBOA and PRUA. . . . . . . . . . . . . . . . . . . . . . . . 60 4.11 Uniform capacity of the example network of Fig. 4.1 versus the number of minislot
pairs m for PBOA and PRUA. For each value of m, the optimal value of p is chosen. For PRUA, PT = 1011 W. (a) PBOA, RP-10. (b) PBOA, RP-30. (c) PBOA, RP-120. (d) PRUA, RP-10. (e) PRUA, RP-30. (f) PRUA, RP-120.
. . . . 63
4.12 A wireless ad hoc network for which a graph channel model is assumed: Two
nodes connected by a straight line can communicate directly in the absence of interference. All other node pairs do not interfere with each others transmissions. Dashed arrows represent a transmission scheme that consists of four transmissions that can coexist. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1
A fundamental tradeoff between the packet delays and the per-node achievable throughput.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2 5.3
Partition of the two-dimensional wireless ad hoc network in a regular lattice of cells. 81 Examples of the routing algorithm.
. . . . . . . . . . . . . . . . . . . . . . . 83 xvii
5.4
One of the 9 sub-lattices of cells appears shaded. Only nodes in the sub-lattice are allowed to receive in the corresponding slot, and only from nodes in the same or neighboring cells. The neighbors of cell c0 appear striped. . . . . . . . . . . . . 89
5.5
(a) The frame structure used in the xed-delay and polynomial-delay schemes. (b) The frame structure of each of the N virtual channels. . . . . . . . . . . . . 95
xviii
Chapter 1 Introduction
1.1 Wireless Ad Hoc Networks
IRELESS AD HOC NETWORKS
consist of mobile nodes communicating over a
wireless channel. Any node can in principle transmit data to any other; how-
ever, because of the nature of the wireless channel, each node can effectively transmit to only some of the others, typically those that lie in its vicinity. On the other hand, the trafc requirements are taken to be arbitrary, therefore it is necessary that nodes cooperate to forward each others packets to their nal destinations. If some of the nodes need to communicate with destinations on other networks, for example the Public Switched Telephone Network (PSTN) or the Internet, they will do so by routing their trafc through nodes that lie on the boundary of the two networks and act as interfaces. The concept of wireless ad hoc networks is certainly not new. Because of their suitability for the battleeld environment, they had been the subject of intense research led by DARPA through the 1970s [1] and the early 1980s [2]. At the time, they were referred to as packet radio networks. Research in this area subsided during the late 1980s, however by the mid-90s the interest of the research community in wireless ad hoc networks intensied [3, 4, 5, 6]. We can best justify the interest in the ad hoc topology by comparing it with the ubiquitous cellular topology. Cellular networks consist of a number of mobile nodes, and a 1
CHAPTER 1. INTRODUCTION
Figure 1.1: The ad hoc topology.
number of strategically placed immobile base stations. The base stations do not have communication needs of their own, but are there to serve the mobiles. They communicate with each other, and with the PSTN, through high-speed wired (or wireless) connections. Each base station and its surrounding mobiles and physical space constitute a cell. A mobile node typically only communicates with the base station that lies in its cell, and will have to use that base station to send data both to other networks, such as the PSTN, and to other mobile users that belong to the same network, no matter how close these users are. An obvious advantage of wireless ad hoc networks over cellular networks is their robustness: If a node dissipates its power supply, or malfunctions, or otherwise disappears, the nodes in its vicinity will take over its routing responsibilities. On the other hand, if a base station becomes inoperable, all the nodes in its cell will lose their connection to the network. A second advantage is the superior resource utilization that ad hoc networks achieve: in cellular networks, users that dont have a good wireless link with any base station are either denied service, or the system consumes a lot of resources (bandwidth and energy) to support their operation. On the other hand, in ad hoc networks there are many different paths with which a packet can reach its destination. If the channel between two nodes is in a deep fade, the network will work around that link (provided there are other nodes around to handle the trafc). Contrary to cellular networks, wireless ad hoc networks are totally wireless. They also have a at and redundant topology. Therefore, they can be deployed very fast, and so are a natural match for applications such as search and rescue operations. Moreover, there are many applications where it is simply not feasible to set up a wired infrastructure. A rather
1.2. CURRENT RESEARCH ON WIRELESS AD HOC NETWORKS
Internet, PSTN, etc.
Figure 1.2: The cellular topology. exotic example of this is a network of spacecrafts orbiting Mars. (It is expected that eventually a few more of these spacecrafts will avoid crashing on the surface or disappearing into deep space.) There are many other reasons for why the ad hoc topology may be more appropriate than the cellular topology. For example, in sensor networks [7, 8, 9, 10] where low node densities are preferable, we would need as many base stations as there are users. In home environments with a few smart appliances [4] that only need to communicate among themselves, a gateway to other networks is not needed, so a base station would be redundant. Since ad hoc networks of hundreds or thousands of nodes are envisioned for some applications (for example networks of automobiles cruising on highways), it is imperative that all algorithms employed by the nodes be distributed. This will greatly improve their usefulness, however it will clearly complicate their design. To conclude, we can summarize the most important features of wireless ad hoc networks, by describing them as mobile distributed multihop wireless networks [11].
1.2 Current Research on Wireless Ad Hoc Networks

Classifying research is very difcult in general; more so in wireless ad hoc networks, where most of the body of research is very recent, and things are uid. However, we present a broad overview of the state of the art in wireless ad hoc network design, within the context of the layered architecture. We focus on the research that is most relevant to the work in this thesis.
Layered Architecture:
Modern communication networks are systems of enormous com-
plexity. To facilitate their design and development, communication engineers have used with great success the concept of the layered architecture. Specically, each node is envisioned as consisting of a stack of relatively independent protocols. Each protocol is supposed to handle one task well and is meant to interact only with the protocols that lie immediately above and below in the stack. By precisely specifying what happens at the interface, designers can work on one protocol at a time, instead of the whole protocol stack simultaneously. Fig. 1.3 represents the Open Systems Interconnection (OSI) stack, that was developed as a reference model by the International Standards Organization (ISO) [12]. Although most data networks (for example the Internet) deviate in some of their details from this reference model, its overall structure and principles are almost universally followed. High on the stack (the Application, Presentation, and Session layers) are protocols related to the applications, for example protocols that handle session initiation and termination, data encryption, and compression. In the middle of the stack (the Transport and Network layers) lie the network protocols that handle network function such as routing. At the bottom of the stack (the Data Link Control and Physical Interface, also the Media Access layer for the case of wireless communications) lie protocols that control the operation of point-to-point links. Physical Interface Layer: The physical interface protocol controls the hardware that
transmits and receives bits across the channel. It species the modulation and coding strategies used to communicate bits, and the power used by the transmitter. Contrary to most other media, the wireless channel is susceptible to time-variable fading. In addition, because of its broadcast nature, users interfere with each other. For the physical layer protocol to achieve a satisfactory performance, it is imperative that it dynamically adapt to the fading and interference. This is achieved by adapting the rate of communication, and also the transmitter power. Another characteristic of wireless communications is that very often the transmitter is powered by a nite energy source, i.e., a battery. In these cases, it is the role of the physical layer protocols to choose a modulation and coding scheme which is energy-efcient, within the constraints placed by the layers higher in the stack.
APPLICATION
PRESENTATION
APPLICATION LEVEL
SESSION
TRANSPORT NETWORK PROTOCOLS NETWORK
DATA LINK CONTROL
MEDIA ACCESS
POINT-TO-POINT FUNCTIONS
PHYSICAL INTERFACE
Figure 1.3: The OSI protocol stack. Media Access Layer: In the wireless broadcast medium, nodes need to coordinate so
that packet collisions are avoided, while at the same time the channel is utilized efciently. This is the task of the medium access control (MAC) layer. MAC protocols can be roughly divided in three groups, depending on the strategy they use for determining access rights. In contention based protocols, such as Aloha [12, 13], CSMA [12, 14], BTMA [15], MACA [16], MACAW [17], FAMA [18, 19], and IEEE 802.11 [20, 21], nodes contend for the channel for each packet they need to transmit. In most protocols specically designed for use in ad hoc networks [16, 17, 18, 19, 20], the actual data packet transmission is preceded by an exchange of two small control packets that aim to inform the other nodes that a node is attempting to reserve the channel. Nodes that either have received a control packet or sense the presence of carrier remain silent. This mechanism is typically referred to as Carrier Sense Medium Access with Collision Avoidance (CSMA/CA). Because of its popularity, and also its connection to our proposed protocols, we study CSMA/CA in greater depth in Chapter 4. The second group of MAC protocols perform transmission scheduling. Specically, time is divided into slots, and each node is scheduled to transmit during some of these slots. Transmissions scheduled for the same slot are sufciently separated in space so that they do
not interfere with each other. With these protocols, a node does not have to contend for each packet it receives, but will rather wait until the time it is scheduled to transmit. The simplest such protocol is pure TDMA, i.e., one node is transmitting per slot. With pure TDMA there is no spatial reuse, so its performance can be very poor. In [22, 23, 24, 25] distributed protocols have been proposed that achieve a tight packing of transmissions with spatial reuse. Frequently, transmission scheduling protocols use contention based mechanism to build their schedules [24]. Hybrid protocols, under which nodes both schedule transmissions and may contend for unclaimed slots, have also been developed. In [26] the authors propose a slotted-time algorithm that performs as pure TDMA under high loads (thus having no stability problems) and as CSMA/CA under low loads (so that data packets experience small delays). In [27], the authors extend the reservation mechanism of R-ALOHA [28] and PRMA [29] in the ad hoc environment. Specically, they propose a slotted-time algorithm, in which slots are grouped in frames and nodes that successfully contend for a slot have the option of reserving the corresponding slots in subsequent frames. Under high loads, the nodes essentially discover a transmission schedule. Routing: In a network where each node can only communicate with a subset of the rest
of the nodes, it is necessary that packets are routed through intermediate nodes before they reach their nal destination. But rst, routes have to be discovered; this is the task of the routing layer. Routing rst attracted attention in the context of wired networks, and the Internet in particular, and the existing body of research in this context is extensive [12]. In the context of wireless ad hoc networks, research on routing started very early [1], however recently the interest of researchers has intensied [30, 31, 32, 33]. As expected, researchers have tried to leverage the vast body of knowledge that already existed in the wired setting in the new wireless setting. Although this is a valid approach, it creates a number of challenges. For example, many routing protocols designed for wired networks do not handle topology changes gracefully. However, due to node mobility and fast fading, topology changes are very frequent in wireless ad hoc networks. As another example, delay in wireless ad hoc networks is typically caused by interference and contention, and not congestion. As a result, algorithms that adapt to delays, such as TCP, will behave poorly because they will be misinterpreting the source of delays.
Network Capacity: Broadly speaking, research on capacity attempts to identify theoretical upper and lower bounds on the achievable performance of networks. Conceptually, network capacity cannot be associated with a particular layer, but rather occupies an extra dimension. Given its obvious importance, it is no surprise that this issue has attracted signicant research interest over the years [34, 35]. A recent breakthrough was achieved by Gupta and Kumar [36, 37]. Their critical idea is to allow the number of nodes to go to innity, thereby allowing statistical averaging to have an effect. Consequently, they are able to derive upper and lower bounds on the capacity that hold for all networks, in the limit of a large number of nodes, since, in that limit, all networks are essentially the same. Grossglauser and Tse employ this approach to show that dramatic gains in the throughput can be achieved if nodes are mobile [38]. The down-side of this result is that these large gains are only possible if large packet delays are tolerated. It should be noted that these works are decidedly non-information theoretic: information theoretic considerations are ushed out by requiring that interference be treated as noise. A comprehensive information theoretic approach is intensely pursued by the research community [39], but does not seem to be within our grasp in the immediate future, because of the enormous complexity of the problem. Cross-Layer Design: Although the use of the protocol stack has been an unqualied
success, it does come with a price. It is true that, by creating many protocols independently of each other, designers have a much simpler task. On the other hand, there are issues that naturally span many layers, especially in wireless networks. As an example, consider the power control algorithm that species the power with which signals are transmitted. It is tempting to associate it exclusively with the physical layer. Although such an association would be correct in the case of wired networks, in the wireless environment the transmit power of a particular node will affect all other nodes in the network, by changing the levels of interference experienced by these nodes and the connectivity of the network, which impacts routing. Therefore, in the wireless environment, power control is not conned to the physical layer, but in reality can affect the operation across the physical, MAC, and routing layers. This fact should not be viewed as a complication, but rather as an opportunity for systemlevel cross-layer design, i.e., a design that spans multiple layers of the protocol stack. For
example, good distributed algorithms exist that allow users to pick transmitter powers that ensure different quality of service in each link in the network [40, 41, 42]. As another example, power control can been used to perform channel access [43, 44]. Another example is the coupling between the medium access control and the routing protocol: Certain routing protocols require that nodes broadcast control messages to all neighbors and wait for all of them to reply. In wireless networks, this imposes a tremendous strain on the medium access layer. Also, routing protocols tend to treat all neighbors as equally accessible although in reality those that lie on the fringes of the communication range of the node are much more susceptible to interference and should be avoided, as they represent a greater challenge to the medium access control layer. In other words, it is very important that the designers of the routing protocol of a system are aware of the limitations of the MAC protocol of that system, and vice versa. More examples will appear later on; in any case the list of examples is endless, and they all point to the same fact: in wireless networks, the separation of protocol layers is not as clean as it is in their wired counterparts. As a result, if wireless ad hoc networks are to achieve their full potential, it is imperative that a cross-layer design approach be adopted.
1.3 Contributions
The subject of this thesis is the capacity of wireless ad hoc networks. Our work on capacity naturally leads us to the study of cross-layer design issues for such networks, so part of the thesis is devoted to that topic. Our contributions per chapter are as follows: In Chapter 3 we introduce and study capacity regions of wireless ad hoc networks. Each point in the capacity region represents a combination of rates between arbitrary source-destination pairs that is achievable by an optimal time division schedule. Therefore, the capacity region of a network species upper bounds on its trafccarrying capabilities that are achievable by an optimal application of the network resources. The formulation is remarkably versatile, allowing us to study networks where nodes are capable of multihop routing, successive interference cancellation, rate adaptation, and power control. Networks with nite energy constraints (such as sensor networks) and multihop cellular networks can be treated as special cases. The
1.3. CONTRIBUTIONS
formulation can also handle at fading and node mobility. In Chapter 4 we study the effects on the performance of wireless ad hoc networks of various choices that span different layers of the protocol stack. In particular we study the routing protocol, the queuing discipline, the power control algorithm, and the medium access control layer. Contrary to the more common approach, which is to focus on a particular layer and compare two alternative protocol designs, we take a more holistic approach: we study different layers at the same time, and we compare their performance with the theoretical optimum, which can be determined by the methods of Chapter 3. We conclude by introducing two protocols that jointly perform power control and medium access control, are integrated with efcient queuing algorithms, and are robust with respect to the choice of the routing protocols. In Chapter 5 we concentrate on networks with a large number of nodes. Following [36], we make use of statistical averaging, so that our results only hold with lower bounds on the per-node achievable throughput, with and without fading, when nodes are immobile. We proceed to derive lower bounds on the achievable per-node throughput when nodes are mobile, and under various constraints on the packet delay. In the process, we identify a fundamental delay-throughput tradeoff. probability approaching unity as the number of nodes goes to innity. We derive
Chapter 2 System Model
ESEARCH IN WIRELINE NETWORKS
has been facilitated by the existence of a
mathematical tool of signicant power: graph theory [45]. In short, the network is
mapped to a graph, with users corresponding to nodes and wired connections corresponding to arcs (or edges). The graph can be unidirectional or bidirectional (depending on whether the underlying links are unidirectional or bidirectional), and possibly weighted, if we are interested in the communication rates per link. Given the mathematical appeal of graph theory, and the fact that most research in wireless networks can be traced back to their wired counterparts, it is no surprise that a signicant body of work exists in wireless ad hoc networks that is, to some extent, based on graphs [22, 23, 24, 36]. In such works, it is typically assumed that whether an link exists between two nodes or not will depend on the physical distance between them. However, graph theory is not such a natural t for wireless networks. In such networks, whether or not two users can communicate, and with what rate, mostly depends on the interference created by competing transmissions, some of which are possibly many users away. In other words, one cannot abstract the network by a graph, since there is no hard notion of connectivity but rather a soft notion of connectivity, with varying link qualities which depend on the competing transmissions. One partial solution is to use many graphs: One to denote nodes that can communicate with each other, one to denote nodes that can interfere with each other, and possibly other graphs, depending on the application [46]. Still, such a model fails to capture the additive effects of interference: maybe one interferer is tolerable, but not two or more. 10
11
Things become even more complicated when it is assumed that nodes can control their power. In this case, the existence of a graph edge will depend on the power levels used [47]. It is therefore evident that graph theory cannot provide us with simple and accurate models in a wireless setting. A more accurate network model is the one used in more traditional areas of wireless research, such as power control [42]. Following the terminology of Gupta and Kumar [36], we call it the physical model.1 (Although it is more general that the model in [36], and it predates that work.) Its main components are the channel gain matrix, which describes the propagation of signal power from node to node, and the receiver function fR (), which captures the capabilities of the receivers. We now introduce the physical model in detail. We consider an ad hoc network with n nodes X1 , X2 , . . . , Xn . Each node has a transmitter, a receiver, and an innite buffer, and wishes to communicate with some or all of the other nodes, possibly by multihop routing. We assume that nodes cannot transmit and receive at the same time, and that transmissions occupy the full bandwidth W of the channel. signal with power Gij Pi , where Gij denotes the channel gain between nodes Xi and Xj . Node Xi can transmit with any power Pi Pimax . When Xi transmits, Xj receives the
the diagonal are unimportant, and are set to Gii = 0. We assume that the elements of the gain matrix are frequency-independent, but may be time-varying. The receiver of each node is subject to thermal noise, background interference from
We dene the channel gain matrix to be the n n matrix G = [Gij ]. The elements along
various noise sources such as other networks, and interference from other users, where the interference caused by Xi to Xj is also determined by the link gain Gij . We model thermal noise and background interference jointly, as a single source of noise, with power spectral density i for node Xi . We dene the noise vector H = [1 2 . . . n ]T . transmitting with power Pt . Let us assume that node Xj , j J is attempting to receive at node Xj will be information from node Xi , i J . Then the signal to interference and noise ratio (SINR) ij =
1
Let {Xt : t J } be the set of transmitting nodes at a given time, each node Xt
Gij Pi j W +
kJ ,k=i Gkj Pk
(2.1)
Throughout the paper, terms being dened are set in boldface.
12
CHAPTER 2. SYSTEM MODEL
We assume that the rate of transmission from node Xi to node Xj , under the SINR ij given by (2.1), is given by Rij = fR (ij ), (2.2)
where the receiver function fR () reects the capabilities of the receiver and the performance metric. Many choices are possible for fR (). For example, we can set fR (ij ) = W log2 (1 + 1 ij ) (2.3)
where log2 (x) denotes the logarithm of x, in base 2. Equation (2.3) implies that the transmitter continuously adapts the transmission rate to the receiver SINR. With = 1, the receiver achieves Shannons capacity. With > 1, (2.3) approximates the maximum data rate that meets a given BER requirement under a specic modulation and coding scheme such as coded MQAM [48, 49]. The precise value of depends on the choice of coding and modulation parameters, and the BER requirement. Alternatively, we can use a receiver function that is more common in studies of medium access control and wireless LANs. In particular, we can assume that nodes use for all transmissions a common transmission rate R, and that the receiving node Xj will be able to decode the signal from Xi , with a negligible probability of error, provided the signal Otherwise, the signal is lost. In other words, to interference and noise ratio ij is constantly above a given threshold T : ij T .
fR (ij ) = R 1{ } . ij T
(2.4)
This is a somewhat pessimistic assumption, as typically there is a non-negligible probability that a packet that is below but close to the SINR threshold will be correctly decoded. However, we will use this assumption for clarity of exposition, and none of our results will be signicantly affected if a more accurate model is used. Regarding the trafc model, we make the fundamental assumption that all packets that are created have a single destination. Therefore, we do not consider the cases of multicasting or broadcasting [23, 25]. We make no assumption on the trafc needs of the network, which are taken to be arbitrary for now.
13
Note that we have not specied any model for determining the positions of the nodes. We have also not specied any model related to the propagation of signals. In our context, these two models will ultimately specify the gain matrix. They are of secondary importance, since the quality of the communication does not depend on the location of the users; it only depends on the power of the received signals. In the following chapters we will base our analysis and simulations on the physical model, as described in this chapter, possibly with minor modications. Notably, for the study of successive interference cancellation performed in Chapter 3, some sources of interference will not be treated as noise. Additional assumptions will be stated in the sections where they are relevant.
Chapter 3 Capacity Regions for Wireless Ad Hoc Networks

3.1 Introduction
N THIS CHAPTER
we dene and study capacity regions for wireless ad hoc networks
with an arbitrary number of nodes and an arbitrary topology. These multidimensional
regions provide detailed information about the capabilities of the network, since they contain all achievable combinations of rates between the nodes, under various transmission protocols. The chapter is organized as follows: In Section 3.2 we dene rate matrices and capacity regions for wireless ad hoc networks. In Section 3.3 we dene capacity regions for a sequence of ve transmission protocols of increasing sophistication, and study them in the context of a random network topology. In Section 3.4 we study the performance of these protocols in the context of two canonical topologies, and in Section 3.5 we study their performance under various assumptions on the capabilities of the receivers. In Section 3.6 we discuss computational issues. In Section 3.7 we modify the formulation to study multihop cellular networks and in Section 3.8 we extend the formulation to estimate the effects of time-varying at fading and node mobility on the capacity of the network. In Section 3.9 we present another modication, suitable for the study of energy constrained networks. We conclude in Section 3.10. 14
3.2. RATE MATRICES AND CAPACITY REGIONS
15
X1
X2
X1 X1, 10
X2
X1 X1, 10
X2
X4
(i)
X3
X4
X3, 10
(ii)
X3, 10 X3 X4
(iii)
X3
Transmission scheme S2 . Transmit-receive node pairs in operation are connected by arrows. The node whose information is being transmitted and the link transmission rate are shown next to the link arrows. The rates are dictated by the SINR of each link and the receiver function fR ().
Figure 3.1: (i) A small wireless ad hoc network of 4 nodes. (ii) Transmission scheme S1 . (iii)
3.2 Rate Matrices and Capacity Regions

In this section we dene the capacity region of a network as a set of rate matrices. Rate matrices provide a mathematical framework for describing the transmission schemes and time division schedules used by a network.
3.2.1 Transmission Schemes and Time Division Schedules

A transmission scheme S is a complete description of the information ow between dif-
ferent nodes in the network at a given time instant. Therefore, the transmission scheme at a given time consists of all transmit-receive node pairs in operation at that time and, for each of these pairs, the transmission rate and the original source node of the transmitted information. Note that, as discussed in Chapter 2, nodes cannot transmit and receive simultaneously, and that the rates used in a transmission scheme are set to Rij = fR (ij ). As an example, consider the wireless network of Fig. 3.1(i), where the node pairs (X1 , X2 ), (X2 , X3 ), (X3 , X4 ), (X4 , X1 ) can all communicate directly but the node pairs (X1 , X3 ) and (X2 , X4 ) cannot, perhaps because of an obstruction along their line of sight. Therefore, if X1 wants to communicate with X3 , it must do so by forwarding data via X2 or X4 , and similarly for trafc between nodes X2 and X4 . Two example schemes for this network are S1 and S2 , depicted in Figs. 3.1(ii) and 3.1(iii) respectively. In S1 , node
16
CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS
X1 5 5 X4
(i)
X2
X1
5 2.5 2.5
X2
X3
X4
5
(ii)
X3
Figure 3.2: End-to-end information ows for two time division schedules that use the schemes
of Fig. 3.1: (i) T1 = 0.5S1 + 0.5S2 . (ii) T2 = 0.75S1 + 0.25S2 . Arrows signify the end-to-end information ows, and numbers denote the overall communication rates.
X1 sends its own information to node X2 and node X3 sends its own information to node X4 . As discussed, for each of these transmissions the bit rate is set to f (ij ) where ij is the SINR of the link. In the second scheme, X2 is forwarding X1 s data to X3 , and X4 is forwarding X3 s data to X1 (presumably these data were received at a previous time interval.) At different times, networks may operate under different transmission schemes, for example in order to provide multihop routing. We assume that the network operation is organized in identical consecutive frames of some xed duration. Within each frame, the network operates using successively schemes S1 , . . . , Sk , with scheme Si operating during a fraction of the frame equal to ai , where the time division schedule T =
k i=1 k i=1
ai = 1. We say that the network is using
as the weights of the time division schedule. For the network of Fig. 3.1 two possible time division schedules are T1 = 0.5S1 + 0.5S2 and T2 = 0.75S1 + 0.25S2 . The resulting end-
ai Si , and we refer to the fractions ai , i = 1, . . . , k
to-end information ows when the network operates under these time division schedules appear in Fig. 3.2. Depending on the ordering of schemes within a time division schedule, the schedule may imply non-causal routing, so that within a frame a node may forward trafc from another node before that trafc actually arrives. This situation does not pose a problem since we can place before the sequence of frames a time period of nite duration during which some data are backlogged in the intermediate nodes before their nal destination. Since the
17
initialization period will have nite duration, the overall performance of the network will not be affected, in the limit of a large number of frames. Therefore, without compromising the generality of our results, we neglect causality in our routing since it signicantly complicates the problem and obscures our main results.
3.2.2 Rate Matrices

Although transmission schemes are useful for describing the state of the network at a given time, they are not convenient for mathematical manipulation. We therefore introduce rate matrices. For a network with n nodes we dene the rate matrix R(S) of a transmission scheme S as an n n square matrix with entries rij such that: r, rij = r, 0, If node Xj receives information at rate r with node Xi as the original information source. If node Xj transmits information at rate r with node Xi as the original information source. otherwise.
Each non-zero rate matrix entry denotes a transfer of data. The row index of the entry corresponds to the original source of the data. The column index species the receiver or transmitter of the data. Specically, a negative entry rij < 0 in row i corresponds to the rate at which node Xj transmits information that originated at node Xi . This entry is negative to reect the fact that the data forwarded cannot be counted in the data that Xj receives from Xi . A positive entry rij > 0 correspond to the rate at which node Xj receives information that originated at node Xi , directly or from another node Xk , in which case rik = rij . For example, the rate matrices of schemes S1 and S2 are, respectively: 10 10 0 0 0 0 0 0 0
0 R1 = 0 0
0 , R2 = 0 10 10 10 0 0 0
10 10 0 0 0 0 0 0
0 . 10 0
Rate matrices mathematically capture all the information needed to describe the state
18
of the network at a given time: namely which nodes transmit or receive, at what rate, and from which nodes the data originate. We note, that since information must be preserved, i.e. each transmission originates at one node and is received by one node, the elements along any row of a rate matrix must sum to zero. Up to now, we have associated rate matrices with transmission schemes, which describe the state of the network at a given point in time. However, as time progresses, the network will operate under a time division schedule that alternates between different schemes. By construction, a time division schedule of transmission schemes is described by the weighted sum of the corresponding rate matrices with weights equal to the percentage of time that each scheme is in operation. Therefore if
N i=1
a time division schedule, then its rate matrix will be R = property holds:
ai Si (with ai 0 and
N i=1
N i=1
ai = 1) is
ai Ri , where R1 , . . . , RN
are the rate matrices of the schemes S1 , . . . , SN . More succinctly, the following linearity
n n
R(
i=1
ai Si ) =
ai R(Si ).
i=1
For example, the rate matrices of schedules T1 = 0.5S1 + 0.5S2 and T2 = 0.75S1 + 0.25S2 will be 7.5 5 0 5 0 0 0 0 0 , 0.75R1 + 0.25R2 = 0 0.5R1 + 0.5R2 = 2.5 5 0 5 0 0 0 0 0 0 ows of the respective schedules. 5 0 2.5 0 0
0 , 0 7.5 5 0 0 0
and by comparing them with Fig. 3.2 we see that they correctly reproduce the information
3.2.3 Ad Hoc Network Capacity Regions and Uniform Capacity

We use the term transmission protocol to describe a collection of rules that nodes must satisfy when transmitting. For example, a transmission protocol could be that nodes are only allowed to transmit their own information, must transmit with their maximum power, can transmit simultaneously with other nodes, and treat interfering transmissions as noise (successive interference cancellation is not allowed). Under a given transmission protocol,
19
a collection of schemes is available to the network. Each of these schemes has a rate matrix. We refer to these as the basic rate matrices. Clearly, the less restrictive the transmission protocol, the more schemes are available, and the larger the collection of basic rate matrices. Since weighted sums of rate matrices describe the net ow of information in the network under a corresponding time division schedule, we could dene the capacity of the network under time division and a given transmission protocol as the set of weighted sums of all basic rate matrices of this protocol (with the coefcients being positive, and their sum being equal to unity). However, some weighted sums of rate matrices have off-diagonal elements that are negative. Such rate matrices correspond to scenarios where some nodes forward more information from a source than they receive from that source (possibly indirectly, through routing). Clearly, this is not a stable condition, and we therefore exclude these sums from the capacity region. All other weighted sums of basic rate matrices can be included in the network capacity region. Formally, based on the above discussion, we dene the capacity region of the wireless ad hoc network, under time division routing and a given transmission protocol, as the convex hull of the basic rate matrices with the restriction that the weighted sums must have nonnegative off-diagonal elements. Specically, if {R1 , . . . , RN } denotes the set of basic rate matrices, the capacity region is C = C({Ri })
N N
Co({Ri }) Pn = {
i=1
ai Ri : ai 0,
i=1
ai = 1} Pn
negative and Co{Ri } denotes the convex hull of the set {Ri } of basic matrices. The shape network topology and parameters, and also on the transmission protocol. The total number of basic rate matrices N is a function on the number of nodes and the transmission protocol. As will be shown in Section 3.6, to calculate if a point belongs in the capacity region it is necessary that all basic rate matrices that contribute to the capacity
where Pn is the subset of all n n matrices with all their off-diagonal elements non-
of the capacity region depends on the pool of basic rate matrices. This pool depends on the
region be created explicitly. It is therefore critical that their number is nite. Consequently, it is necessary that the transmission protocol does not permit a continuous variation of
20
transmission parameters, if this would lead to an innite number of matrices that contribute to the capacity region. Although the formulation can handle an arbitrarily large number of basic rate matrices, practical computational issues place a limitation on their number. We will elaborate on this point in Section 3.6. The meaning of the capacity region is the following: Let R = [rij ] be a matrix in the capacity region. Then there is a time division schedule of transmission schemes that are valid under the transmission protocol such that when the network operates under this time division and i = j, rij is the rate with which node Xi sends its own information to node information to all other nodes. Xj , possibly through multiple hops, and rii is the total rate with which node Xi is passing Since the elements in each row of all matrices in the capacity region must sum to 0, the capacity region is contained in a n(n1) subspace. This dimensionality is expected, since there are n nodes, each with (n 1) other nodes with which it may want to communicate. n(n 1) off-diagonal elements. Each of the n(n 1) possible communication pairs corresponds to exactly one of the
Note that the shape of the capacity region does not depend on the exact model of the
packet arrival process: For any rate matrix that belongs in the capacity region, and for any (probabilistic or not) arrival process for which the time averaged arrival rates at the source division and multihop routing schedule such that the packet delivery rates at the destination nodes also converge to the corresponding elements of the rate matrix. The reason for this lack of dependence on the arrival process is that the time division and multihop routing schedule is computed centrally, is based only on the long term averages of the arrival process and is not affected by isolated arrivals or even secondary statistics. Thus for every source-destination pair there is a guaranteed virtual pipeline of packets of capacity equal to the corresponding rate matrix entry. If a node is scheduled to transmit but has no packet, its opportunity will be wasted, however such lost opportunities will be occurring so infrequently that the long term delivery rates will not be affected. In addition, we have placed no delay constraint, and all nodes are equipped with buffers of innite length, so packets can remain in the system awaiting delivery indenitely. To capture the capacity of an ad hoc network with a simple gure of merit, we dene the uniform capacity Cu of a network under time division routing and a given transmission nodes for the n(n 1) streams converge to the elements of the rate matrix, there is a time
3.3. CAPACITY REGIONS FOR VARIOUS TRANSMISSION PROTOCOLS
21
5
0.5
y(m)
4 1 3
0.5
1 10 5 0 5 10
x(m)
Figure 3.3: An example wireless ad hoc network with 5 nodes. protocol as the maximum aggregate communication rate of the network assuming that all nodes wish to communicate with all other nodes at a common rate. The uniform capacity off-diagonal elements equal to r belongs to the capacity region, and n(n 1) is the total number of source-destination pairs for a network of n nodes. The uniform capacity may be is equal to rmax n(n 1), where rmax is the largest r for which the matrix with all its
thought of as the maximum possible aggregate throughput under uniform trafc conditions.
3.3 Capacity Regions for Various Transmission Protocols

In this section we dene a sequence of capacity regions, each corresponding to a progressively more sophisticated transmission protocol (and consequently a richer pool of basic rate matrices), and then study them in the context of an example ad hoc network. This will illustrate the capacity gains that can be obtained by using these protocols. The example network, shown in Fig. 3.3, consists of 5 nodes, and has a random topology ob10 m, 10 m y 10 m}. The power gains between nodes Xi and Xj are given by Gij = KSij ( d0 ) dij tained by uniformly and independently distributing ve nodes in the box {10 m x
(3.1)
22
where dij is the distance between the nodes, K and d0 are normalization constants set to K = 106 and d0 = 10 m respectively, the path loss exponent is set to = 4, and the shadowing factors Sij = Sji are random, independent, and identically generated from a lognormal distribution with a mean of 0 dB and variance = 8 dB (so Sij = 10
Nij 10
and Nij is Gaussian with expectation E[Nij ] = 0 and standard deviation Nij = 8). The receiver function is given by (2.3), with = 1. The power gains are assumed to be time invariant: this assumption will be dropped in Section 3.8. The transmitter powers are xed at Pi = 0.1 W (this restriction is relaxed later, in Section 3.3.4), all receivers are subject to thermal noise with the same power spectral density = 1010 W/Hz, and the bandwidth is W = 106 Hz.
3.3.1 Single Hop Routing, No Spatial Reuse

We rst determine the capacity region when only single hop routing is allowed (no forwarding) and only one node is transmitting at any time. By only allowing one active node at a time, link data rates are higher since there is no interference, but the network does not take advantage of spatial reuse. Since there are n nodes in the system, and each of them has n 1 possible receivers, the network has N a = n(n 1) + 1 transmission schemes (including the one in which all nodes remain silent), and associated basic rate matrices
a Ri , i = 1, . . . , N a . One of these, the zero rate matrix, will correspond to the scheme
where all nodes remain silent. All elements of the zero rate matrix will be identically equal to zero. Determining the rest of the basic matrices is straightforward by applying (2.1) and (2.3). The capacity region will therefore be:
C a = Co{Ri , i = 1, . . . , N a } Pn .
In Fig. 3.4, we have drawn a two-dimensional slice of C a along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j (line (a)). This line captures a background rate of zero for send data: the other nodes never transmit since they do not have data of their own to send and, in single hop routing, they cannot help in forwarding. Note that the slice is a straight line, as expected, since without spatial reuse at any one time the network supports the all node pairs other than (X1 X2 ) and (X3 X4 ). Therefore only nodes X1 and X3

2.5 (d) 2 (c) r34 (Mbps) 1.5 (b) 1 (e)
23
0.5
(a)
0 0
0.5
1.5 r12 (Mbps)
2.5
Figure 3.4: Capacity region slices of the example ad hoc network along the plane rij = 0, (i, j) =
(1, 2), (3, 4), i = j: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c).
information transfer of only one source-destination pair. By changing the time percentages that are devoted to each of the two node pairs, different points on the straight line (a) will
a be achieved. The uniform capacity of the network is Cu = 0.83 Mbps.
3.3.2 Multihop Routing, No Spatial Reuse

Next we consider the case where multihop routing is allowed, but spatial reuse is not, so only one node is transmitting at a given time. Since there are n nodes in the systems, and each has n 1 different possible receivers and n possible nodes to forward data for schemes (including the one in which all nodes remain silent) and associated rate matrices (2.3). The capacity region under these assumptions will be:
b C b = Co{Ri , i = 1, . . . , N b } Pn .
(including itself and the receiver), there are now N b = n2 (n 1) + 1 possible transmission
b Ri , i = 1, . . . , N b . Determining these matrices is straightforward by applying (2.1) and
In Fig. 3.4 we have drawn a slice (line (b)) of C b along the plane rij = 0, (i, j) =
24
(1, 2), (3, 4), i = j. We note that this slice is again a straight line, as expected, since without spatial reuse at any one time the network supports information transfer between only one source-destination pair. We also note a signicant increase in the size of the capacity region as compared with the previous case. This is due to the fact that under multihop routing the nodes can avoid transmitting directly to their destination over paths with small gains, and instead use multiple hops over channels with much more favorable gains and correspondingly higher rates. Uniform capacity is also increased by 242%, to
b Cu = 2.85 Mbps.
3.3.3 Multihop Routing with Spatial Reuse

We now allow both multihop routing and spatial reuse. In this case a network of n nodes will have Nc =
n 2 i=1
n(n 1) . . . (n 2i + 1) i n +1 i!
(3.2)
distinct transmission schemes. Indeed, the i th term in the above sum is the total number choices for the 2i nodes that are involved, however this number must be divided by i! to
of schemes having i transmit-receive pairs. There are n(n 1) . . . (n 2i + 1) distinct
account for the fact that pair orderings are unimportant. The total number of pair combinations is multiplied by ni to account for the different possibilities in the choice of information
c sources, for each of the pairs. Denoting the basic rate matrices by Ri , i = 1, . . . , N c , we
can dene the capacity region C c as

c C c = Co{Ri , i = 1, . . . , N c } Pn .
In Fig. 3.4 we have drawn a slice of C c (line (c)) along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j. We note that the slice is no longer a straight line, as the network can now use spatial reuse to maintain multiple active transmissions, and at any time instant more than one stream may be serviced (directly or along a multihop route). The introducc tion of spatial reuse increases uniform capacity by 26% to Cu = 3.58 Mbps, even for this
small network of ve nodes.
25
3.3.4 Power Control

We have so far assumed that nodes either transmit at their maximum power or remain silent. If we relax this condition and allow each node to transmit at different power levels below the maximum power, then we increase the set of basic rate matrices and thereby, possibly, augment the capacity region. Since an innite number of power levels implies an innite number of rate matrices, which our formulation cannot handle, we necessarily restrict ourselves to a nite number of power levels 1 . In particular, node Xi can transmit at any of the p possible power levels of the set { 1 Pi , 2 Pi , . . . , Pi }. The network will then p p have a set of Nd =
n 2 i=1
n(n 1) . . . (n 2i + 1) i i np +1 i!
(3.3)
basic rate matrices. (3.3) is derived from (3.2), with the factor pi being added to account for the different possible power control scenarios, when i pairs are active. Denoting the set
d of basic rate matrices by Ri , i = 1, . . . , N d , the capacity region now becomes: d C d = Co{Ri , i = 1, . . . , N d } Pn .
In line (d) of Fig. 3.4 we have drawn a slice of C d along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j, for power control with two levels (p = 2). We observe that this simple power control strategy does not signicantly change this slice of the capacity region.
d Moreover, the uniform capacity changes less than 1%, to Cu = 3.61 Mbps. As expected,
additional power levels lead to negligible gains. This result is consistent with other results on variable-rate transmission with power control, which indicate that if the transmission rate is adjusted to the link SINR, additional power control does not signicantly improve performance [48]. On the other hand, as we will see later on, power control leads to signicant gains when variable-rate transmission is not possible. In other words, if the system has one degree of freedom that it can use to adapt (rate adaptation), a second degree of freedom (power adaptation) will not lead to dramatic gains.
in Chapter 4 it will be shown that the formulation can handle continuous power control, provided the receiver function fR (), introduced in (2.2), is a step function.
26
3.3.5 Successive Interference Cancellation (SIC)

Until now, we have assumed that under transmission schemes with many simultaneous transmissions each node decodes only its intended signal and treats all other signals as noise. However, under a successive interference cancellation (SIC) strategy, nodes may decode some signals intended for other nodes rst, subtract out this interference, and then decode their own signals. This strategy may cause a node to restrict the transmission rate of an interfering node, since the receiving node must be able to decode the interfering signal. (For example, under Shannon analysis, the interfering signal is assumed to be decoded perfectly as long as the rate at which the interfering signal is set to is less than the capacity of the interfering link.) However, this restriction is balanced by the fact that the receiving nodes rate will increase due to the removal of interference. For example, consider the setup of Fig. 3.1(ii). Node X3 s signal will interfere with node X2 s reception. X2 could decode X1 s signal and treat X3 s signal as noise, or could rst decode and remove the signal from node X3 and then decode the desired signal from node X1 . This second decoding strategy will impose an additional constraint on the transmission rate of node X3 , since its signal will also need to be decoded at X2 . A similar choice of strategies exists for X4 . Assuming a network of n nodes, power control with p levels, and successive interference cancellation, the total number of transmission schemes and associated basic rate matrices will be Ne =
n 2 i=1
n(n 1) . . . (n 2i + 1) i i i n p Z (i) + 1 i!
(3.4)
(3.4) is derived as (3.3), with the factor Z(i)i being added to account for all the possible combinations of successive interference strategies. Z(i) represents the total number of SIC strategies available to a specic receiver in the presence of i transmissions: the user may decode its own signal rst (1 scenario) or second (i1 scenarios, depending on what signal is decoded rst), or third ((i 1)(i 2) scenarios), and so on. Denoting the set of basic
e rate matrices by Ri , i = 1, . . . , N e , the capacity region is
where Z(i) = 1 2 . . . (i 1) + 2 . . . (i 1) + . . . + (i 1) + 1. Equation
e C e = Co{Ri , i = 1, . . . , N e } Pn .
3.4. UNIFORM CAPACITY OF CANONICAL TOPOLOGIES
27
3 (e) 2.5 Uniform Capacity, Cu (Mbps) Uniform Capacity, Cu (Mbps) (c) (d)
3.5 3 (c) 2.5 2 1.5 1 0.5 0 (d) (e) (b) (a)
1.5
(b)
0.5
(a)
4 5 Number of nodes, n
4 5 Number of nodes, n
(i)
(ii)
Figure 3.5: Uniform capacity of two canonical topologies as a function of the number of nodes: (i) Linear networks. (ii) Ring networks. (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c). Line (e) of Fig. 3.4 shows a slice of C e for the example ad hoc network, with no power control (p = 1), along the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j. This slice indicates that SIC signicantly augments the capacity region even without power control. Moreover,
e c the uniform capacity becomes Cu = 4.31 Mbps, 19% greater than Cu that corresponds to
spatial reuse and multiple hops, but no SIC.
3.4 Uniform Capacity of Canonical Topologies

In the previous section we studied the capacity regions of a network with a random topology for a sequence of transmission protocols, and determined that multihop routing, spatial reuse, and successive interference cancellation all improve the performance signicantly. On the other hand, adding power control yields very small gains. We have determined that these results are general, and hold with little variation for a wide range of random or canonical topologies and modeling parameters. For example, in Fig. 3.5 we plot the uniform capacity versus the number of nodes for the ve transmission protocols introduced in Sections 3.3.1 to 3.3.5, and for two canonical topologies: The rst, which we dene as the linear topology, consists of nodes arranged
28
on a linear array, with a spacing of 10 m between them. The second, which we call the ring topology, consists of nodes arranged on a circle with a radius of 10 m. The nodes are separated by arches of equal length. In both cases, the bandwidth available to the system is W = 1 MHz, all nodes have the same maximum power of Pi = 0.1 W, and the thermal noise power spectral density is the same for all receivers and equal to = 1014
W . Hz
We calculate channel gains using the model of (3.1), with the parameters set to = 3, d0 = 10 m, K = 106 and = 0 dB (so that the topology, not random shadowing, dominates the performance). We note that the trend of the curves strongly depends on the topology. In the case of linear networks, the performance of the transmission protocols that do not allow spatial reuse deteriorates fast as the number of nodes increase. The rest of the protocols perform signicantly better, but the overall performance decreases with the number of nodes. This means that the loss from the need for multiple hops is greater than the gain from improved spatial separation. In the case of ring networks comparing lines (a) and (b) reveals that gains from allowing multiple hops are limited. This is due to the fact that nodes are clustered together in a ring formation. Still, the performance of the protocols that allow concurrent transmissions actually improves as we add more nodes, which implies that networks can take advantage of spatial separation even when more nodes are placed in the same area. This nding is reminiscent of the results appearing in [36]. We have also established that the relative performance shown in the previous gures does not change under a wide range of the transmitter powers, thermal noise powers and the exponential decay parameter . Successive interference cancellation was found to be particularly effective in increasing capacity in the case where channels have comparable link gains, for example when is small ( 2).
3.5 Discrete-Rate Transmission

Until now, we have assumed that the transceivers used are capable of variable-rate adaptation, as dened by (2.3) with = 1, meaning that transmitters automatically adjust the transmission rate to match the SINR at the receiver and achieve the Shannon bound. Such an assumption implies that the available rates are not restricted to a nite set of values. We now consider the effects of a discrete-rate restriction on our capacity calculations.
3.5. DISCRETE-RATE TRANSMISSION
29
7 6
5 4.5 4 (e)
Achievable rate, R (bps/Hz)
5 4 (e) 3 2 1
(d)
Uniform capacity, Cu (Mbps)
(c)
3.5 3 2.5 2 1.5 1 0.5 (d)
(d, d)
(b)
(b) (a)
(c) (a)
0 10
5 SINR (dB)
10
15
20
7 Rate levels, L
15
(i)
(ii)
Figure 3.6: (i) A class of stepwise receiver functions with 1, 3, 7, and 15 rate levels (plots (a), (b),
(c), and (d) respectively). The Shannon bound of (2.3) is also plotted (plot (e)). (ii) Uniform capacity versus the number of available rate levels for the example ad hoc network. Each curve corresponds to a different transmission protocol: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d,d,d) Power control with 2, 3, and 4 levels added to (c). (e) Successive interference cancellation added to (c).
We consider as possible candidates for fR () the four step-wise functions of Fig. 3.6(i). Each has a nite number of steps L, which is set to L = 1, 3, 7, 15 respectively. Each step corresponds to a distinct rate, which has a minimum SINR requirement based on the Shannon capacity formula (2.3). The transmitter chooses the highest rate within the set of possible rates for which the SINR requirement is met. In Fig. 3.6(ii) we plot the uniform capacities of the example ad hoc network of 5 nodes, for the protocols introduced in Sections 3.3.1 to 3.3.5, and for the four different receiver functions fR () of Fig. 3.6(i). (receiver function were introduced in (2.2).) We compare their performance with the performance achieved when there is no rate restriction, as in Sections 3.3.1 to 3.3.5. As expected, increasing the number of available rates improves capacity. More importantly, power control now leads to signicant gains: for example when only one rate is available, if added to spatial reuse and multihop routing, power control with 2 power levels raises the uniform capacity by 26%, and outperforms successive interference cancellation. Power control with 3 levels raises the uniform capacity by 31%, but using more than three levels leads to negligible gains. This result is consistent with other ndings in the literature [48] and shows that power control can be interchanged to some extent with
30
rate adaptation. The performance enhancements achieved by using power control when time division is not allowed have been studied extensively [42]; our results also shows that power control leads to improved network capacity, even when time division is factored in, provided that variable-rate transmission is not being used.
3.6 Computational Issues

convex hull of the networks basic rate matrices R1 , R2 , . . . , RN . Therefore, checking if a point is in the networks capacity region is equivalent to checking if the point belongs to matrices. Since rate matrices are isomorphic to vectors of length n(n1), this is a standard Pn , which is trivial, and checking if it belongs to the convex hull of the networks basic rate We have dened the capacity region of a network as the intersection of the set Pn with the
problem in computational geometry, and can be solved by a variety of different techniques. space: We choose to cast it as the following linear program in the (N 1)-dimensional Euclidean minimize: g(x) = subject to:
N 1 i=1
xi (3.5)
R=
0 xi 1
N 1 i=1
xi Ri .
zero rate matrix (that corresponds to all nodes being silent). If the problem is feasible (i.e.
{R1 , . . . , RN } is the set of all basic rate matrices for the network, and RN specically is the
the set of points satisfying the constraints is not empty) with g(xopt ) 1, then R belongs to the capacity region, and can be obtained via a time division schedule of the basic rate matrices with the zero rate matrix RN being active 100(1 g(xopt )) percent of the time.
If the problem is infeasible (i.e. the constraints cannot be satised for any point x) or it is feasible with g(xopt ) > 1, R does not belong to the capacity region. By iteratively solving the linear problem (3.5) we can determine boundary points of the capacity regions. These boundary points correspond to optimal modes of operation for the network. Such optimal modes require coordination among nodes on a global level. Indeed,
each mode of operation is a time division schedule consisting of a number of schemes, each being active for a certain time percentage. While a particular scheme is active, all nodes, however distant, must transmit with the power and to the destination specied by
3.6. COMPUTATIONAL ISSUES
31
SCHEME A (40.64%) 2 5
SCHEME B (26.88%) 2 5
6.16 0.63
SCHEME C (6.68%) 2 5
1.60
SCHEME D (25.78%) 2 (decodes 5

4 first) 3.47 1.90
1.60
2.45
4
(decodes 1 first)
4
7.32
Figure 3.7: The time division schedule that maximizes r12 = r34 , if the rest of the node pairs have
no communication requirements, for the example ad hoc network of Fig. 3.3. Arrows signify actual transmissions. Numbers denote link rates (in Mbps).
that scheme. Therefore, the optimal schedules must be computed in a central location and then distributed to the nodes; distributed routing and media access algorithms in general cannot achieve the capacity. It should also not be a surprise that the optimal schedules will typically involve non-FIFO queuing. In addition, packets belonging to a given sourcedestination will typically travel through many different routes. These points will become more clear in the following chapter. For example, if we wish to maximize rates r12 and r34 with r12 = r34 and allow spatial separation, multiple hops, and successive interference cancellation, but no power control, the optimal time division schedule is that of Fig. 3.7 that achieves r12 = r34 = 1.6564 Mbps. This gure shows that the developed method may be viewed as a solution to the optimal routing problem when links interfere with each other, and the communication needs of the network (in terms of the streams that must be serviced and the rates that are required) are arbitrary. Note that in order to determine the capacity region, the set of all basic rate matrices that can potentially affect the region must rst be calculated. As the number of nodes increases, the number of basic rate matrices increases very fast (for example, factorially for the protocol that allows multihop routing and spatial reuse), as shown in Fig. 3.8. Since not all matrices will contribute to the capacity region, signicant speed gains can be realized by carefully constructing a minimal set of rate matrices that sufciently describes it. When the transmitters follow the Shannon capacity limit of (2.3), the MATLAB/C routines we have developed become impractical for networks with more than 7-8 nodes when successive interference cancellation is not used, or 5-6 nodes when successive interference
32

25
log10(# of rate matrices)
20
(d)
15
10
(e)
5
(c) (b) (a)
10
Number of nodes, n
Figure 3.8: The total number of rate matrices that exist versus the number of nodes n and various
protocols: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse and 1 4 distinct power levels. (d) Successive interference cancellation added to (c). (e) The maximum number of rate matrices that our current software and hardware can create. Although the number of total rate matrices increases very fast with the number of nodes, it is not necessary to explicitly create all of them. A small subset of them will in reality contribute to the capacity region, and only those must be explicitly created.
cancellation is used. When variable-rate transmission is not used, and the transmitters can only operate with a single xed rate (provided the SINR is larger than some threshold), the developed routines become impractical for networks with more than around 15 nodes. We are currently developing more sophisticated software that, together with faster hardware, will increase these bounds by 3-4 nodes in all cases.
3.7 Multihop Cellular Networks

We dene multihop cellular networks as networks in which one node (we assume Xn ) acts as a base station, and the rest of the nodes (X1 to Xn1 ) can transmit to each other, but are only interested in sending to or receiving from the base station. The developed formulation can be readily applied to the case of multihop cellular networks. However, because of the special structure of the network, a simplied formulation can be used to solve for the capacity region. For the case of the uplink communication, instead of dening the rate matrix of a scheme S, we dene the rate vector R(S) = [r1 r2 . . . rn1 ]T in the following
3.7. MULTIHOP CELLULAR NETWORKS
33
manner: r, ri = r, 0, If node Xi transmits information to a node (ultimately intended for the base station Xn ) at rate r. If node Xi receives information from a node (ultimately intended for the base station Xn ) at rate r. otherwise.
A similar formulation exists for the downlink case. We note that, contrary to the rate matrix formulation, a positive entry signies transmission of information rather than reception. As in the rate matrix formulation, rate vectors contain all the information that is essential for describing the state of the network at a given time. The rest of the theory can be developed in a manner entirely analogous to the general case. In particular, because of the construction of rate vectors, a weighted time division schedule of transmission schemes is described by a rate vector equal to the weighted sum of the corresponding rate vectors, with the same weights. Therefore, if a multihop cellular network with n nodes operates under a transmission protocol associated with a set of basic rate vectors {R1 , . . . , RN }, the capacity region is
N N
C = C({Ri })
Co({Ri }) Pn1 = {
i=1
ai Ri : ai 0,
i=1
ai = 1} Pn1
with Pn1 being the set of all vectors of length n 1 with non-negative elements. Rate
vectors with negative elements are excluded because they correspond to schedules in which some nodes consistently receive more information than they transmit, so data ultimately intended for the base station are accumulated to intermediate nodes. The feasibility problem can be solved with a linear program identical to (3.5). As in the general case, the capacity region depends on the repertoire of basic rate vectors. The set of basic vectors depends on the network topology and parameters, and also on the transmission protocol. The meaning of the capacity region is the following: Let R = [ri ] be a vector in the capacity region. Then there is a time division of schemes, allowed by the transmission protocol, such that when the network operates under this time division, ri is the rate with
34

2.5
10 3 2 5 4
r5 (Mbps) 2 (e) 1.5
5 y (m) 1 0 7(BS) 5 6 10 10 5 0 x (m) 5 10
(d) (b) (c)
0.5
(a)
0 0
3 4 r (Mbps)
3
(i)
(ii)
Figure 3.9: (i) An example multihop cellular network with 6 nodes and a base station. (ii) Capacity
region slices of the example multihop cellular network along the plane ri = 0, i = 3, 5: (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c).
which node Xi sends its own information to the base station, possibly through multiple hops and time division. To capture the capacity of a multihop network with a simple gure of merit, we dene the uniform capacity Cu of a network as the maximum aggregate communication rate, if all nodes wish to communicate with the base station with a common rate. The uniform capacity is equal to rmax (n 1), where rmax is the largest r for which the vector with all its elements equal to r belongs to the capacity region. As an example, in Fig. 3.9(i) we plot a multihop cellular network, created by placing 6 y 10 m} and the base station at the origin. The rest of the parameters were chosen as nodes randomly, independently and uniformly in the box {10 m x 10 m, 10 m
in the network of Section 3.3. Slices of the capacity regions for the transmission protocols introduced in Sections 3.3.1 to 3.3.5 for the general case appear in Fig. 3.9(ii). The uniform
a b capacities for the same transmission protocols were Cu = 1.52 Mbps, Cu = 3.01 Mbps,
c d e Cu = 3.63 Mbps, Cu = 3.67 Mbps, and Cu = 4.22 Mbps.
We have also studied various other random and canonical topologies and for a wide range of the modeling parameters. In all cases, it was determined that multihop cellular
3.8. FADING AND MOBILITY INCREASE CAPACITY
35
networks behave similarly to ad hoc networks: multihop routing, spatial reuse and successive interference cancellation greatly increase the capacity of the network; improvements from power control are signicant only if variable-rate transmission is not used.
3.8 Fading and Mobility Increase Capacity

3.8.1 Time-Varying Flat Fading Channels
With minor modications, our formulation can handle the case of time-varying at-fading channels. We return to the general case of ad hoc networks and start by considering a network with a channel that alternates between M different fading states, each being comcollection of basic rate matrices for the i-th fading state (or a sub-collection of them with the same capacity region, as discussed in Section 3.6). Assuming that, over a long period of time, each state will be present for a fraction of time equal to the network can be dened as
M J(i) 1 , M
pletely described by a gain matrix Gi , i = 1, . . . , M. Let {Rij , j = 1, . . . , J(i)} be the
the capacity region of
C={
aij R
i=1 j=1
ij
1 : aij 0, aij , M j=1

J(i) j=1
J(i)
J(i)
i=i j=1
aij = 1} Pn .
(3.6)
The additional constraint
should use the basic matrices of a given state for longer than the time for which the state is available. The capacity region is no longer dened as the convex hull of a collection of matrices, however we can easily see that it is still convex. In dening the capacity region as in (3.6), we have intrinsically assumed that nodes schedule their transmissions to take advantage of the different fading states and so are willing to tolerate the random delays associated with waiting for all the required states to come up. Depending on the fading statistics, these delays can be large. Also note that the formulation can be trivially generalized to describe the more general case where each fading state is active for an arbitrary fraction of time, other than must sum up to unity.) Assuming that RiJ(i) = R0 , i = 1, . . . , M with R0 being the zero rate matrix, the
1 . M
aij
1 M
simply means that no time division schedule
(Of course all fractions
36

6
5 (b) (e) (c) (d) (a)
4 r34 (Mbps)
0 0
0.5
1.5 2 r12 (Mbps)
2.5
3.5
Figure 3.10: A slice of the capacity region C c for the example ad hoc network of Section 3.3 along
the plane rij = 0, (i, j) = (1, 2), (3, 4), i = j for different combinations of fading states: (a) Fading state F 1. (b) Fading state F 2. (c) Fading states F 1, F 2. (d) Fading states F 1, . . . , F 10. (e) Fading states F 1, . . . , F 15.
feasibility problem can be solved by a linear program similar to (3.5):

J(i)1 xij minimize: g(x) = M j=1 i=1 J(i)1 R= M xij Rij , j=1 i=1 1 subject to: 0 xi M , 1 i M, J(i)1 1 xij M , 1 i M. j=1
If the linear program is feasible (i.e. the set of points satisfying the constraints is not empty) with g(xopt ) 1, then R can be obtained via a time-division schedule of the basic rate matrices {Rij } and the silent rate matrix R0 that will be active 100(1 g(xopt )) percent of the time. Otherwise, R does not belong to the capacity region.
Until now, we assumed that the network operates under a nite number of fading states. This assumption is restrictive, since in most realistic fading models the fading gains can have arbitrary values, so that the total number of fading states is innite, and the developed formulation becomes inadequate. Furthermore, even if we use a model with a nite number of fading states s, because the gain matrix has n(n 1) non-trivial entries, the resulting number of possible states is sn(n1) , which is intractable for non-trivial values of s and n.
37
In these cases, the capacity region may be estimated by using a large, but tractable, number of fading states. In this vein, we rst create a sequence of independent fading states, according to the statistics of the fading. Then, we use a Monte-Carlo approach, and calculate the capacity region as the number of fading states increases, until we observe convergence. Since capacity regions are multi-dimensional, we use convergence of the uniform capacity to indicate capacity region convergence2 . In Fig. 3.10 we display the convergence of a slice of the capacity region C c for the network of Section 3.3 as the number of fading states increases. The gain matrices for each of these fading states were created independently, by using the model of (3.1), with
c = 8 dB. The uniform capacity converges at the value Cu = 5.9 Mbps which is 120%
larger than the value of the uniform capacity with no time-varying at fading and = 0 dB,
c which is Cu = 2.7 Mbps. This value is also 90% larger than the value of the uniform c capacity Cu = 3.1 Mbps achieved when there is time-varying at fading with = 8 dB,
but the nodes do not schedule their transmissions across different fading states, and pick an optimal schedule within each fading state (i.e. as if the current fading state is the only fading state available). The 90% increase is possible because nodes are willing to tolerate large delays in order to exploit the changing fading states. In Fig. 3.11 we plot the uniform capacity as a function of the shadowing standard deviation , in the case of a ring ad hoc network with 5 nodes and in the presence of time-varying at fading. We note that the uniform capacity increases with the standard deviation, which is a measure of the severity of the fading. This result should actually be expected, since in the presence of fading the network has more degrees of freedom when deciding on the optimal transmission schedule. From another perspective, each fading state will be more favorable for certain combinations of transmissions, and the network will use this state for these combinations. We note that allowing successive interference cancellation in addition to spatial reuse yields limited gains when fading is severe ( = 10 dB). This is because under severe fading for most of the time the network has the opportunity to operate using simultaneous transmissions with very little cross interference. In these cases, successive interference
Since the load of calculations required for the determination of the uniform capacity increases as more states are added, we are practically limited to around 40 states for protocols that allow concurrent transmissions. In all cases we were able to determine the limit with a tolerance of 5% or better.
2
38

9 8 Uniform capacity, Cu (Mbps) 7 6 (c) 5 4 3 2 1 0 0 5 Fading standard deviation, (db) 10 (e) (a) (d) (b)
Figure 3.11: Estimates of the uniform capacity of various protocols, for a ring ad hoc network with 5 nodes, as a function of the standard deviation of the shadowing component. Estimates are accurate within 0.2 Mbps. (a) Single hop routing, no spatial reuse. (b) Multihop routing, no spatial reuse. (c) Multihop routing with spatial reuse. (d) Two level power control added to (c). (e) Successive interference cancellation added to (c). cancellation is ineffective because it imposes very stringent upper bounds on the rates of the interferers.
3.8.2 Node Mobility

We can study node mobility using the framework developed for the case of time-varying at fading channels. Indeed, the critical element of both settings is that the gain matrix alternates between a number of different values, as time progresses. Consider an ad hoc network of n nodes X1 , X2 , . . . , Xn where the channel gains are deterministic and depend only on distances (for example, path gains are given by (3.1) with = 0 dB) but the spatial conguration alternates between M different states. Assuming that, after long periods of time, each state is active for a fraction of time equal to
1 , M
the
capacity region of the network is again given by (3.6), where each of the M sets of rate spatial conguration. matrices {Rij , j = 1, . . . , J(i)} does not describe a different fading state, but a different The assumption that there is a nite number M of congurations is rather articial, and a more realistic model would be that nodes move continuously. This would in turn
39
12 11 Uniform capacity estimate (Mbps) 10 9 8 7 6 5 4 3 2 0 5 10 15 20 25 Spatial configurations 30 35 40 Cc (no spatial scheduling) = 3.7 Mbps u Cc = 11.1 Mbps u
c Figure 3.12: Convergence of the estimate for the uniform capacity Cu with node mobility, for the
example ad hoc network of Section 3.3, with = 0 dB.
imply that the number of possible states (and hence the number of rate matrices) is innite, and as a consequence the developed formulation is insufcient. However, the capacity region can be estimated by using a large, but nite, number of spatial congurations, in the same way as a nite number of fading states was used in the case of time-varying at fading. Specically, let us assume that the vector stochastic process describing the node movement is stationary and ergodic (for example, two-dimensional independent Brownian walks constrained to a box). Then, instead of using all states, we can use a large number of independent realizations of the node positions to estimate the capacity region. In Fig. 3.12 we plot the estimate for the uniform capacity versus the number of spatial congurations used for an ad hoc network of ve nodes. All parameters are set as in the example ad hoc network of Section 3.3 with = 0 dB (i.e., no time-varying at fading). The transmission protocol allows multiple hop routing and spatial reuse, but no power control or successive interference cancellation. The node movements are modeled as independent two-dimensional standard Brownian motions constrained to the box {10 m x 10 m, 10 m y 10 m}, so we can create a realization of the process at a xed time by distributing the nodes randomly, uniformly and independently in the box.
c The uniform capacity converges to the value Cu = 11.1 Mbps, after including around 40
40
spatial congurations. In the gure we also plot the uniform capacity that is achieved by the network if scheduling across different spatial congurations is not used, so that at any time instant nodes are using only the current conguration in the most efcient way to comc municate, and are not exploiting node mobility. In this case the capacity is Cu = 3.7 Mbps.
However, the 200% increase achieved by using scheduling across spatial congurations is only possible if nodes are willing to tolerate large delays. This tradeoff between delay and throughput, under node mobility, is reminiscent of the results in [38].
3.9 Energy Constrained Networks

We now apply the methodology of the previous sections to develop a formulation suitable for the study of energy constrained networks. In such networks, nodes have a nite amount of energy that they can use for transmitting or receiving [7]. As a consequence, continuous communication with a constant non-zero rate is not possible, but rather the goal is to maximize the total number of bits communicated.
3.9.1 System Model

We consider a network of n nodes X1 , X2 , . . . , Xn , described by the system model of Chapter 2. For simplicity, we assume that Xn acts as a base station, and the rest of the nodes wish to send information to it (modifying the formulation to treat the downlink case or the more general ad hoc case is straightforward). In addition, each node Xi , i = 1, . . . , n 1 has some nite amount of energy Ei to use for its communication needs. Let E = [E1 E2 . . . En1 ]T be the energy vector of the network. We make the assumption that, under a given transmission scheme T , when node Xi
transmits to node Xj with power Pi for t seconds (so that the dissipated energy is Ei = Pi t), the number of bits transmitted is B = fR (ij )t, where ij is the link SINR and fR () is the receiver function, as described in Chapter 2. We assume that the relation B = f (ij )t holds true for any duration t, which implies that the achievable data rate is independent of the number of bits transmitted. Our formulation is not consistent with previous information theoretic approaches [50, 51], since we do not permit the coherent combining of signals, however it is well justied for any practical modulation and coding
3.9. ENERGY CONSTRAINED NETWORKS
41
strategy, as long as the number of bits transmitted is much greater than the bits per symbol (for uncoded modulation), the block length (for block-coded modulation), or the memory of the convolutional coder (for convolutional coding). Note that nodes expend energy not only for transmitting, but also for receiving. We make the assumption that the energy spent to decode a sequence of bits is proportional to its length. Our transmission strategy aims to maximize the number of bits communicated between the nodes and does not consider delay. Thus, since any transmitting node will eventually expend its energy and become silent, there is no gain by allowing concurrent transmissions: If there are non-zero cross-channel gains, some transmissions will experience interference and fewer than the maximum possible bits will be transmitted. If all cross-channel gains are zero, spatial reuse will not hurt, but it will not produce any gains either. Therefore, for our capacity analysis, we assume that no concurrent transmissions take place in the network. On the other hand, using multiple hops still makes sense and can augment the capacity region signicantly.
3.9.2 Capacity Region Formulation

We still dene transmission schemes as in Section 3.2, and describe them in terms of rate vectors, as dened in Section 3.7. In addition, if qi (S) is the power consumed by node Xi vector Q(S) = [q1 (S) q2 (S) . . . qn1 (S)]T . Nodes will typically need most of their power for transmission, however the above denition allows us to factor in the power required to receive signals. We assume that the network operates under a transmission protocol that allows schemes S1 , . . . , SN , corresponding to the rate vectors R1 , R2 , . . . , RN and power dissipation vectors Q1 , Q2 , . . . , QN . Based on the discussion of Section 3.9.1, we assume that the transmission protocol does not permit concurrent transmissions. Depending on the particular communication requirements of the nodes, the network will operate under a succession of some of the N different schemes, each being active for a specic amount of time. With no loss of generality, we assume that each scheme will appear at most once in this succession (separate occurrences can be lumped together). Any succession of schemes is acceptable, during scheme S, then we dene the power dissipation vector of the scheme S to be the
42
as long as the energy dissipation associated with this succession does not exceed the energy resources of the network. This leads to the following denition of the capacity region of the network:
N N
C = C({Ri }, {Qi})
i=1
ti Ri : ti 0,
i=1
ti Qi E} Pn1 .
The capacity region is not a convex hull, but it is trivial to check that it is still convex. actual times.
N i=1 ti Ri ,
Contrary to the previous cases, the coefcients {ti } do not represent time percentages, but The meaning of the capacity region is the following: If B = [bj ] C, with B =
then there is a strategy under which each of the nodes Xj will send bj bits
of data to the base station. The strategy consists of using the schemes corresponding to each Ri for exactly ti seconds. Since in each scheme only one node will be active, we are guaranteed that the schemes can be temporally placed so that nodes used for forwarding data will receive the data before they are required to forward them. As in the previous cases, the feasibility problem can be formulated as a linear program in the N-dimensional Euclidean space. Moreover, the linear program has no objective function: Point B belongs to the capacity region if and only if the following system has a solution {t1 , . . . , tN }: B=
i=1 N N
ti Ri ,
i=1
ti Qi E.
3.9.3 A Numerical Example

In Fig. 3.13(i) we display an energy constrained network with 15 nodes and a random topology, created as the example multihop cellular network of Section 3.7. However, there is no at fading (i.e., = 0 dB) and now the base station is placed on the point {x = 10 m, y = 10 m}. All nodes transmit with power Pi = 0.1 W, but waste no power when they receive. The energy available to each node is Ei = 10 J. We assume that error is always less than 104 , and trellis coding with a coding gain of 3.5 dB is used. Under these assumptions, the achievable rate f (ij ) is well approximated [49] by (2.3) with the transmitters use variable-rate M-QAM, with the requirement that the probability of symbol
3.9. ENERGY CONSTRAINED NETWORKS

8 10 10 5 12 4 y (m) 9 0 2 14 5 1 15(BS) 10 10 6 5 0 x (m) 5 10 0 0 2 4 b5 (Mbits) 6 3 7 11 1 8 13 b10 (Mbits) 7 5 6 5 4 3 2 (a) (b)
43
(i)
(ii)
Figure 3.13: (i) An example energy constrained network with 15 nodes. (ii)Capacity region slices
of the network of Fig. 3.13 along the plane bi = 0, i = 5, 10: (a) Single hop routing. (b) Multihop routing.
gap equal to = 3 dB. is equal to the gap between uncoded M-QAM and capacity (6.5 dB), minus the coding gain (3.5 dB). All other system parameters are chosen as in the network of Section 3.7. In Fig. 3.13(ii) we plot a slice of two capacity regions, along the plane bi = 0, i = 5, 10. The capacity regions correspond to a single hop and a multiple hop transmission protocol, with no spatial reuse, successive interference cancellation or power control. In the single hop case the capacity region is rectangular, since in this case no forwarding is allowed, and the best (and only) strategy is for the two nodes to transmit sequentially to the base station. As expected, multihop routing greatly improves performance. As an example, the point on the boundary of the multihop capacity region for which b5 = b10 is achieved by a transmission strategy consisting of 23 transmissions that actually depletes the energy reserves of all nodes. This result is typical, and shows that optimum strategies aggressively take advantage of all available energy resources to service all data streams. With this transmission strategy both nodes send b5 = b10 = 5.60 Mbps to the base station, which is only slightly less than the number of bits either node would transmit if it were the only one using the network (7.62 Mbps and 7.45 Mbps respectively), although the two
44
nodes are placed close to each other. This result can be explained by noting that there are enough nodes between X5 , X10 and the base station, such that two roughly independent paths to the base station can exist. We have arrived at similar results for various transceiver models, and for a wide range of modeling parameters.
3.10 Conclusions
In this chapter, we have developed a mathematical framework for nding the capacity region of an ad hoc or multihop cellular wireless network under time division routing and a given transmission protocol, possibly in the presence of time-varying at fading or node mobility. We use this framework to determine the network performance that can be obtained using various transmission protocols for a number of different network topologies. We show that multihop routing, spatial reuse and successive interference cancellation all lead to signicant gains, but gains from power control are signicant only if very limited or no rate adaptation is used. We also determine that fading and node mobility can actually improve network capacity. Finally, we introduce a formulation suitable for the study of energy constrained networks. A major limitation of our capacity formulation is the very fast increase of basic rate matrices as the number of nodes increases. As was discussed, the capacity region can be completely reproduced by a much smaller set of rate matrices. Based on this observation, a possible future research direction could be toward developing methods for determining a good enough subset of rate matrices, that will be manageable and also reproduce most of the capacity region. These methods may be deterministic (for example requiring all such rate matrices to have a certain set of characteristics) or stochastic (for example genetic algorithms). Such techniques will allow us to study capacity regions of larger networks.
Chapter 4 Cross-Layer Design of Wireless Ad Hoc Networks

4.1 Introduction
N THE PREVIOUS CHAPTER
we established a formulation that allows us to determine
theoretical upper bounds on the performance of wireless ad hoc networks, in terms
of their capacity regions. As noted, points on the boundary of the capacity region are achieved by optimal time division schedules. Such schedules can only be determined by an algorithm that has global information about the network and its trafc requirements. However, algorithms that require global information are not practical; distributed algorithms are more appealing because they scale better. In principle, such algorithms will be not be able to achieve the whole of the capacity region. Therefore, a very natural question to ask at this point is how the performance of such algorithms compares with the theoretical upper bounds; that is the subject of this chapter. In particular, we focus on (i) the power control algorithm (ii) the routing layer, (iii) the queuing discipline and (iv) the medium access control (MAC) layer. Our work culminates in the introduction of two contention based MAC protocols, the Progressive Back Off Algorithm (PBOA), and the Progressive Ramp Up Algorithm (PRUA). PBOA performs jointly medium access and power control. PRUA sacrices energy efciency in favor of a tighter packing of transmissions. Both protocols are integrated 45
46
CHAPTER 4. CROSS-LAYER DESIGN OF WIRELESS AD HOC NETWORKS
with a queuing discipline that is more relaxed than the First In First Out (FIFO) rule, thus achieving a tighter packing of transmissions. Both protocols are very robust with respect to the choice of the routing protocol. Time is divided in frames, and each frame consists of a contention slot and a data slot. Because time is slotted, our protocols can easily be used for transmission scheduling, by assigning more than one data slot to each contention slot. The rest of this chapter is organized as follows: In Section 4.2 we specify the system model. In Section 4.3 we present an extension of the formulation of the previous chapter that allows us to determine the effects of continuous power control on the capacity. In Section 4.4 we dene a class of routing protocols and evaluate their performance. In Section 4.5 we study queuing disciplines in the context of ad hoc networks. In Section 4.6 we turn our attention to Medium Access Control (MAC) and in particular study the performance of Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA). The PBOA and PRUA protocols are introduced and analyzed in Sections 4.7 and 4.8 respectively. In Section 4.9 we return to the subject of distributed versus centralized protocols for one more time. We conclude in Section 4.10.
4.2 System Model

In this chapter we will use the physical model, as described in Chapter 2, with the receiver function fR () given by Equation (2.4), which is repeated here for convenience: fR (ij ) = R 1{ } . ij T In words, we assume that nodes use for all transmissions a common transmission rate R, and that the receiving node Xj will be able to decode the signal from Xi , with no probability of error, provided the signal to interference and noise ratio ij is greater than or equal to a given threshold T . Otherwise, the signal is lost. Node Xi can transmit with any power between 0 and a maximum Pimax . As noted in Chapter 2, this is a somewhat pessimistic assumption, as typically there is a non-negligible probability that a packet that is below but close to the SINR threshold will be correctly decoded. However, we will use this assumption for clarity of exposition, as the design of our protocols will not be affected and the numerical results will not change
4.2. SYSTEM MODEL
47
100 y (m)
1 2 3 6
50 100
12
5 8
10 11
50
7 9
150
0 0
200 x (m)
250
300
350
400
Figure 4.1: A wireless ad hoc network of 12 nodes. A signal transmitted by a node reaches all others. However, signal strength decays with distance, and only nodes connected by a straight line can communicate directly with each other, in the absence of interference from competing transmissions.
signicantly if a more accurate model is used. The common transmission rate R is a function of T . The specic dependence of R on T is not important, however their relation in practical systems can be closely approximated by a function similar to (2.3): R = W log2 (1 + 1 T ),
where W is the bandwidth of the system and > 1 depends on the particular choice of modulation and coding. As an example, let us consider the wireless ad hoc network of Fig. 4.1. The 12 nodes with any power equal to or less than Pimax = 0.3 W, the thermal noise power is equal are placed in the box {0 x 400 m, 0 y 100 m} randomly. Nodes can transmit
communication rate is R = 1 Mbps. The link gain between nodes Xi and Xj is set to signal decays exponentially with distance. We choose not to model shadowing and fast fading for the sake of simplicity, although the extension would be straightforward. In Fig. 4.1, nodes that can communicate with each other, in the absence of interference
to i W = 1012 W = 90 dBm, the threshold SINR is T = 10 dB and the common
d0 Gij = K( dij ) , with K = 108 , d0 = 1 m, and = 4, so that the power of the transmitted
and with the transmitters using the maximum available power, are connected by straight lines. Nodes that can communicate directly but are not very close are more sensitive to competing transmissions.
48
0.6
0.5 (a) (g) (d) (b) (c) (e) 0.2
0.4 (Mbps)
R 4,8
0.3
0.1
(f)
0 0
0.1
0.2
0.3
R3,12
0.4 (Mbps)
0.5
0.6
Figure 4.2: Capacity region slices of the example ad hoc network along the plane Rij = 0, (i, j) =
(3, 12), (4, 8), i = j: (a) Power control, optimal routing. (b) No power control, optimal routing. (c) Power control, RP-120. (d) No power control, RP-120. (e) RP-120, CSMA/CA. (f) RP-120, PBOA, with m = 15, p = 0.8. (g) RP-120, PRUA, with m = 8, p = 0.3, PT = 1011 W.
4.3 Effects of Power Control on the Capacity

A fundamental restriction of our capacity region formulation is the requirement that the number of transmission schemes that are allowed by the network is nite. Therefore, in general, our formulation cannot handle the case where nodes can transmit with any power level; if this was allowed, the number of different schemes would be innite, because there is an innite number of combinations of transmitter powers, each of which can in principle affect the capacity region. However, in our particular setting, the functional form of the receiver function, as given by (2.4) allows us to bypass the problem. In particular, the receiver function fR () is stepwise with only two steps, on the values 0 and R. Because of this, it is intuitively clear that the innite number of transmission schemes that exist can be mapped to a nite number of rate matrices. In particular, we execute the following algorithm: First, we enumerate all different
4.3. EFFECTS OF POWER CONTROL ON THE CAPACITY
49
combinations of transmitters, receivers, and sources that can exist in the network. The total number of such combinations is large but nite. Secondly, for each of these combinations we determine if there is a power vector that can ensure that all the transmissions can go through simultaneously. For this, the SINR at each receiver must be above the threshold T . This power vector can be calculated by the very well known power control algorithm that is studied in [40, 41, 42]. If all transmissions can go through, the associated basic rate matrix is created. The set of basic rate matrices thus created will give the capacity region in the usual manner. As an example, let us consider the capacity region of the network of Fig. 4.1. Since there are 12 nodes, its dimensionality will be 12(121) = 132, and it clearly cannot be plotted. Therefore, in Fig. 4.2 we plot a slice (line (a)) along which only the rate components R3,12 and R4,8 are non-zero. This slice describes the ability of the network to simultaneously forward packets from node 3 to node 12, and from node 4 to node 8, using multihop routing, time division and power control. The uniform capacity of the network is Cu = 0.93 Mbps. An important advantage of our formulation is that it allows us to change fundamental assumptions about the network, and nd how these changes affect its inherent transport capabilities, in terms of its capacity region. As a simple application of this idea, let us assume that the nodes can no longer control their power, i.e., they either transmit with maximum power or not at all. The formulation of Chapter 3 can be readily applied: In line (b) of Fig. 4.2 we plot the new slice of the capacity region. The uniform capacity now
N becomes Cu P C = 0.86 Mbps. The 7.5% difference is the price we pay for not performing
power control. As discussed in Chapter 3, the gains by allowing power control are not negligible, because there is no continuous rate adaptation. However, for the example network, and also for all other networks generated in a similar manner, these gains appear quite modest. This should be attributed to the multihop nature of the network. Here, it is not only important that a node transmits, but also that it transmits to a distant node, so that the data packet will move signicantly closer to its nal destination. On the other hand, as was also observed in [52], power control mainly facilitates the communication between nodes that are close. In this case the transmitter can use much less power than the maximum and still achieve successful reception, while at the same time creating a negligible amount of interference to other transmitters, thus maximizing the level of spatial reuse.
50
Note that the gains from using power control are modest only when assuming that the network can achieve capacity. On the other hand, if the network is operating under a suite of protocols whose performance is signicantly below the capacity, then in theory enhancing the suite of protocols with power control capabilities could improve the capacity signicantly. In addition, power control also implies energy conservation: When the nodes of the example network achieve the uniform capacity, the energy dissipated for the end-toend transmission of a single packet is 5.95 mJ without power control, and 3.25 mJ with power control. To summarize, employing power control leads to a modest improvement of the network capacity, but to signicant energy savings. This is true for the example network, but also for all other networks with random topologies that we create in a similar manner.
4.4 Effects of the Routing Protocol on Capacity

Over the last few years the development of routing protocols for use in wireless ad hoc networks has attracted signicant research interest and the related body of literature is extensive [5]. Most proposed routing protocols aim to discover a single route with the fewest possible hops between the two nodes that wish to communicate [32, 53]. However, there is no guarantee that this is the best route. Indeed, the best single route will depend on the trafc requirements of the network [54]. Furthermore, typically the optimal routing strategy does not consist of a single route. If two routes are used simultaneously, packets can move to their ultimate destination at a higher rate, provided the two routes are sufciently decoupled [55], as illustrated in Fig. 4.3. As expected, the optimal routing strategy, as determined with the methods of Chapter 3, typically requires that a large number of nodes lying between the source and destination cooperate, the result being that a packet may travel along any of a large number of possible routes. Even in the very small ad hoc network of 5 nodes of Fig. 3.3, the optimal time division schedule of Fig. 3.7 involves multiple routes for the same source-destination pair. It would be interesting to quantify the effect of using simplistic, non-adapting, singlehop routing algorithms on the capacity region of the example network of Fig. 4.1. In particular, let us dene the following three routing protocols: Two nodes will use a single, minimum-hop route to communicate, with the requirement that they only use links whose
4.4. EFFECTS OF THE ROUTING PROTOCOL ON CAPACITY
51
111 000 111 000 111 000 111 000
111 000 111 000 111 000 111 000 11 00 11 00 11 00 11 00
11 00 11 00 11 00 11 00 111 000 111 000 111 000 111 000 111 000 111 000
Figure 4.3: A small network of 6 nodes. Only nodes connected by a dashed line can communicate directly. The rest do not interfere at all. The strategy that maximizes the end-to-end trafc from node A to node B divides the time equally in three transmission schemes, and employs two different routes. SNR (in the absence of competing transmissions) is above a given threshold. For the three routing protocols, the thresholds are set to 10 dB (so that all links may be used for routing), 30 dB, and 120 dB (so that only the strongest links are used), respectively. We call these protocols RP-10, RP-30 and RP-120, respectively. In Fig. 4.2 we plot capacity region slices of the example network, assuming the nodes control their power and operate under RP-120 (line (c)) and assuming no power control and RP-120 (line (d)). In Fig. 4.4 we plot the uniform capacities of the network for all three routing protocols, with (line (c)) and without (line (d)) power control. These plots were determined by using the formulation of Chapter 3, with minor modications that ensure that transmission schemes which are incompatible with the routing protocol used in each case are excluded. As nodes use stronger links, more transmissions may take place simultaneously. On the other hand, these transmissions will be over smaller distances, and more hops will be needed for the nodes to reach their nal destination. For the example network, these effects roughly cancel out, so that the uniform capacity remains roughly the same for all three routing protocols. However, there is a signicant gap between the capacity of the network, and the capacity when a particular routing strategy is imposed. In particular, the uniform capacity is reduced by about 30% for all three routing protocols. As discussed, the reason for this reduction is the requirement that a single route be used per source-destination pair, and that the routing paths do not adapt to the trafc needs of the network Note that, since we are calculating capacities, we are ignoring the overhead associated with discovering new routes and using obsolete routing information. If we include this
52
1 0.9 0.8 Uniform capacity, Cu (Mbps) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 RP10 RP30 Routing Protocol RP120 (e) (f) (d) (g) (c) (a) (b)
Figure 4.4: Uniform capacities versus the routing protocol for the example network: (a) Power
control, optimal routing. (b) No power control, optimal routing. (c) Power control, minimum-hop routing. (d) No power control, minimum-hop routing. (e) Minimum-hop routing, CSMA/CA. (f) Minimum-hop routing, PBOA, m = 15, p = 0.8. (g) Minimum-hop routing, PRUA, m = 8, p = 0.3, PT = 1011 W.
overhead then, depending on the level of node mobility, the actual throughput achieved by the network will further decrease.
4.5 When not to FIFO

Following [56], we refer to the set of rules that governs the order with which a node attempts to forward the packets in its queue as the nodes queuing discipline. The most common queuing discipline is the First In First Out (FIFO) rule, i.e., the node attempts to forward rst the packet that was rst in the queue. Alternative rules have attracted significant research interest in the past in the context of wired networks. Typical applications include prioritizing users [12] and achieving fair bandwidth distribution among competing ows [57]. However, with a few exceptions [58], queuing disciplines other than FIFO have not been studied extensively in the context of wireless ad hoc networks, although they can increase the degree of spatial reuse, as we demonstrate with the following example. Consider the simple network of Fig. 4.5, which consists of the four nodes, A, B, C and
4.6. CARRIER SENSE MULTIPLE ACCESS WITH COLLISION AVOIDANCE
53
First In First Out

C A B D A B D C A B D C
non-FIFO
C A B D A B D C
Figure 4.5: A simple example where employing a non-FIFO queuing discipline is advantageous. D. Only node pairs connected by a straight line can communicate directly with each other, and the rest of the node pairs do not interfere at all. Suppose that only node C is transmitting, to node D, and node B has two packets in its queue, one for A and one for C. If B follows the FIFO discipline and the packet at the head of its queue is intended for C, B will have to wait for C to nish its transmission. However, if there is no FIFO restriction, B can go ahead and transmit the packet intended for A rst without harming Cs transmission (both SINR requirements are satised), and then wait for C to become available (if it has not already become available), thus saving time and improving the channel utilization. In the following we will not assume that nodes follow the FIFO rule. Rather, the queuing discipline will be a subject for optimization.
4.6 Carrier Sense Multiple Access with Collision Avoidance

As shown in Fig. 4.6, under Carrier Sense Multiple Access (CSMA/CA) [17, 18, 19, 20], a node that wants to send a data packet will rst wait for the channel to become available
54
and then transmit a short RTS (Request To Send) packet. The potential receiver, assuming it perceives an available channel, will immediately respond with a CTS (Clear To Send) packet that authorizes the initiating node to transmit, and also informs neighboring hidden nodes (i.e., nodes that are outside the communication range of the transmitter but within the communication range of the receiver) that they will have to remain silent for the duration of the transmission. Nodes that overhear the RTS packet will refrain from transmitting until the CTS packet can be received, and those overhearing the CTS packet will refrain from transmitting until the whole data packet is sent. A node can only send an RTS packet if it perceives an idle channel and has not been silenced by another control packet. A node will only transmit a CTS packet if it has not been silenced by another control packet. The RTS/CTS handshake is coupled with a backoff mechanism that ensures nodes do not transmit immediately after the channel becomes available, but rather wait for a random period, to minimize the probability of a collision. In some incarnations [17, 20], after the data packet is correctly decoded, the receiver sends an ACK (Acknowledgment) packet to verify the reception on the link layer. In these cases, this exchange of packets is referred to as the 4-way handshake. CSMA/CA is a substantial improvement over plain CSMA. However, it does not completely eliminate the hidden node problem, since, for example, it is still possible that CTS packets will collide. In addition, hidden nodes (which we have dened as nodes within the communication range of the receiver but not within the communication range of the transmitter) are not the only nodes that can interfere with, and potentially destroy, the ongoing transmission. In other words, because the interference range is different from the communication range, not all potential interferers will be silenced by the CTS packet. This fact is often suppressed in studies of CSMA/CA that use a more simplistic graph model for the network [59]. Moreover, since the potential transmitter must receive a control packet, it is necessary not only for nodes around the receiver to remain silent, but also for nodes around the transmitter. This articial restriction leads to reduced spatial reuse. For example, in the context of Fig. 4.7(i), if B is already transmitting a data packet to A, C will not be able to initiate a data packet transmission to D, because it will perceive a busy channel. In Fig 4.7(ii), if A is transmitting a packet to B, D will not be able to initiate a data packet transmission to D, because C is silenced by B and can not respond with a CTS packet.
55
1) RTS (Request To Send) B A RTS 3) DATA B A DATA C D C D
2) CTS (Clear To Send) B A CTS C D
4) ACK (Acknowledge) B A ACK C D
Figure 4.6: The 4-way handshake employed in CSMA/CA.
It is also well understood [26] that under heavy trafc, too much time (equivalently bandwidth) is wasted on collision avoidance and resolution, either because nodes back off excessively or because of frequent data and control packet collisions. Finally, there is no known simple way to integrate CSMA/CA with a power control algorithm. As an example, let us study the performance of CSMA/CA, as it is implemented in the ad hoc portion of the IEEE 802.11 protocol (the Distributed Coordination Function) [20], for the example network of Fig. 4.1. We assume Poisson arrivals of packets of xed size (set to 10 Kbits). We refer to the set of all possible combinations of rates that can be achieved by the n(n 1) source-destination pairs of the network when the nodes operate according to a MAC protocol as the capacity region of the network under the given MAC protocol. Contrary to the capacity regions dened previously, its shape will in general depend on the exact model of the packet arrival process. In addition, the effects of the control overhead required by the MAC protocol and the effects of imperfect routing are factored in. Such capacity regions will be obtained here by simulation, as the accurate mathematical model is intractable for all MAC protocols we consider. In line (e) of Fig. 4.2 we plot the slice of
56
B A C D A
B D C
Figure 4.7: CSMA/CA imposes articial restrictions on how close transmissions are placed. (i)
If B is transmitting a packet to A, C will not be able to initiate a data packet transmission to D. (ii) If A is transmitting a packet to B, D will not be able to initiate a data packet transmission to C. However, if time is slotted and all RT S packets are transmitted simultaneously, the articial restriction on how close transmission can be is lifted.
the capacity region that corresponds to the two streams 3 12, 4 8 when nodes operate under CSMA/CA and RP-120, assuming Poisson arrivals of packets with size 10 Kbits. We refer to the maximum aggregate communication rate achieved by the network, under uniform trafc conditions, when it is operating according to a MAC protocol as the uniform capacity of the network under the given MAC protocol. It corresponds to a particular point in the capacity region of the MAC protocol, so all the restrictions under which the capacity region is calculated are carried over in the case of the uniform capacity. In Fig. 4.4 we plot the uniform capacities (line (e)) of the network under CSMA/CA and for the three routing protocols of Section 4.4, assuming Poisson arrivals of packets of length equal to 10 Kbits. In Fig. 4.8 we draw the throughput-delay curve of CSMA/CA under RP-120 and with a uniform trafc load (line (a)). Finally, in Fig. 4.9 we plot the average energy consumed for the transportation of a packet to its nal destination with CSMA/CA, and under the three routing protocols RP-10, RP-30, RP-120 (lines (a), (d), (g) respectively). Again, we assume Poisson arrivals of packets with length equal to 10 Kbits. From Figs. 4.2, 4.4 and 4.8 we see that, because of the problems of CSMA/CA that we have identied, the throughput achieved by CSMA/CA is only 50% of the theoretical maximum, in the case of RP-30 and RP-120, i.e., those protocols that avoid weak links. More dramatically, CSMA/CA performs very poorly when coupled with the protocol that is using all links, RP-10, achieving less than 20% of the theoretical maximum. The reason for this very poor performance is that, under this routing strategy, for the transmissions over
57
1000 900 800 700 Delay, D (msec) 600 500 400 300 200 100 0 0 0.1 0.2 0.3 0.4 Throughput, T (Mbps) (f) 0.5 0.6 (c) (e) (a) (b) (d)
Figure 4.8: Throughput-delay curves for the example network, under RP-120. (a) CSMA/CA. (b)
PBOA with m = 15, p = 0.8. (c) PRUA with m = 8, p = 0.3, PT = 1011 W. (d) Uniform capacity with no power control. (e) Uniform capacity with power control. (f) Lower bound on delay, caused by data packetization and the transmission of each packet over multiple hops.
the weaker links there are many nodes that can potentially destroy the ongoing transmission if they transmit themselves. Many of these nodes lie outside the communication range of the receiver, so they never receive the CTS packet that instructs them to remain silent. Therefore, packets run a very high risk of being destroyed. Consequently, the average energy required for the successful relaying of a packet becomes very large when the network operates close to its capacity (see line (a) of Fig. 4.9). The above result is not an artifact of the particular topology used, but rather highlights a general problem: CSMA/CA (and consequently IEEE 802.11) should not be used for communication over relatively weak links. Rather, it should be coupled with a routing protocol that actively avoids such links. Even if this is the case, using CSMA/CA is still problematic. For example, a strong link may simply not exist. Secondly, even if it does, the routing protocol may fail to nd it because of collisions. We can summarize this discussion in the following statement: CSMA/CA is not robust with respect to the network topology and the choice of routing protocol. A careful look at the RTS/CTS handshake reveals that its effectiveness can be improved signicantly if time is slotted, and nodes start the contention at prearranged times. In the
58
context of Fig. 4.7(i), let us assume that at some point in time nodes B and C have one packet each awaiting transmission, for nodes A and D respectively. If the two nodes transmit their RTS packets simultaneously, they will hear the CTS packets also simultaneously, and then will start transmitting at the same time (the geometry is such that all SINR requirements are satised). The situation is similar if node A has a packet for B and node D has a packet for C. Essentially, slotting time permits a tighter packing of transmissions. In addition, by imposing structure on the contention period, i.e., by requiring that RTS packets only compete with other RTS packets, the efciency of the contention periods should improve, by eliminating partial packet overlaps, and transmissions over weaker links can be better protected. In the next sections we present two protocol that are based on these critical observations.
4.7 Progressive Back Off Algorithm (PBOA)

4.7.1 Frame Format
We start by dening the frame format used by PBOA. As shown in Fig. 4.10, time is partitioned in consecutive frames of xed duration and each frame consists of a contention slot, followed by a data slot. In turn, the contention slot is divided into m pairs of minislots. The rst minislot of each pair is reserved for the transmission of RTS packets, and the second slot for the transmission of CTS packets. The nodes contend for channel access during the minislot period and, depending on the outcome of the contention, some will transmit a data packet during the data slot. As a consequence, all data packets are required to have length equal to that of the data slot. The time that must be devoted to a minislot will depend on the system parameters. As a typical case, let us assume that nodes transmit with a rate of 1 Mbps, the size of data packets is 10 Kbits, and the amount of time required by a transceiver to switch from receive mode to transmit mode is 10 s. Also, let us assume a guard time of 3 s, a maximum channel propagation delay of 2 s, and that the time required for receiver synchronization is another 5 s. As will be shown, each RTS packet will contain the MAC addresses of the source and the destination. Each CTS packet will also contain the source and destination MAC addresses, and in addition a scaling factor used for power control. Assuming that each of
4.7. PROGRESSIVE BACK OFF ALGORITHM (PBOA)
59
18 16
(a) (g) (h) (d) (f) (c) (b)

0.3 0.4 0.5 0.6
Energy per Packet, Ep (mJoule)
14 12 10 8 6 4 2 0 0 0.1 0.2
(i)
(e)
Throughput, T (Mbps)
Figure 4.9: Average energy consumed for the transportation of a packet to its nal destination
as a function of the achieved throughput, for the example network: (a) CSMA/CA, RP-10. (b) PBOA, RP-10. (c) PRUA, RP-10. (d) CSMA/CA, RP-30. (e) PBOA, RP-30. (f) PRUA, RP-30. (g) CSMA/CA, RP-120. (h) PBOA, RP-120. (i) PRUA, RP-120. For PBOA, m = 15, p = 0.8. For PRUA, m = 8, p = 0.3, PT = 1011 W.
the addresses and the scaling factor are 8 bits long, the combined payload of an RTS-CTS pair is 40 bits. This brings the duration of a minislot pair to 80 s, or 0.8% of the duration of the data packet. This calculation is only indicative and the actual overhead may be different if other system requirements are set. For example, if the communication rate is on the order of a few tens of Kbps, as in most deployed (or proposed) Wide Area Networks (WANs), the overhead will be signicantly less. If, on the other hand, the protocol is used in a high speed LAN with a communication rate of 10 Mbps, the duration of each minislot as a percentage of the data packet transmission time will be larger, and care must be taken to minimize it, for example by maintaining a tighter synchronization. From the above discussion, it is clear that nodes must be equipped with clocks that allow global coordination on the level of a few microseconds. For a detailed discussion on minislot overheads and time synchronization we refer the reader to [24] and the references therein.
60
FRAME n1
FRAME n
FRAME n+1
DATA SLOT RTS m CTS m RTS 1 CTS 1 RTS 2 CTS 2
Figure 4.10: Frame format of PBOA and PRUA. Finally, we assume that the mobility of nodes does not cause the channel gains to uctuate signicantly over the duration of a contention slot and the accompanying data slot. Using the parameters above, this duration is on the order of milliseconds, so our assumption is well justied for moderate levels of mobility. For faster changing channels, the suppression of fast fading by the use of techniques such as polarization, frequency, or space diversity can eliminate most fast fading effects.
4.7.2 Protocol Operation

Before going into the details of how PBOA operates, we present the main idea: At the beginning of the contention period, all nodes that have a packet awaiting transmission will start contending for the channel. As the contention period progresses, nodes that are unsuccessful progressively back off, so that others may access the channel. At the same time, nodes that have succeeded in capturing the channel take advantage of the remaining time in the contention period to discover the minimum amount of power that is needed to transmit their data packet. The benets are two-fold: (i) energy is conserved and (ii) other nodes may nd that interference is reduced and so may also succeed in capturing the channel. We now specify the operation of the protocol during the contention period. At any time, a node will belong to one of three groups: the contending nodes, the locked nodes and the silent nodes. Contending nodes have not secured the channel, but continue to contend. Locked nodes have been successful and will transmit a data packet once the contention period is over. Silent nodes have withdrawn from contention. At the beginning of the contention period each node that has a packet awaiting transmission will be placed in the contending group with probability p, where p is a system parameter. All these nodes will
61
initially try to transmit the packet at the head of their queue (to the node that represents the nal destination of the packet or the next hop along its route). The rest of the nodes form the silent group. At the beginning of the i-th RTS minislot: 1. Silent nodes listen to the channel. 2. Contending nodes transmit an RTS packet to their intended receiver with maximum power. 3. Locked nodes transmit an RTS packet to their intended receivers. The power they use for transmitting was specied by the intended receiver at a previous CTS minislot. At the beginning of the i-th CTS minislot: 1. Silent nodes that received an RTS packet from a contending node successfully ( i.e., the SINR i was greater than or equal to the threshold SINR T ) will send, with maximum power, a CTS packet to notify that node of its success. Each of them will specify a factor equal to g = min((1 + ) T , 1) by which the potential transmitter i must lower its power, where > 0 is a small design parameter that compensates for imperfect SINR estimation and changing channel conditions. Therefore, provided the potential transmitter decodes the CTS packet successfully, it will transmit in the current RTS minislot. The success in the next RTS minislot is guaranteed, as no other node will be powering up, transmissions with a SINR above the threshold are always successful, and the parameter adequately compensates for changing channel conditions. 2. Silent nodes that received an RTS packet (intended for them) from a locked node with an SINR i > (1 + )T will transmit a CTS packet notifying the transmitter that it can power down to Pi+1 = Pi (1 + ) T . We require that > . i 3. All other nodes remain silent. At the end of the i-th CTS minislot: 1. Contending nodes that received a CTS packet (intended for them) become locked. next RTS slot with power Pi+1 = gPi Pi where Pi is the power used in the
62
2. Contending nodes that did not receive a CTS packet will remain in the contending group with probability p, or will move in the silent group with probability 1 p. Any node that remains in the contending group tries to pick a new potential receiver. Specically, it moves down its queue of buffered packets awaiting transmission (possibly returning to the head once the tail is reached) and searches for a packet with a different destination from the current. If such a packet does not exist, the node will contend for the same packet in the next minislot pair. At the start of the data slot period, all locked nodes transmit their data packet, with the power level that was last specied to them by their intended receivers. The protocol guarantees that the data packets will be received successfully. Note that the nodes do not use the FIFO queuing discipline. Rather, nodes that are not successful try to pick a new potential receiver, for which the conditions may be more favorable (for example, maybe the rst potential receiver is also contending, but the second is not). Also, a node that receives an RTS packet from a locked node will only respond with a CTS if the SINR is above a threshold (1 + )T , with > . The reason for not always sending a CTS packet is for minimizing congestion during the CTS minislot. Note that we have integrated the RTS/CTS mechanism with the power control algorithm of [40, 41, 42]. We achieved this in a very straightforward manner, without having to resort to using a separate control channel [52] or a centralized scheduling algorithm [60]. Also, we have totally dispensed with the carrier sense mechanism.
4.7.3 Parameter Selection

The number of minislot pairs m and the probability of retransmission p are design variables. Given m and a network topology, there is an optimal value popt for p, but we have experimentally determined that the performance of the protocol is rather insensitive to p (values of p within 0.1 of popt that are not too close to 0 or 1 typically achieve 90% of the maximum performance). As m increases, so does popt , since when there are more minislots it is better for the nodes to be more persistent, so that they back off progressively over the whole minislot period. In Fig. 4.11 we plot the uniform capacity of the example network of Fig. 4.1 under PBOA and the routing protocols RP-10 (line (a)), RP-30 (line (b)), and RP-120 (line (c))
63
0.55 0.5 Uniform Capacity, Cu (Mbps) 0.45 0.4 0.35 0.3 0.25 0.2 0.15 (e)
(a)
(d)
(f)
(b)
(c)
10 15 Minislot pairs, m
20
25
Figure 4.11: Uniform capacity of the example network of Fig. 4.1 versus the number of minislot
pairs m for PBOA and PRUA. For each value of m, the optimal value of p is chosen. For PRUA, PT = 1011 W. (a) PBOA, RP-10. (b) PBOA, RP-30. (c) PBOA, RP-120. (d) PRUA, RP-10. (e) PRUA, RP-30. (f) PRUA, RP-120.
as a function of the number of minislot pairs m. For each value of m, the optimal value of p is chosen. As m increases the performance of the protocol at rst increases, since the nodes are given more time to resolve the congestion efciently. However, after some value of m the incremental improvement in the resolution of conicts is outweighed by the cost of reducing the time devoted to the transmission of data by an extra minislot. The optimal value of m depends on the control overhead incurred by each minislot. For the example network of Fig. 4.1, using the transmitter parameters of Section 4.7.1, and under any trafc condition, a near-optimal choice of parameters is m = 15, p = 0.8. The case m = 1 merits further discussion: from Fig. 4.11 we see that in this case PBOA performs very poorly, and actually worse than CSMA/CA. The explanation is that, with only one minislot pair, nodes have only one opportunity to transmit. If they decide to transmit and there is a collision, the whole data slot will be wasted. In other words, in the case m = 1 PBOA degenerates to slotted Aloha. We also note that this case is the least common denominator of PBOA and CATA [27]: That protocol has a single RTS/CTS minislot pair, but also other minislots for performing reservations and multicasting. When used without reservations or multicasting, CATA is identical to PBOA with m = 1 (so that
64
it too degenerates to slotted Aloha).
4.7.4 Performance of PBOA

Line (f) of Fig. 4.2 is a slice of the capacity region of the network of Fig. 4.1, under PBOA and RP-120. We assume the system parameters of Section 4.7.1 and set m = 15, p = 0.8. In line (f) of Fig. 4.4 we plot the uniform capacity of PBOA for each of the three routing protocols RP-10, RP-30, RP-120. In line (b) of Fig. 4.8 we plot the throughput-delay curve of PBOA under RP-120. This curve is obtained assuming that all of the 132 possible sourcedestination pairs communicate with a common rate. The maximum throughput is now Tmax = 0.403 Mbps, as we see from the gure. All curves are determined by simulation. The performance of PBOA in terms of throughput is superior to that of CSMA/CA for all three routing protocols. This is expected since, as discussed in the previous section, slotting time allows a closer packing of transmissions and a more efcient contention period. PBOA performs much better than CSMA/CA especially in the case of the routing protocol that uses weak links, RP-10. PBOA is therefore robust with respect to the routing protocol. In addition, as Fig. 4.9 shows, PBOA is far superior to CSMA/CA in terms of energy efciency. For example, under RP-10, PBOA uses 2.24 mJ per packet under low trafc conditions and 2.53 mJ under heavy trafc conditions. (Under low trafc conditions there are fewer transmissions going on and lower power levels can be used.) On the other hand CSMA/CA needs 6.35 mJ per packet under low trafc and 16.23 mJ under heavy trafc. (Under heavy trafc data packet collisions are frequent, so data packets need to be retransmitted, and more energy is lost.) However, the performance of PBOA is still considerably below the theoretical maximum. There is no single source for this gap, however it is intuitively clear that one of the reasons is the following: During the contention period there are nodes that contend for the channel in vain. For example, a node may be sending an RTS to a node that has already sent an RTS and received a CTS in reply. The rst node would have no possibility of success, but it would still be transmitting, thus hurting other nodes that still have a chance. If the nodes could make more informed decisions, the contention for the channel would be more efcient. This idea is the motivation for the Progressive Ramp Up Algorithm (PRUA).
4.8. PROGRESSIVE RAMP UP ALGORITHM (PRUA)
65
4.8 Progressive Ramp Up Algorithm (PRUA)

4.8.1 Protocol Operation
PRUA uses the same frame format as PBOA (see Fig. 4.10). However, the rules of contention are very different. Before we specify its operation in detail, we briey discuss the main idea: During the contention period, a node that has a packet awaiting transmission monitors the channel. Whenever it senses favorable conditions for one of its potential destinations, it transmits an RTS with some small probability at the beginning of the next RTS minislot. If it is successful, it will persist sending RTS packets (so that it makes it known to competitors that it was successful). Otherwise, it backs off and may try again at a later RTS minislot. Nodes do not control their transmitter power. We now specify in detail the operation of the protocol. During the i-th RTS minislot, a node A will transmit an RTS packet to a destination B if, in the previous minislot pair, it sent an RTS packet and received a CTS packet from B in reply. Alternatively, node A will send an RTS packet if all of the following conditions are met: 1. A has not sent a CTS packet in the previous minislot (i.e., it is not awaiting a packet in the coming data slot). 2. During the previous CTS slot, the combined power of all signals received at A did not exceed a predetermined threshold PT (Otherwise, A may interfere with other transmissions). 3. If A has not decoded an RTS packet in the previous RTS minislot, then A must have a non-empty queue. It will contend for the packet at the head of the queue. If A has correctly decoded an RTS packet in the previous RTS minislot from some node C, then there must be a packet in the queue, intended for some node B, such that B is able to decode the packet from A in the presence of interference from node C (specically, B must be different from C). A will choose to contend for the packet that is closest to the queue head and satises this condition. 4. A must perform a biased coin toss with probability of success p, and succeed.
66
During the i-th CTS minislot, all nodes that successfully decoded in the i-th RTS minislot an RTS packet intended for them reply with a CTS packet. In order to transmit a data packet, a node must receive a CTS packet in the last CTS minislot. Nodes do not necessarily transmit the packet at the head of their queue, but rather the packet closest to the head of the queue for which the conditions appear more favorable. As in CSMA, nodes employ carrier sense. However, nodes now sense the power transmitted from other receivers, as in [61, 52], and not the power of ongoing transmissions, as in CSMA, so the decisions they make are based on much more relevant information.
4.8.2 Parameter Selection

The number of minislots m, the probability p, and the carrier sense threshold PT are design variables. If p is set too high, then collisions of RTS packets will be frequent. If, on the other hand, p is set to a very small value, too many minislot pairs will be wasted. Given m and a network topology, there is an optimal choice of p, popt , but we have determined experimentally that the performance is rather insensitive to the actual choice of p (values of p within 0.1 of popt typically achieve 95% of the maximum performance). As m increases, popt decreases, since now the nodes can afford to be less persistent, thus reducing the probability of RTS packet collisions while having ample opportunity to capture the channel. In Fig. 4.11 we plot the uniform capacity of the example network of Fig. 4.1 under PRUA and the routing protocols RP-10 (line (d)), RP-30 (line (e)), RP-120 (line (f)), versus the number of minislot pairs m. For each value of m, the optimal value of p is chosen. As in PBOA, as m increases the performance of the protocol at rst increases, since the nodes are given more time to resolve congestion efciently. However, after some value of m the incremental improvement in the resolution of conicts is outweighed by the cost of the time that must be set aside for an extra minislot. For the example network of Fig. 4.1, using the transmitter parameters of Section 4.7.1, and under any trafc condition, a near-optimal choice of parameters is m = 8, p = 0.3. Because of its design, PRUA typically needs fewer minislot pairs than PBOA to resolve contentions. Therefore, PRUA has an advantage over PBOA when the system parameters are such that the minislot overhead is signicant.
opt The optimal value for PT , PT , depends on the topology and the routing protocol, but opt is relatively insensitive to both. For the example network, PT 1011 W = 80 dBm.
4.8. PROGRESSIVE RAMP UP ALGORITHM (PRUA)
67
Figure 4.12: A wireless ad hoc network for which a graph channel model is assumed: Two nodes
connected by a straight line can communicate directly in the absence of interference. All other node pairs do not interfere with each others transmissions. Dashed arrows represent a transmission scheme that consists of four transmissions that can coexist.
4.8.3 Performance of PRUA

Line (g) of Fig. 4.2 is a slice of the capacity region of the example network under PRUA and RP-120. We assume the system parameters of Section 4.7.1 and set m = 8, p = 0.3, PT = 1011 W. In line (g) of Fig. 4.4 we plot the uniform capacities of the network for the three routing protocols RP-10, RP-30, RP-120, and under the same parameters. In line (c) of Fig. 4.8 we plot the throughput-delay curve of PRUA, assuming the same parameters, RP-120, and uniform trafc conditions. Finally, in Fig. 4.9 we show the average energy consumed per packet under PRUA and for all three routing protocols RP-10, RP-30, RP120 (lines (c), (f), (i)). All curves are determined by simulation. Since transmitters always transmit with their maximum power, and transmissions are always successful, the average energy does not depend on the throughput, and is simply equal to the energy consumed per transmission multiplied by the average number of hops. The results show that PRUA is superior to PBOA in terms of throughput performance; allowing nodes to make more informed decisions helps marginally more than power control. On the other hand, PRUA is more wasteful in power: With RP-120, PRUA requires 10.44 mJ per packet under any trafc intensity, whereas we have seen that PBOA requires 0.94 mJ per packet under low trafc, and 1.18 mJ under heavy trafc. On the other hand, because with PRUA there are no data packet collisions, PRUA does not require as much power as CSMA/CA does, especially under RP-10 (compare lines (a) and (c) of Fig. 4.9). Although PRUA outperforms PBOA and CSMA/CA in terms of achievable throughput, the above gures show that there is clearly a large gap between the performance of PRUA and the theoretical maximum, in terms, for example, of uniform capacities. This is rather
68
disappointing, given the relative advantages that PRUA has over CSMA/CA. The existence of this gap, as will be discussed in detail in the following Section, is in large part due to the distributed nature of our protocols. Specically, in order to achieve capacity, it is necessary that arbitrarily distant nodes coordinate their transmissions. This behavior is not feasible by PBOA and PRUA or any other distributed MAC protocol.
4.9 Fundamental Limits of Distributed Protocols

In the previous sections, we rst studied CSMA/CA and highlighted some of its shortcomings. Motivated by this discussion, we developed two protocols whose performance is clearly superior to that of CSMA/CA. But as we have seen, both protocols fail to achieve a performance close to the theoretical maximum. For example, under uniform trafc conditions and RP-120, PRUA achieves only 71% of the uniform capacity. This gap is largely due to the distributed nature of our protocols. Indeed, as was shown in Chapter 3, to achieve the capacity, the nodes need to use an optimal time division schedule that can only be calculated if we have global information about the network topology and trafc requirements. In addition, under this schedule it is imperative that arbitrarily distant nodes coordinate their transmissions. To illustrate this point, let us examine the network of Fig. 4.12, for which we assume a graph channel model: Node pairs that can communicate directly are connected by a straight line, and the rest of the pairs do not interfere at all. Any two packets that overlap at the receiver, even partially, are both destroyed. Because the topology is highly symmetric, it is very easy to guess the uniform capacity of this network, and how it is achieved: Under uniform trafc conditions, the optimal strategy for the network would be to divide its time equally between the following transmissions schemes: A: 1 2, 4 3, 5 6, 8 7 B: 2 1, 3 4, 6 5, 7 8 C: 2 3, 5 4, 6 7, 1 8 D: 3 2, 4 5, 7 6, 8 1
4.9. FUNDAMENTAL LIMITS OF DISTRIBUTED PROTOCOLS
69
In Fig. 4.12, we have drawn the rst of these schemes using dashed arrows. At any time, there will be s = 4 simultaneous transmissions. The average number of hops that a packet needs to arrive at its nal destination can be easily found to be h = communication rate used by the transmitters is r = 1 network is Cu =
s rh packet , sec 16 . 7
Assuming that the
the uniform capacity of the
7 packets . 4 sec
Let us assume that the network operates under CSMA/CA, and there is an ongoing transmission from 2 to 3. Clearly, 1 cannot be receiving a packet at the same time. In addition, it cannot be transmitting (necessarily to node 8) because that would imply that nodes 1 and 2 transmit their RTS packets simultaneously. This kind of behavior is very unlikely in CSMA/CA for any reasonable value for the propagation delay (impossible if we ignore propagation delays). So 1 does not participate in any transmission. By a similar argument, 4 does not participate in any transmission also. On the other hand, the rest of the nodes are perfectly free to initiate a single other transmission. Therefore, we can see that under CSMA/CA there can be no more than s = 2 simultaneous transmissions, so the uniform capacity of CSMA/CA cannot be more than Cu = r s = h
7 packets , 8 sec
or 50%
of the theoretical maximum. In reality, because of control overhead, backing off, and data packet collisions, the uniform capacity of CSMA/CA, estimated by simulation, is only Cu
CSMA/CA
= 0.635
packets . slot
Let us consider how PRUA or PBOA would perform. In theory, both protocols can operate under any of the 4 schemes A,B,C,D that are required to achieve capacity. However, for this to happen it is necessary that distant nodes coordinate: Whenever 1 captures the channel with the intention of sending a packet to 2, 5 must also capture the channel and send a packet to 6. This kind of behavior cannot be forced on the nodes by PBOA or PRUA, because under these protocols, nodes make decisions on whether they access the channel or not, based on local information. In other words, our protocols are distributed. The result is that most of the time there will be only two or three simultaneous transmissions. Indeed,
PBOA by simulation we nd that the uniform capacity under PBOA is Cu = 1.05 P the uniform capacity under PRUA is Cu RU A = 1.145 packets . sec packets sec
and
To summarize, for an ad hoc network to operate at capacity, arbitrarily distant nodes will need to coordinate their transmissions. By denition, this behavior is not possible when nodes use distributed MAC protocols, such as CSMA/CA, PBOA, PRUA and practically all protocols proposed in the literature. In the language of [22], such protocols do not discover
70
a maximum slot assignment (although it is intuitively clear that, in the limit of a large number of minislots, PRUA in particular does converge to a maximal slot assignment [22]). Therefore, for all such protocols there will be a gap between the capacity and what the protocols achieve. This gap will depend on the topology of the network, the properties of the transceivers, the trafc requirements of the nodes and the routing protocol. Note that this gap is fundamentally different from the gap that results from imperfect routing, that we have identied in Section 4.4.
4.10 Conclusions
In this chapter we study the performance of wireless ad hoc networks in a multihop trafc environment, with an emphasis on the medium access control sub-layer. Our approach is to evaluate the effects of various design choices by using theoretical upper bounds on the performance that are given in terms of capacity regions. Applying this framework we arrive at a number of important observations: 1. Power control can be very helpful in terms of energy efciency, but its gains are limited in multihop environments. 2. The First In First Out (FIFO) queuing discipline is suboptimal for use in wireless ad hoc networks. A more relaxed rule can lead to a tighter packing of transmissions. 3. The choice of routing protocol can signicantly reduce the capacity of the network. Specically, any routing strategy that assigns a single route to a node pair typically suffers a penalty in its performance. 4. CSMA/CA is not well-suited for communication over weak links. Therefore, it should be coupled with routing protocols that avoid, if possible, such links. 5. Medium access protocols that are distributed inherently carry a penalty in their performance since, to achieve capacity, it is necessary that nodes that are separated by an arbitrarily large distance coordinate their transmissions. By denition, this is impossible in a distributed algorithm.
4.10. CONCLUSIONS
71
We also introduce two contention based protocols that perform medium access control in wireless ad hoc networks. The Progressive Back Off Algorithm (PBOA) performs medium access jointly with power control. The Progressive Ramp Up Algorithm (PRUA) sacrices energy efciency in favor of higher throughput. With both protocols time is divided in frames, each frame consisting of a number of RTS-CTS minislot pairs used for contention and a data slot, used for the transmission of data. Both protocols use queuing disciplines that are more relaxed than the First In First Out (FIFO) rule and promote dense spatial reuse. They are totally distributed and their control overhead does not scale with the size and node density of the network. Both protocols are superior to CSMA/CA (and protocols that are based on CSMA/CA, such as IEEE 802.11) both in terms of achievable throughput and energy efciency. In particular, they are robust with respect to the choice of routing protocol and network topology. Although we only present results for an example topology, we have also studied a large number of other topologies and have arrived at very similar results in all cases. Our contention based protocols may be trivially converted to transmission scheduling protocols, by using the frame format of [24]: k contention slots, each consisting of m minislot pairs, are followed by l frames, each consisting of k data slots. Each contention slot is used to reserve l data slots, one in each of the l frames. Alternatively, PRUA in particular can be trivially modied to support reservations, with a mechanism similar to that of CATA [27]. Specically, during the rst RTS minislot of the n-th frame only nodes that have successfully reserved the (n k)-th frame are allowed to transmit an RTS. Therefore, once a node contends successfully for the channel, it has guaranteed access every k frames. If a node does not make use of its priority (by remaining silent during the rst RTS minislot), the rest of the nodes will contend as usual, starting from the second RTS minislot. Finally, PBOA and PRUA are not the only protocols that could follow the frame format of Fig. 4.10. Within this framework, other designs are certainly possible.
Chapter 5 Large Wireless Networks

5.1 Introduction
HE CAPACITY REGIONS
introduced in Chapter 3 constitute a precise description of
the trafc-carrying capabilities of wireless ad hoc networks. However, as was noted
in that chapter, the formulation suffers from a computational complexity which explodes with the number of nodes. If we are to study arbitrarily large networks, it is imperative that an alternative approach be adopted that is perhaps less precise, but captures the main trends and tradeoffs of large networks. One such approach in the one introduced in a landmark paper by Gupta and Kumar [36]. There, the authors investigate the asymptotic behavior of the capacity of a class of two-dimensional random networks as the number of nodes n approaches innity, under a uniform trafc assumption. Here, nodes are assumed immobile. The authors present a scheme that achieves with high probability, i.e., with probability approaching 1 as n approaches innity, a per-node communication rate equal to (n log n) 2 , up to a multiplicative constant, from each node to its randomly chosen destination. This result hinges on the fact that, if every node transmits to a close neighbor, around n/ log n simultaneous transmissions are possible. On the other hand, for a packet to reach its destination node, which is chosen in random, around (n/ log n) 2 such transmissions will be needed. Consequently, the aggregate trafc that will be carried by the network will grow as (n/ log n)/(n/ log n) 2 = (n/ log n) 2 , so that the end-to-end rate of communication to 72
1 1 1 1
5.2. NETWORK MODEL AND OVERVIEW OF RESULTS
73
each destination will be around (n log n) 2 . The authors also show that, with high probability, the n nodes cannot send data to their destinations with a per-node communication rate greater than (n log n) 2 , up to a (different) constant. The last result is disheartening since it suggests that, as the number of nodes n goes to innity, the per-node communication rate will necessarily go to zero. However, node mobility can lead to dramatic improvements: In [38] the authors concentrate on a wireless network with mobile nodes. They show that, with high probability, in the absence of any constraint on the delay in the delivery of a packet, each node is guaranteed a xed rate of communication to its destination, irrespective of the number of users in the channel. The downside of this result is that, as the number of nodes increases, so will the expected delay experienced by the packets. In this paper, we continue the investigation along the lines of [36] and [38]. In Section 5.2 we introduce our model for wireless ad hoc networks, and present all our results. The next sections are devoted to the derivation of these results. In Section 5.3 we start by concentrating on networks with immobile nodes and with no fading. In Section 5.4 we extend our formulation to include the effects of fading. In Section 5.5 we derive a per-node achievable rate of communication when nodes are mobile and there is a xed bound on the delay, dmax , which does not depend on the size of the network n. In Section 5.6 we derive an achievable per-node communication rate when nodes are mobile, but assuming that packet delays dmax (n) that increase with n are tolerated. We conclude in Section 5.7. Throughout the text, terms being dened are set in boldface.
1
5.2 Network Model and Overview of Results

5.2.1 Immobile nodes with no fading
1 1 pendently, in the two-dimensional area {(x, y) : 2 |x|, |y| 2 }. For now, we take the
Let N be a collection of n nodes X1 , X2 , . . . , Xn , placed randomly, uniformly and inde-
nodes to be immobile.
Among the n nodes there are n trafc streams, each described in terms of its source node and destination node, and created according to one of the following three rules: Trafc Model 1 (TM1): Each node is the source of a single stream, and the destination of a
74
CHAPTER 5. LARGE WIRELESS NETWORKS
single stream. A node cannot be the source and destination of the same stream. Apart from this restriction, all other combinations of sources and destinations are equally probable. Trafc Model 2 (TM2): Each node is the source of a single stream. The destination of the stream can then be any of the rest of the nodes. Therefore, a node is the destination of a random number of streams. (This trafc model bears the greatest resemblance to the one used in [36].) Trafc Model 3 (TM3): The source of each of the n streams is chosen randomly, uniformly and independently of the other sources. The destination of that stream is then chosen randomly and uniformly, among the rest of the streams. Therefore, a node is the source, and destination, of a random number of streams. Nodes communicate over a wireless channel of bandwidth W . Half-duplex transmission is assumed, i.e., nodes cannot transmit and receive simultaneously. Each node can transmit with any power Pi , provided it is less than the maximum P0 . When node Xi transmits with power Pi , node Xj receives the transmitted signal with power Gij Pi , where Gij = Kd . K is a constant, the same for all nodes, dij is the distance between nodes Xi ij and Xj , and > 1 is the decay exponent. For now, we do not incorporate fading in our model. transmitting with power Pt . Let us assume that node Xj , j T is receiving information be Gij Pi +
kT , k=i
Let {Xt : t T } be the set of transmitting nodes at a given time, each node Xt
from Xi , i T . Then the signal to interference and noise ratio (SINR) at node Xj will Cj = Gkj Pk ,
where is the thermal noise power at the receiver, which is assumed the same for all nodes. We assume that the transmission of the packet will be successful if and only if the transmission rate used, Rj , satises the inequality Rj fR (Cj ), where fR () is a function that reects the quality of the receiver and the performance metric. In particular, we use the formula fR (Cj ) = W log2 (1 + 1 Cj ) (5.1)
where log2 (x) denotes the logarithm of x, of base 2. With = 1, the receiver achieves Shannons capacity. With > 1, (5.1) approximates the maximum data rate that meets
75
a given BER requirement under a specic modulation and coding scheme such as coded MQAM [48]. The precise value of depends on the choice of coding and modulation parameters, and the BER requirement. Each node is equipped with a buffer of innite length, for storing packets intended for other nodes. We assume omniscient nodes with perfect knowledge of the topology. This is a realistic assumption, as we take them to be immobile (for now). The transmission strategy for all nodes is agreed to in advance. Thus, no overhead is needed to determine the topology or the transmission strategy. Under the above set of assumptions, the following can be proved:
Theorem 1 (Immobile nodes with no fading) For all three trafc models TM1, TM2, TM3, each stream can achieve, with high probability, i.e., with probability approaching 1 as the number of nodes approaches innity, a rate of communication equal to: 1 [log (1 + 1 36 10 1 )(n log n) 2 2 2 35
32
if > 2, if = 2, if 1 < < 2. (5.2)
W 3 1 [100]1n 2 (log n) 2 1 (n) = 144 10 (2)10 1 3 1 [ ]n 2 (log n) 2
Note that, for the case > 2, a similar result was shown, under a very similar setting, in [36]. In Section 5.3, we present a proof that is signicantly shorter and more intuitive. In addition, the methodology used can be easily extended to handle fading and mobility.
5.2.2 Immobile nodes with fading

We now assume that when node Xi transmits with power Pi , node Xj receives the transmitted signal with power Gij Pi , where Gij = Kd fij . The extra factor fij is the fading ij coefcient, a non-negative random variable that models fading, and does not change with time. We assume that the expectation E[fij ] = 1, and that fij = fji . We take the distinct
n(n1) 2
fading coefcients to be independent and identically distributed (iid). We also
76
assume that their complementary cumulative distribution function F c (x) has a thin, exponentially decaying tail. Formally: F c (x) P [fij > x] exp[qx] x > x1 , (5.3)
for some real and positive parameters q, x1 . In addition, we make the very mild assumption that there is a median value fm > 0 such that: 1 P [fij fm ] . 2 (5.4)
Both of these assumptions are satised by most distributions used to model fading. In particular, we show in the appendix that they are satised for the Nakagami, Ricean and Rayleigh distributions. In this setting, we devise a scheme that guarantees a per-stream communication rate that, compared to the previous case, is not reduced by more than a logarithmic factor. In particular, the following holds (the proof appears in Section 5.4):
Theorem 2 (Immobile nodes with fading) For all three trafc models TM1, TM2, TM3, each stream can achieve, with high probability, i.e., with probability approaching unity as the number of nodes approaches innity, a rate of communication equal to: a+3 [ 10 2 36 ]n 1 (log n) 3 2 2 1296 35 5 W qfm 10 2 1 5 2 (n) = [ 1296 ]n 2 (log n) 2 1 (2)10 2 3 +1 [ ]n 2 (log n) 2
32 1296
if > 2, if = 2, if 1 < < 2. (5.5)
5.2.3 Mobile nodes with fading and delay constraints

Let us assume that nodes are no longer immobile, but rather move inside the square region being uniformly distributed at any time, and according to a stationary and ergodic process that satises the following assumptions:
1 {(x, y) : 1 |x|, |y| 2 }, independently of each other, with the position of each node 2
77
(i) There is a duration of time s such that within any interval of this duration all nodes remain immobile, so that the power gains of all links remain constant, even in the presence of fading. (ii) There is a duration of time S, such that after the passing of a time interval of length S, the positions of all nodes become perfectly reshufed. In other words, after the lapse of an interval equal to S, the position of each node becomes randomly, uniformly and independently redistributed, irrespective of the initial node placement. The rst assumption is a reasonable approximation of any realistic model of mobility, provided we take s to be sufciently small. The second assumption holds for all stochastic mobility models that have nite memory. However it is also a reasonable approximation for trafc models in which the distribution of the position of a node that was at a particular location at time t = 0 approaches the uniform distribution as t . For convenience, we take S to be a large integer multiple of s, i.e., S = Ns. Let the rest of the assumptions of Sections 5.2.1 and 5.2.2 continue to hold, with the exception that nodes no longer have global topology information. Also, the whole communication period will last for a time interval of length equal to 2nD S, where D is an arbitrary integer, greater than unity. Before formally presenting the two capacity results under node mobility, it is worthwhile discussing the basic idea behind them, which in fact is very simple: It was shown in [38] that a number of simultaneous transmissions on the order of n is possible, with a xed transmission rate that does not decrease with n, if all nodes transmit to their nearest neighbors. In addition, because of the mobility of the nodes, two transmissions are enough for a packet: once to a relay (that happens to be the nearest neighbor of the source) and once more to the nal destination, whenever this happens to be the nearest neighbor of the relay. (One transmission would also be enough, but the number of source-destination pairs that are also nearest neighbors is not on the order of n.) Therefore, an aggregate throughput on the order of n is possible, so that the per-node throughput is xed, and independent of the number of nodes n. However, under this scheme each packet will have to remain in its relaying node for a time increasing proportionally with n, since the chance of the relay
1 being the closest neighbor with the destination is n , and that is excluding the queuing delay.
(Of course, to make this argument formal we need to specify the mobility model in detail.)
78
Delay exponent, d
0 -1
A
Throughput exponent, t 0
Figure 5.1: A fundamental tradeoff between the packet delays and the per-node achievable throughput.
The motivation for our schemes is simple: If more nodes could receive the packet when it is transmitted by its source (and act as potential relays), rather than just one, maybe the packet would not have to wait that much time. But for more nodes to receive the packet, it is necessary that fewer nodes transmit, so that the transmitted signals experience less interference and reach further. However, this will reduce the throughput. In other words, it is clear that, in principle, there can be a trade-off between delay and throughput. Indeed, we prove the following two theorems (whose proofs appear in Sections 5.5 and 5.6 respectively): Theorem 3 (Mobile nodes with a xed upper bound on packet delay) Under the trafc model TM1, node mobility, fading, and an upper bound on the acceptable packet delay equal to dmax = (N + 1)s, each node can communicate, with high probability, with its destination with a rate equal to: 3 (n) = W qfm 72(D + 3)(10D + 20) 2 [ 36 10 a 1 ]n 1 (log n) 3 2 2 2 35 1 5 1 [ 100 ]n 2 (log n) 2 (2)10 1 (10D+20)1 3 1+ 2 2 [ ]n 2 (log n) 2
32
1
if > 2, if = 2, (5.6) if 1 < < 2.
79
This result reects a stringent requirement on the packet delay. If larger delays are tolerated, the per-node communication rate can improve, as the following theorem shows:
Theorem 4 (Mobile nodes with a polynomial upper bound on delay) Under the trafc model TM1, node mobility, fading, and an upper bound on the acceptable packet delay equal to dmax = (N + 1)s, each node can communicate with its destination, with high probability, with a rate equal to: 4 (n) = W fm q 3456(D + 3)(11D) 2 [ 36 10 a 1 ]n d1 (log n) 5 2 2 2 35 1d 7 1 [ 100 ]n 2 (log n) 2 (2)10 1 (11D)1 +d3 3+ 2 2 [ ]n 2 (log n) 2 32
3
if > 2, if = 2, (5.7)
if 1 < < 2.
Theorem 4 suggests a fundamental tradeoff between the acceptable packet delay and the achievable per-node throughput: If we are willing to tolerate larger delays, our throughput will increase. To better illustrate this trade-off, let us say that a particular communication scheme has a throughput exponent t if, with high probability, the per-node throughput is equal to (n) = Rnt , with R being a constant, modulo arbitrary powers of logarithms. Similarly, let us say that a communication scheme will have a delay exponent d if, with high probability, the maximum packet delay is bounded by dmax (n) < Rnd , where R is a constant, modulo powers of logarithms. Using this language, Theorem 4 implies that there are communication schemes whose throughput exponent and delay exponent pair lies on the open line interval AB of Fig. 5.1. Theorem 3 shows that point A is also achievable. Point B can be achieved by a scheme very similar to that of [38].
80
5.3 Immobile Nodes with no Fading

In this section we prove Theorem 1, which gives a lower bound on the communication rate available to each stream, assuming that nodes are immobile and that there is no fading. The proof is constructive, i.e., we will determine the lower bound by creating a communications scheme, and showing that the scheme will work with high probability.
5.3.1 The Cell Lattice

n Let x be the greatest odd multiple of 3 less than or equal to x. Let ( k1 log n ) 2 =
1
2r + 1, where k1 is a constant to be specied later. We divide the space occupied by the network into a regular lattice of g(n) = (2r+1)2 cells, as shown in Fig. 5.2. The boundaries of the cells are formed by the lines 1 i x= + , i = 0, . . . , 2r + 1, 2 2r + 1 1 j y= + , j = 0, . . . , 2r + 1. 2 2r + 1
Points on the boundary lines are assigned arbitrarily to one of the cells forming the boundary. We denote the cells by c1 , c2 , . . . , cg(n) . In addition, each cell is identied by its node to be the coordinates of the cell in which the node lies (so two nodes may have the same coordinates). We call two cells neighbors if they share a common boundary edge, so that each cell has at most four neighbors. We call two nodes neighbors, if they lie in the same or neighboring cells. will only hold asymptotically, i.e., for sufciently large n. For example, f (n) <a g(n) means that there is a n0 such that f (n) < g(n) for all n > n0 . Using this notation, we have that n n <a g(n) . 2k1 log n k1 log n (5.8) We will use the symbols <a , >a , a , a to denote that the corresponding inequality coordinates (v1 , v2 ) in the lattice, where r v1 , v2 r. We dene the coordinates of a
5.3. IMMOBILE NODES WITH NO FADING
81
y
(-r,r) (r,r)
(2r+1) cells
(-r,-r)
(2r+1) cells
(r,-r)
Figure 5.2: Partition of the two-dimensional wireless ad hoc network in a regular lattice of cells.
n Indeed, g(n) = ( k1 log n ) 2 2 n , 2k1 log n
1 1 1 n n n . Also, let x(n) = ( k1 log n ) 2 ( k1 log n ) 2 , and k1 log n 1 1 n n n = [( k1 log n ) 2 x(n)]2 = [1 x(n)( k1 log n ) 2 ]2 k1 log n >a
note that 0 x(n) < 6. So g(n) and (5.8) is proved.
Let mi be the number of nodes in cell ci . Because the nodes are placed randomly, independently and uniformly, mi is binomially distributed, with the n Bernoulli experiments and the probability of each experiment equal to E(mi ) =
n g(n) 1 . g(n)
By (5.8), the expected value
grows with n roughly as k1 log n. However, we need uniform bounds on the
actual number of nodes of any cell. In fact, we have the following:
Lemma 1 (Number of nodes in cells) 2n1 8 k1 log n mi 4k1 log n i] >a 1 . P[ 2 k1 log n
k1
Proof: We make use of Chernoffs bounds [62]: Let X be a binomially distributed random variable, with parameters k (the number of Bernoulli experiments) and p (the probability
82
of success of each Bernoulli experiment). Then, for any (0, 1], P [X < (1 )kp] < exp(kp Also, for any > 0, P [X > (1 + )kp] < exp[kpf ()], (5.10) 2 ). 2 (5.9)
that f () > 0 for > 0.
where f () = (1 + ) log(1 + ) . By calculating the derivative, we immediately have Let us concentrate on a particular cell ci . Then: P [mi < < = Similarly, P [mi > 4k1 log n] a P [mi > 2 < = n ] (Using (5.8)) g(n) nf (1) ] (Using (5.10) with = 1) exp[ g(n) exp[k1 f (1) log n] (Using (5.8)) nk1 f (1) n 8 .
n i=1
k1
k1 log n ] 2 n ] (Using (5.8)) P [mi < 2g(n) n 1 exp[ ] (Using (5.9) with = ) 8g(n) 2 k1 exp[ log n] (Using (5.8)) 8 k1 8 n .
(5.11)
(5.12) P (Ei ), typically referred to as
the union bound. Combining (5.11) and (5.12) with the union bound, and using (5.8), we arrive at the result.
We also note the basic inequality P (n Ei ) i=1
83
y
r
A1 S1 D2
A2
A3
v2
A8
ci
A4 x
D1 A7
-r
S2 A6
v1
-r
A5
r
Figure 5.3: Examples of the routing algorithm. Note that this lemma is closely related to the well-known Coupon Collectors Problem [63]. To see the connection, let n be the number of coupons the collector purchases and g(n) be the number of distinct types of coupons. The lemma gives uniform upper and lower bounds on the number of coupons collected from each type.
5.3.2 The routing rules

Lemma 1 shows that, with high probability, all cells are guaranteed to have at least one node (in fact many more), provided k1 is large enough. We are therefore justied to dene the following rules that govern the routing of data to their nal destination: (i) Direct transmission of data is allowed only between nodes that lie in the same cell, or in neighboring cells (as specied, cells are considered neighbors if they share a common edge, so that each cell has at most four neighbors). (ii) Nodes that do not lie in the same or neighboring cells communicate by using nodes in intermediate cells as relays. The message is rst transmitted along cells whose x-coordinate
84
is the same as the x-coordinate of the source, until it arrives at a cell whose y-coordinate is the same as the y-coordinate of the destination. Then, the message is transmitted along cells whose y-coordinate is the same as the y-coordinate of the destination. In each of the intermediate cells, one of the nodes in the cell is arbitrarily selected to act as the relay of the packet. Examples of this routing rule appear in Fig. 5.3.
Since the streams that are routed through a particular cell create load for the nodes in that cell, it is important to have an estimation of their number. We have the following:
Lemma 2 (Expected number of streams arriving at a cell) Let (v1 , v2 ) be the coordinates of cell ci . Let si be the number of streams arriving at (and possibly going through) cell ci . For all three trafc models TM1, TM2, TM3, the expected value of si is:
2 2 1 + 2(2r + 1)(2r 2 v1 v2 + 2r) E[si ] = n . (2r + 1)4
Proof:
We break the rest of the cells in the 8 areas A1 , . . . , A8 , as shown in Fig. 5.3.
Denoting by N(Ai ) the number of nodes in Ai , we have: N(A2 ) = r v2 , N(A4 ) = r v1 , N(A6 ) = r + v2 , N(A8 ) = r + v1 , N(A1 ) = N(A2 )N(A8 ), N(A3 ) = N(A2 )N(A4 ), N(A5 ) = N(A4 )N(A6 ), N(A7 ) = N(A6 )N(A8 ). According to the routing rules, a packet will have to be received by a node lying in ci if and only if its source and destination nodes lie in one of the following pairs of areas:
85
Source A1 A2 A3 A4 A5 A6 A7 A8 ci
Destination A4 , ci A4 , A5 , A6 , A7 , A8 , ci A8 , ci A8 , ci A8 , ci A1 , A2 , A3 , A4 , A8 , ci A4 , ci A4 , ci ci
Since we have assumed that the nodes are uniformly distributed, calculating the probabilities of the 8 cases is straightforward. Adding them up, and multiplying by n to account for the n streams, we arrive at the result. Note that, although the algorithm makes some effort to avoid routing through the center of the network, some cells will inadvertently be required to receive more trafc than others. For example, the corner cell with coordinates (r, r) will be required to receive packets from n 1+4r+8r streams on the average. On the other hand, the node at the center will be (2r+1)4 required to receive packets from n 8r
1 2 3 +12r 2 +4r+1 2
(2r+1)4
streams on the average. Since (2r + 1) =

1
n ( k1 log n ) , the streams received in the corner will approximately grow like 2k1 log n,
while the streams received at the center will grow approximately like (k1 n log n) 2 . A hot
spot [36] is formed at the center of the network. However, even with a perfectly balanced routing algorithm, the nodes at the center will not be required to handle much less trafc. Indeed, ignoring constants, with roughly Divided among the
1
n log n
n cells, each stream will need roughly ( log n ) 2

1 1
n relays, even with minimum-hop routing. With n streams, there is need for n( log n ) 2 relays. n log n
cells, each cell will be relaying (n log n) 2 streams on average.
Therefore, even if the routing algorithm created a perfectly balanced load, and required
n only around ( log n ) 2 relays per stream, the cells at the center would experience a trafc load
equal (up to a constant which is independent of n) to that of the current routing algorithm. Knowing the expected number of streams that a given cell will be receiving is useful, however we still need to bound the actual number of streams that will be arriving at any cell. We have:
86
Lemma 3 (Bounds on the actual number of streams arriving at a cell) Under all three trafc models TM1, TM2, TM3, P [si 16(k1 n log n)
1 2
i] a 1
4n 2 k1 f (1) (k1 log n) 2

3
(5.13)
Proof: Because of the routing protocol, for each stream that passes through ci , at least one of the following two conditions must be satised: either the source and ci share the same x-coordinate, or the destination and ci share the same y-coordinate. It is also possible that both conditions are satised. Let si1 be the number of streams that satisfy the rst condition, and si2 be the number of streams that satisfy the second condition. Clearly, si si1 + si2 . Also, let Mi be the number of nodes that have the same x-coordinate with ci , and Ni
the number of nodes that have the same y-coordinate with ci . We have P [Mi > 4(k1 n log n) ] P [Ni > 4(k1 n log n) ] Indeed, we have P [Mi > 4(k1 n log n) 2 ] = P [Mi > 4k1 log n(
1 n )2 ] k1 log n 1 1 1 2 1 2
n 2 k1 f (1) (k1 log n) 2 n 2 k1 f (1) (k1 log n) 2

1 1 1
, .
(5.14) (5.15)
(g(n)) 2 nk1 f (1) (union bound, symmetry, (5.12)) n 2 k1 f (1)

1
P [Mi > 4k1 log n(g(n)) 2 ] (using (5.8))

1
(k1 log n) 2
, (using (5.8))
and we arrive at (5.14). The proof of (5.15) is similar. Now we concentrate on TM1. By denition, each node is the source of a single stream, so si1 Mi . Similarly, each node is the destination of a single stream, so si2 Ni .
87
Combining these inequalities with the inequality si si1 + si2 , we have that si Mi + Ni . Applying the union bound and (5.14), (5.15), we have: P [si > 8(k1n log n) ] 2
1 2
n 2 k1 f (1) (k1 log n) 2

1
By symmetry, this result holds for all cells ci . Therefore, applying the union bound once more we arrive at: P [ i : si > 8(k1 n log n) ] which is stronger than (5.13).
1 2
2n 2 k1 f (1) (k1 log n) 2

3
Next, we consider the case of TM2. We rst concentrate on si2 : P [si2 > 8(k1 n log n) 2 ]
n
1
=
m=1
P [si2 > 8(k1 n log n) 2 |Ni = m]P [Ni = m]

1
4(k1 n log n) 2 m=1 4(k1 n log n) 2 m=1 4(k1 n log n) 2 m=1

1 1
P [si2 > 8(k1 n log n) |Ni = m] + P [X1 > 8(k1 n log n) |Ni = m] + P [X2 > 8(k1 n log n) |Ni = m] +
1 1 2 1 2 1 2 1 2
1 2
n 2 k1 f (1) (k1 log n) 2 n 2 k1 f (1) (k1 log n) 2 n 2 k1 f (1) (k1 log n) 2

1 1 1 1 1 1 1
(using (5.15))
4(k1 n log n) P [X3 > 8(k1 n log n) ] +
n 2 k1 f (1) (k1 log n) 2
1 1 3 n n 2 k1 f (1) a 4(k1 n log n) 2 P [X3 > 4(k1 n log n) 2 ]+ 1 2 n1 (k1 log n) 2 1
<a
n 2 k1 f (1) n f (0.5)] + (using (5.10)) 4(k1 n log n) exp[4(k1 n log n) 1 n1 (k1 log n) 2
1 2 1 2
2n 2 k1 f (1)
(k1 log n) 2
(5.16)
88
In the second inequality, we dene X1 as the number of streams that end up in one of the Ni cells who share the same y-coordinate with ci . Clearly X1 si2 . Conditioned on success equal to success
m1 . n m , n1
Ni = m, X1 is the sum of n m Bernoulli random variables, each with probability of and another m Bernoulli random variables, each with probability of
m n1
The third inequality then follows if we dene X2 as a binomial random variand number of experiments n. The fourth inequality
4(k1 n log n) 2 n1
1
able, with probability of success
comes by dening X3 as a binomial with probability of success rst.
and number of
experiments n. The last inequality comes from noting that the second term dominates the
So now we have: P [si > 12(k1 n log n) 2 ] P [si1 > 4(k1 n log n) 2 ] + P [si2 > 8(k1 n log n) 2 ] (union bound)
1 2 1 1 1
<a P [si1 > 4(k1 n log n) ] + P [Mi > 4(k1 n log n) ] + 3n 2 k1 f (1) (k1 log n) 2
1 1 1 2
2n 2 k1 f (1) (k1 log n) 2 2n 2 k1 f (1) (k1 log n) 2

1 1 1
(using (5.16)) (trafc model)
. (using (5.14))
The third inequality comes from noting that, because of the trafc model, each of the si1 streams will correspond to a distinct node with the same x-coordinate with ci . Applying the union bound, we arrive at: P [ i : si > 12(k1 n log n) ] <a which is stronger than (5.13).
1 2
3n 2 k1 f (1) (k1 log n) 2

3
Finally, we consider TM3. By using a sequence of inequalities similar to that leading to (5.16), in the case of TM2, we have that (5.16) continues to hold. Furthermore, si1 will
89
A1
A2
A3
A8
1111 0000 1111 0000 1111 0000 1111 0000 111 111 000 c 000 111 111 000 0 000 1111 0000 111 111 000 000 1111 0000 1111 0000 1111 0000
A4
A7
A6
A5
Figure 5.4: One of the 9 sub-lattices of cells appears shaded. Only nodes in the sub-lattice are
allowed to receive in the corresponding slot, and only from nodes in the same or neighboring cells. The neighbors of cell c0 appear striped.
also satisfy a similar inequality: P [si1 > 8(k1 n log n) ] a We then have: P [si > 16(k1 n log n) 2 ] a P [si1 > 8(k1 n log n) 2 ] + P [si2 > 8(k1 n log n) 2 ] (union bound) 4n 2 k1 f (1) (k1 log n) 2
1 1 1 1 1 1 2
2n 2 k1 f (1) (k1 log n) 2

1
(5.17)
. (using (5.16), (5.17))
Applying the union bound, we arrive at (5.13).
90
5.3.3 The time division scheme

By the construction of the cell lattice, if n is the number of nodes in the network, there will
n be g(n) = ( k1 log n ) 2 2 = (2r + 1)2 cells, g(n) being a multiple of 9. Therefore, we
1
may divide the g(n) cells perfectly into nine sub-lattices. We index the nine sub-lattices by whose coordinates may be expressed as (i + 3k1 , j + 3k2 ), for some k1 , k2 Z. In Fig. 5.4
the pairs (i, j), where 1 i, j 1. The cells belonging to sub-lattice (i, j) are all those
we have shaded the cells belonging to one of the 9 sub-lattices.
We divide time into frames, and each frame into nine slots. Each slot corresponds to a sub-lattice. At any time during that slot, only one node from each cell of the sub-lattice is allowed to receive (but many nodes of that cell may receive consecutively in the same slot). As specied by the routing protocol, the transmitter of that transmission will have to lie in the same cell, or in one of the four neighboring cells. All transmissions will be with the maximum power P0 . We will clearly need a lower bound on the Signal to Interference and Noise Ratio (SINR) at each receiver. Intuitively, such a bound exists, as the time division scheme spaces out the interferers. In fact, we can prove the following: Lemma 4 (Lower bound on the SINR) In the absence of fading, the SINR Cj at node Xj , where j = 1, . . . , n, is asymptotically lower bounded by 36 10 1 2 35 Cj a Cmin (n) (100 log n)1 2 1 n 10 2 ( k1 log n ) 2 1 32 Proof: We rst bound the interference. Let x0 =
1 2r+1
if > 2, if = 2, if 1 < < 2. (5.18)
be the length of the sides of the cells,
and let c0 be the cell in which the receiver lies. Working as in [64], we note that the rest of the cells in the same sub-lattice are located along the perimeters of concentric squares, whose center is cell c0 . For example, there are 8 cells along the perimeter of the rst square (fewer if the cell is at the edge of the square region). Irrespective of the coordinates of c0 , all the cells of its sub-lattice are located along the perimeters of at most 2r squares. There
91
are at most 8i interferers corresponding to the i-th square, whose distance from the receiver by will be at least x0 (3i 2). Consequently, the interference at the receiver is upper bounded
2r
Ij
= < < =
i=1
8iKP0 [x0 (3i 2)]

2r1 i=1 2r 0 2r 0
8KP0 [1 + x 0 8KP0 [1 + x 0 8KP0 [1 + x 0
i+1 ] (3i + 1) x+1 dx] ( > 1) (3x + 1)
(3x + 1)1 dx] 1 + (6r+1)2 + 1 3(2) 3(2) 8KP0 (2r + 1) 1 + 1 log(6r + 1) 3 35 ( n ) if > 2, 36 k1 log n 2 n 8KP0 k if = 2, 1 2 3 n if 1 < < 2.
2 k1 log n
if = 2 if = 2
(5.19)
We also need a lower bound on the power of the useful signal. Clearly, since the maximum possible distance that the useful signal will need to travel, under the routing assump tions, is 5x0 , the power of the useful signal Sj is bounded as follows: Sj KP0 ( 5x0 ) >a 10 2 KP0 (
n )2 . k1 log n
(5.20)
Combining (5.19) with (5.20), and noting that the thermal noise remains bounded, and therefore becomes negligible as n , we arrive at the result.
5.3.4 Performance of the basic scheme

We will refer to the algorithm specied in Sections 5.3.1-5.3.3 as the basic scheme. We are now ready to prove our rst main result, Theorem 1:
92
Proof of Theorem 1: We set k1 = 10 > max(8, 2f3 ). By Lemma 1, with high prob(1) ability, i.e., with probability approaching 1 as n , each cell will contain a node that will be used for routing. In addition, by Lemma 3 no cell will need to serve more than 16(k1 n log n) 2 routes, with high probability. By Lemma 4, each receiver is guaranteed a reception rate of at least Rmin (n) = fR (Cmin (n)) = W log2 (1 + 1 Cmin (n)), where Cmin (n) is given by (5.18). Noting that, due to the time division scheme, each cell will be able to receive packets during only one out of nine slots, each stream is guaranteed, Substituting Cmin (n) from Lemma 4 and using the limit lim with probability approaching 1 as n , a rate equal to (n) = log2 (1 + x) = log2 e > 1, x0 x we arrive at the result.
W log2 (1+1 Cmin (n)) 916(k1 n log n) 2
1 1
(5.21)
5.4 Immobile Nodes with Fading

Under the basic scheme, each node with a packet waiting to be relayed through a neighboring cell will arbitrarily pick one of the nodes in that cell as the relay. In the absence of fading, the choice of relay is not critical. With fading, the choice is clearly important, as the power gains from one node to various nodes of the neighboring cell may be drastically different. Fortunately, by Lemma 1 there are many potential receivers for a transmitter to choose from, provided k1 is large enough. We therefore use the basic scheme, with a value for k1 to be determined, and with a single modication on the routing rules. Specically, when it is the turn of a node to transmit, the transmitter will pick as the receiver any of the nodes in the target cell, among those whose link with the transmitter has a fading coefcient greater than or equal to the median of the distribution, fm . (By (5.4), the probability that a node in the cell will satisfy
1 this requirement is at least 2 .) If such a node does not exist, the packet is dropped. This
rule is extremely intuitive: nodes plainly avoid transmitting to nodes with poor fading coefcient. There is, however, a complication. Once a packet arrives at its destination cell, it may not nd itself at its destination node. This can easily be amended as follows: Let the packet be in node X1 , and have a destination node X2 , both being in the same cell ci .
5.4. IMMOBILE NODES WITH FADING
93
The packet will be transmitted two more times, once to an intermediate node X3 , and then from the intermediate node to the destination node X2 . The intermediate node is chosen among the nodes in cell ci , such that the fading coefcients of both links are greater than fm . If no such node exists, the packet is dropped. We will refer to this algorithm as the spatial-diversity scheme. For a given number of nodes n, there is a positive probability that the packets of some of the streams will have to be dropped, because at some point along the way no node can be found to satisfy the restriction we place on the fading coefcients. However, as the number of nodes increases this probability goes to zero, if k1 is chosen large enough. To show this, let us rst condition the discussion on the event E = [ k1 log n mi 4k1 log n i]. 2 The probability that a given node will not be able to send a packet to any of the nodes in
1 a neighboring cell is at most ( 2 )
k1 log n 2
. There are n nodes, each with at most 4 neighboring
cells. Using the union bound, we have that the probability that any of the nodes cannot forward a packet to one of its neighboring cells, because of fading, is upper bounded by 4n1
log(2)k1 2
.
k1
Also, note that the probability that a given node will not be able to send a packet to
3 another node in the same cell by means of a relay in that cell is at most ( 4 ) 2 log n
. There
are n nodes, each with at most 4k1 log n potential destinations in the same cell. Using the union bound, we have that the probability that any one of the nodes cannot forward a packet to one of the other nodes in its own cell by means of a relay is upper bounded by 4k1 n1 2
k1 4 log( 3 )
log n.
k 1 1
By Lemma 1, the probability that the event E will not occur is upper bounded by 2n log8n . k1 Let D be the probability that the packets of one or more of the streams are dropped. Then: P [D] = P [D|E]P [E] + P [D|E c]P [E c ] P [D|E] + P [E c ]
1
log(2)k1 2
a 4n
+ 4k1 n
k1 2
4 log( 3 )
2n1 8 log n + . k1 log n
k1
(5.22)
Therefore, provided that we select a high enough value for k1 , by taking advantage of the existing spatial diversity, all nodes will be using strong paths and no packet will be dropped. We will also need an upper bound on the value of the power gains fij . Such a bound can
94
be easily derived, by virtue of the exponentially thin tail of the fading distribution. Indeed, exp[qk2 log n]. Noting that there are bound we have: P [max{fij } k2 log n] a 1 n2qk2 . We now are now ready to prove the main result of the section, Theorem 2:
2 2 Proof of Theorem 2: Pick k1 = 10 > max{ log 2 , log 4log 3 , 8}. By (5.22), no packets will
focusing on an arbitrary fading coefcient, we have, using (5.3), that P [fij > k2 log n] a
n(n1) 2
< n2 fading coefcients, by use of the union (5.23)
be dropped, with high probability. Let us next concentrate on the SINR at the receivers. Because of the restriction we place on which paths may be used, if Sj is the received power of the useful signal without we set k2 = 3 , by (5.23), and with high probability, all fading coefcients will be smaller q
F than k2 log n. Therefore, if Ij is the interference power at Xj with no fading and Ij is the F interference power at Xj with fading, we will have that Ij (k2 log n)Ij . This bound will F F fading and Sj is the received power with fading, we will have Sj fm Sj . In addition, if
hold uniformly, i.e. for all receptions, and with high probability. Noting that the thermal
F probability, a minimum SINR Cmin (n) equal to
noise power is not affected by the presence of fading, all nodes are guaranteed, with high
F Cmin (n) =
fm Cmin (n), k2 log n
(5.24)
where Cmin (n) is the minimum SINR in the absence of fading, given by (5.18). ThereF F fore, with high probability, each receiver is guaranteed a rate Rmin (n) = fR (Cmin (n)) = F W log2 (1 + 1 Cmin (n)).
By Lemma 3, a maximum of 16(k1 n log n) 2 streams will be arriving at each cell. For a few of those, specically those whose destination lies in that cell, three receptions will receptions will have to take place in each cell. Noting that each cell will be receiving in only one of every nine slots, we see that each stream is guaranteed, with high probability, a rate equal to
F W log2 (1+1 Cmin (n))
be required, according to the routing rules. Therefore, a maximum of 3 16(k1 n log n) 2
432(k1 n log n) 2
F . Substituting Cmin (n) from (5.24) and (5.18), and using
the limit of (5.21), we arrive at the result.
5.5. NODE MOBILITY WITH A FIXED DELAY CONSTRAINT

Ns
95
} } }
(a)
2n D FRAMES
} }
(b)
2nD MINI-FRAMES
N MINI-FRAMES
s 9
s 9
9 SLOTS
9 SLOTS
Figure 5.5: (a) The frame structure used in the xed-delay and polynomial-delay schemes. (b) The frame structure of each of the N virtual channels.
5.5 Node Mobility with a Fixed Delay Constraint

5.5.1 The xed-delay scheme
As in Sections 5.3 and 5.4, we divide the area of the network into a regular lattice of g(n) later. The cells are again divided into the nine sub-lattices (i, j) where 1 i, j 1. Note
1 1 n 1 n ( ) 2 <a g(n) ( )2 . 2 k1 log n k1 log n
n cells. Now, however, we require that g(n) = ( k1 log n ) 4 2 , where k1 will be specied
that
(5.25)
As shown in Fig. 5.5(a), we divide time into frames, with each frame consisting of N mini-frames. Each mini-frame will have a duration equal to s, and will consist of 9 slots,
s each of duration 9 . In each of the slots, only nodes lying in a corresponding sub-lattice are
allowed to transmit (and only with maximum power). The rest will have to remain silent. In addition to this time division, which is on the 9 slots of each mini-frame, we superimpose another, distinctly different time-division, which is on the N mini-frames of each frame. Specically, the N mini-frames within the frame are logically independent: packets that are transmitted, received, and created in a given mini-frame, are stored in the memory of the nodes for N 1 mini-frames, and then brought back forward for the corresponding mini-frame of the next frame. We thus create N independent virtual channels, with the nodes communicating using each of them independently. As a result, the topology in each of these virtual channels becomes totally decorrelated every other mini-frame, as opposed to every N mini-frames.
96
In the following, we suppress all reference to the outer time division, and we concentrate on a single virtual channel. Therefore, our working model becomes that of Fig. 5.5(b). It should be understood, though, that the nodes will be executing the same algorithm independently, and concurrently, in each of the N virtual channels. The algorithm executed in each of the N virtual channels is as follows: In the rst, third, and generally all odd mini-frames, nodes wait for the slot in which they are allowed to transmit (this depends on the sub-lattice in which the node lies). They then transmit a
F F packet, with rate Rmin (n) = fR (Cmin (n)), where Cmin (n) is given by (5.24), where k2
will be specied shortly, for a time duration equal to

1
1 s (k n log n) 2 . 94 1
Nodes that lie in
the same cell coordinate among themselves, and transmit their packets consecutively. If there are more than 4(k1 n log n) 2 nodes in any cell, then some of the packets will not be transmitted and will be dropped. We will call this a type-I error. The transmitted packets will possibly be received successfully by some of the other nodes lying in the same cell. These will act as relays in the subsequent mini-frame. In the second, fourth, and generally all even-numbered mini-frames, each destination node will wait for the slot in which nodes in its cell can transmit, and then will receive the packet intended for that destination by one of the relays lying in the same cell. That relay will have successfully received the packet in the previous mini-frame, and has a sufciently strong path to the destination. We say that a type-II error occurs if there are more than 4(k1 n log n) 2 nodes in a cell, so that some of them will not have the time to receive their packets. We say that a type-III error occurs when, for a specic packet, no node is able to successfully act as a relay, so that the packet never makes it to the destination. At the end of each even-numbered mini-frame, nodes discard all packets remaining in their buffers. We will refer to this algorithm as the xed-delay scheme.
1
5.5.2 No packets are dropped

We now show that the probability that an error of type-I, II, or III occurs at any time, at any dropped. part of the network, goes to zero as n , so that, with high probability, no packets are We rst bound the number of nodes at each cell of each mini-frame. Let mij be the
97
number of nodes in cell ci , i = 1, . . . , g(n) of mini-frame j, j = 1, . . . , 2NnD . Then:

1 1 P [mij < (k1 n log n) 2 ] 2 n P [mij < ] (Using (5.25)) 2g(n) n 1 < exp[ ] (Using (5.9) with = ) 8g(n) 2 1 1 exp[ (k1 n log n) 2 ] (Using (5.25)) 8
(5.26)
Similarly, P [mij > 4(k1 n log n) 2 ] n <a P [mij > 2 ] (Using (5.25)) g(n) nf (1) < exp[ ] (Using (5.10) with = 1) g(n) exp[f (1)(k1 n log n) 2 ]. (Using (5.25))
1 1
(5.27)
Applying the union bound, we arrive at:

1 1 1 P [ (k1 n log n) 2 mij 4(k1n log n) 2 i, j] 2 1 1 1 1 n a 1 2nD N( ) 2 {exp[ (k1 n log n) 2 ] + exp[f (1)(k1 n log n) 2 ]} k1 log n 8 1 k1 > 0. (5.28)
By (5.28), type-I and type-II errors are excluded with high probability.
Next, we exclude type-III errors. For this, we rst bound the value of the fading coefcients. Let fijk be the fading coefcient between nodes Xi and Xj , at mini-frame k, and let the event F = {max{fijk } < k2 log n}. Noting that there are 2NnD mini-frames, by the union bound and (5.23) we have
P [F ] a 1 2NnD+2qk2 .
(5.29)
98
We now condition the discussion on F . A condition that is sufcient (but not necessary) for a packet to be successfully received when the transmitter and receiver lie in the
F F same cell with rate Rmin (n) = fR (Cmin (n)), where Cmin (n) is given by (5.24), is that
the fading coefcient between the receiver and transmitter is equal to or greater than fm (this happens with conditional probability at least 1 , for sufciently large n). This result 3 follows from noting that the bounds (5.19) and (5.20) that were derived in the context of the basic scheme also apply to the xed-delay scheme with no modication. Indeed, in the xed-delay scheme the transmitter of the useful signal must be placed closer (in the same cell as the receiver), and the interferers must be placed further than in the case of the basic scheme. Excluding the source and the destination, there are n 2 potential relays. The prob-
ability that one of them will be in the same cell as the source in the mini-frame when the packet leaves the source, and will be in the same slot as the destination in the subsequent mini-frame, and will have sufciently strong paths to both source and destination, in both the n 2 potential relays will be successful in relaying the packet is bounded by: pf = (1 k1 log n n2 ) 9n k1 log n )] 9n k1 log n] 10 mini-frames, is p(n)
1 1 11 g(n) g(n) 3 3
1 ( k1 log n ). Therefore, the probability pf that none of 9 n
exp[(n 2) log(1
k1
<a exp[ = n 10 .
Since the total number of packets is NnD+1 , by the union bound, the probability that a type-III error occurs conditioned on the event F , P [III|F ] <a NnD+1 10 . Removing the conditioning on F , we have that the probability of a type-III error occurring, P [III], is bounded by: P [III] = P [III|F ]P [F ] + P [III|F c ]P [F c ] P [III|F ] + P [F c ]
k1 k1
a NnD+1 10 + NnD+2qk2 .
(5.30)
99
Therefore, by setting k1 = 10(D + 2) and k2 = probability.
D+3 , q
type-III errors are excluded with high
5.5.3 Performance of the xed-delay scheme

Proof of Theorem 3: Regarding the packet delay, we assume that packets are created
just before it is their nodes turn to transmit. (Delays due to random arrival times and queuing are beyond the scope of this work such a study is undertaken in [65].) We then note that each packet will arrive at its destination within an interval (N + 1)s after its transmission from the source. Therefore, packet delays are smaller than a maximum dmax = (N + 1)s. This bound does not depend on the number of nodes n, but rather on the underlying mobility model: it is slightly longer than the decorrelation time of the topology. To calculate the per-stream throughput achieved by the scheme, we start by noting that a packet of size
1 s F (k n log n) 2 Rmin (n) 36 1
bits will be arriving at each destination evF = W log2 (1 + 1 Cmin (n)), substituting
ery 2 mini-frames, of duration 2s. Therefore, the per-stream throughput will be (n) =
1 1 F (k n log n) 2 Rmin (n). Noting that Rmin (n) 72 1 F Cmin (n) from (5.24), (5.18), and using the limit
of (5.21), we arrive at the result.
The theorem states that a non-trivial aggregate throughput is possible when nodes are mobile, even with a xed upper bound on the acceptable packet delay and without the nodes using global routing and topology information. For the case 2, this aggregate throughput increases roughly like n 2 (log n) 2 . This happens to be equal, up to a constant,
1 3
to the aggregate throughput that nodes would roughly achieve if they were using the spatialdiversity scheme, properly modied to handle node mobility, and assuming that they had access to global topology information. However, for large networks and high node mobilities, such information comes at a prohibitive cost, or is just impossible to acquire. But network designers need not despair: Theorem 3 states that the same aggregate throughput is achievable, with no need for routing or global topology information, provided nodes coordinate within their single-hop neighborhood, and a packet delay roughly equal to the topology decorrelation time is tolerated.
100
5.6 A Polynomially Increasing Delay Constraint

In Section 5.5 it was shown that a non-trivial per-node throughput is possible even with a xed upper bound on the delay. It is natural to ask how much better we can do if larger delays are tolerated, in particular delays that increase with the number of nodes. Looking back at the xed-delay scheme, it is clear that the throughput is limited by our requirement that the number of relays per packet (which is around (k1 n log n) 2 ) be larger than the
n number of cells (which is around ( k1 log n ) 2 ). If this requirement is relaxed, packets may
1 1
need to spend more than one slot at a relay, but since there are now more cells, and hence more simultaneous transmissions, the per-node throughput increases.
5.6.1 The polynomial-delay scheme

So let us assume that packet delays up to dmax (n) = 2(nd + 1)Ns (4Ns)nd (5.31)
are acceptable, where 0 < d < 1. We assume that packets are created just before their sources turn to transmit. We employ the usual cell lattice, with the number of cells equal to g(n) = kn log n ) 4 2 , where k1 will be determined later. Clearly, 1 nd+1 1 1 nd+1 1 ( ) 2 <a g(n) ( )2 . 2 k1 log n k1 log n (5.32)
d+1 1
Again, we divide time into frames, each frame in N mini-frames, and each mini-frame in
s nine slots of duration 9 , as shown in Fig. 5.5(a). As previously, mini-frames are organized
in N virtual channels. As with the xed-delay scheme, within each of the N virtual channels, odd numbered mini-frames are devoted to source-relay communication, and even-numbered mini-frames are devoted to relay-destination communication. During the odd-numbered slots, nodes behave similarly to the xed-delay scheme: Each one waits for the turn of the slot that
F F corresponds to its cell, and transmits with rate Rmin (n) = fR (Cmin (n)), where Cmin (n) is
given by (5.24), a packet of length
s (k1 n1d 36
log n) 2 (k3 log n)1 , with k3 a constant that
will be specied later. Nodes that lie in the same cell coordinate, and transmit their packets
5.6. A POLYNOMIALLY INCREASING DELAY CONSTRAINT
101
consecutively. If the number of nodes in a cell is greater than 4(k1 n1d log n) 2 (k3 log n), then some of the packets will not be transmitted and will be dropped. We will call this a type-I error. The transmitted packets will be received by some of the other nodes lying in the same cell. These will act as relays in the subsequent mini-frames. Contrary to the xed-delay scheme, where each relay keeps packets for a single miniframe, relays now keep the packets for the next (nd + 1) even-numbered mini-frames. After that time, packets are removed from their buffers, as the delay constraint specied in (5.31) can no longer be satised. In the second, fourth, and generally all even-numbered mini-frames, each destination will wait for the slot in which nodes in its cell can transmit, and then will attempt to receive all packets intended for it, by all relays that happen to be in the same cell. The relays will coordinate among themselves and transmit consecutively. Relays will only transmit if their fading coefcient to the destination is greater than fm . In addition, if more than two relays have the same packet, then only one of them will transmit it. We say that a type-II error occurs if there are more than 4(k1 n1d log n) 2 (k3 log n) distinct packets in a cell, so that some of them may not have the time to be transmitted to their nal destination. We say that a type-III error occurs when, for a specic packet, there is no relay with sufciently strong paths between both itself and the source, and itself and the destination, so that the packet never makes it to the destination, in any of (nd + 1) even-numbered mini-frames as the polynomial-delay scheme. for which the packet is waiting to be transmitted by its relay. We will refer to this algorithm
1
5.6.2 No packets are dropped

Here, we show that no packet is dropped, with high probability. Note that some packets may be received by their destination node multiple times, as relays do not coordinate among themselves and many of them might transmit the same packet to its destination, at different mini-frames. (This implies lost bandwidth, but the per-stream throughput is not reduced by more than a logarithmic factor.) Because of this, the system has natural redundancy, and the occurrence of a type-II error does not imply that a packet is lost. However, to prove that no packet is lost, it sufces to show that no errors of any type will be occurring.
102
Type-I errors are excluding by replicating the steps that led to (5.28), so that we have
1 1 1 P [ (k1 n1d log n) 2 mij 4(k1n1d log n) 2 i, j] 2 1 1 n1+d 1 1 a 1 2nD N( ) 2 {exp[ (k1 n1d log n) 2 ] + exp[f (1)(k1 n1d log n) 2 ]} k1 log n 8 1 k1 > 0. (5.33)
We now exclude type-II errors. Let P (II) be the probability that a type-II error occurs at some mini-frame and cell, pII be the probability of a type-II error in a particular cell and mini-frame, and pm be the same probability conditioned on the fact that there are m II nodes in that cell and mini-frame. We rst focus on bounding pm . Let the nodes be indexed II by X1 , . . . , Xm , and let Y1 , . . . , Ym be the number of distinct packets intended for each of them, and carried by one or more of the rest of the nodes in the cell. Since each packet will require for its transmission a length of time equal to
s (k n1d 36 1
1
log n) 2 (k3 log n)1 , we
must ensure that there are no more than 4(k1 n1d log n) 2 (k3 log n) of them. So we have:
m
pm = P [ II
i=1
Yi > 4(k1 n1d log n) 2 (k3 log n)]

1
4(k1 n1d log n) 2 (k3 log n) > )] m 1 4(k1 n1d log n) 2 (k3 log n) mP [Y1 > ] (union bound, symmetry) (5.34) m P [m (Yi i=1
could be among the Y1 existing in the cell (if the slot is within one of the rst nd evento their relays at distinct slots. Therefore, conditioned on the number of potential relays in the cell. If we remove the conditioning, the independence no longer holds: If a packet exists in the cell, this increases the probability that there are many nodes in the cell, and this in turn increases the probability that other packets will also exist in the cell. Therefore, conditioned on the number of nodes in the cell m, Y1 is binomially distributed, with at most numbered mini-frames, the packets are fewer than nd + 1 ). All of them were offered
Now there are at most nd + 1 packets, indexed by P1 , . . . , P(nd +1) for node X1 that
m, the existence of any of them in the cell is independent of the existence of any other
5.6. A POLYNOMIALLY INCREASING DELAY CONSTRAINT
103
nd + 1 Bernoulli experiments and probability of success per experiment pm , where pm 1 (1 1 m ) g(n) 2(k1 log n) 2
d+1 2 1
a 1 (1 a
a 3(k1 log n) 2 n
n d+1 1 1 exp[3(k1 log n) 2 n 2 m]

1 d+1 2
)m
We can now bound pII . Let qm be the probability that there are m nodes in the cell:
4(k1 n1d log n) 2 1 m=1 4(k1 n1d log n) 2 1 m=1
1 1
pII
pm qm II
+
m=4(k1 n1d log n) 2
1
pm qm II
1
4(k1 n1d log n) 2 (k3 log n) ]qm mP [Y1 > m

1 1
+P [m > 4(k1 n1d log n) 2 ] (using (5.34)) 4(k1 n1d log n) 2 P [Z > k3 log n] + P [m > 4(k1 n1d log n) 2 ],
1
2nd experiments, each with probability of success 12k1nd log n. Setting k3 = 48k1, and using (5.10), we have that P [Z > k3 log n] < exp[12f (1)k1 log n]. Therefore, pII a 4(k1 n1d log n) 2 n12f (1)k1 + P [m > 4(k1 n1d log n) 2 ] a 5(k1 n1d log n) 2 n12f (1)k1
1 1 1
where in the last inequality Z is a binomially distributed random variable, with nd + 1 <
The last inequality comes by noting that, by (5.33), the rst term dominates the second. Applying the union bound, we have that the probability of a type-II error at any cell and even mini-frame is P (II) a 10Nn1+D12f (1)k1 . Setting k1 >
D+1 , 12f (1)
we exclude type-II errors, with high probability.
104
Having excluded type-I and II errors, we know that no packet will be denied transmission because of an overcrowding of nodes. However, we still need to prove that all packets will be given the chance in the rst place, i.e., we need to exclude type-III errors. There is, however, a technical issue that must be dealt with rst: Those packets that will be created toward the end of the experiment, within a time dmax (n) from the end, will not stay in the relays as long as the rest. It is clear that some of them will never reach their destination. However, the fraction of these packets is smaller than
(4N s)nd . (2N s)nD
As this fraction goes to 0 as
n increases, we will tolerate type-III errors for these packets. We concentrate on a particular packet P . Using the previous paragraph for justication, we assume that P is not created within a time period equal to dmax (n) from the end. We also condition the discussion on the event F = {max{fijk } < k2 log n}, for which (5.29) continues to hold.
As in the xed-delay scheme, conditioned on F , a packet will be successfully received if the fading coefcient between the receiver and the transmitter is greater than or equal to
1 fm (and so with probability at least 3 , for sufciently large n). Counting the source and
the destination, there are n potential relays. The probability that one of them in particular, say X1 , will receive the packet from the source is greater than or equal to to the destination is at least [1 (1 x > 0, log(1 + x)
1 . 3g(n)
Assuming
that this happens, the probability that X1 will have the opportunity to forward the packet to successfully act
d 1 )n +1 ]. So the probability that X1 will be able 3g(n) d 1 1 as a relay is p1 3g(n) [1 (1 3g(n) )n +1 ]. Noting that, for small log < x and 1 ex x, we have p1 >a k110n n . However, there are n k1
potential relays, so we have that the probability that packet P will not be relayed by any node is upper bounded by n 10 . Using the union bound, the probability of a type-III error, conditioned on F , P [III|F ], is upper bounded by P [III|F ] a NnD 10 . Removing the conditioning on F as in (5.30), we have: P [III] a NnD 10 + NnD+2qk2 , from which we have that by setting k1 > 10D and k2 > with high probability.
D+2 , q
k1 k1
(5.35)
type-III errors do not occur,
5.7. CONCLUSIONS AND FUTURE WORK
105
5.6.3 Performance of the polynomial-delay scheme

and k2 =
D+1 Proof of Theorem 4: We have set k3 = 48k1 . We also set k1 = 11D > max{10D, 12f (1) } D+3 . q
By Section 5.6.2, errors of all types are excluded with high probability.
1
We still need to calculate the throughput of the scheme. Each packet will have a size
s of Rmin (n) 36 (k1 n1d log n) 2 (k3 log n)1 bits. In each pair of mini-frames, of duration
2s, n such packets are created, all of them arriving at their destination (excluding a negligible fraction of those, all created in the last 4NnD mini-frames). This brings the per1 stream communication rate to 4 (n) = Rmin (n) 72 (k1 n1d log n) 2 (k3 log n)1 . Noting F F that Rmin (n) = W log2 (1 + 1 Cmin (n)), substituting Cmin (n) from (5.24), (5.18), and
1
using the limit of (5.21), we arrive at the result.
5.7 Conclusions and Future Work

We derive lower bounds on the capacity of wireless ad hoc networks with a large number of nodes. Following [36] and [38], our approach is to construct schemes that achieve a given per-stream throughput with probability approaching unity as the number of nodes increases. We rst study networks with immobile nodes, in the absence of fading. We introduce the basic scheme, which achieves a per-stream throughput 1 (n) = K1 (n log n) 2 . Although a similar result rst appeared in [36] under a similar setting, the proof presented here is signicantly shorter and is using only basic tools of probability, most notably the union bound and Chernoffs bounds. We then introduce the spatial-diversity scheme which, in the presence of fading, achieves a per-stream throughput 2 = K2 n 2 (log n) 2 . We then study networks where the nodes are mobile, the channel exhibits fading, and delay constraints are placed on the delivery of packets. We introduce the xed-delay scheme that achieves a per-node throughput 3 = K3 n 2 (log n) 2 with a xed upper bound on packet delay. We then identify a fundamental delay-throughput tradeoff, showing that if the nodes are willing to endure packet delays on the order of nd then, using the polynomial-delay scheme, a per-node throughput 4 = K4 n
d1 2 1 3 1 3 1
(log n) 2 is achiev-
able. With both schemes, and contrary to most traditional routing protocols, nodes need no global topology or routing information and need to coordinate only locally, i.e., only with
106
nodes that are one hop away. It is assumed that the received power (or its expected value, in the case of fading) decays exponentially with distance, with an exponent > 2. Similar results are derived when 1 < < 2. Also, these results hold for an appropriate choice of constants Ki , and with high probability, i.e., with probability approaching 1 as the number of nodes approaches innity. Because we do not resort to the law of large numbers or any other theorem of a similar avor, we are able to identify in all cases the rate with which this probability goes to unity. A possible future line of research would be to investigate the behavior of the capacity when nodes are mobile and the packet delay must remain smaller than an arbitrary bound specied by the nodes. In contrast, the xed-delay scheme guarantees a packet delay which is dictated by the mobility model and may be larger than what the nodes are willing to tolerate. In both the xed-delay and the polynomial-delay scheme, the packet size decreases with n. It is appealing to require that the size of packets remain constant. However, this requirement may lead to increased packet delays. Another possible extension, in a drastically different direction, would be to incorporate queuing delays in the formulation. Note that, in all our schemes, we assume that the creation of packets follows a regular and deterministic rate. If we assume that the arrival times are random, queuing delays are introduced. The use of the regular cell lattice will be particularly useful in this setting. In the spatial-diversity scheme, the fading coefcients were taken to be constant. It is straightforward to show that all results continue to hold, with minor modications, even when the fading coefcients are time-varying. This assumption would be most relevant in the case where nodes are allowed to move, but only within the connes of a limited area. All our derivations can be somewhat simplied if we tolerate losses in the achievable throughputs of our schemes on the order of n , and ignore logarithmic factors. For example it may be shown, using proofs somewhat shorter than the ones presented here, that both the basic and the spatial-diversity scheme achieve a per-node throughput greater than n 2 for all . Proofs become shorter because the values for multiplicative constants are no longer critical, as happens for example with the constants k1 and k2 that appears in (5.8) and (5.23), respectively. Moreover, it is reasonable to expect that in some settings, there is a
1
5.7. CONCLUSIONS AND FUTURE WORK
107
xed exponent o (the order) such that throughputs of the form f (log n)no for any rational function f(n) are not achievable, but throughputs of the form no are achievable for all > 0. In such settings, sacricing a factor equal to n is even more appealing. As an example, let us consider a setting in which nodes have a complete knowledge of the current and future positions of all other nodes. Although this assumption initially appears unrealistic, it may be relevant in applications in which the movement of nodes is periodic. In such a case, we conjecture that the aggregate throughput can increases as fast as n1 , for all > 0, even with a xed upper bound on the packet delay. Intuitively, if nodes can communicate only with n of their closest neighbors, any two of them, being mobile, will be within a nite number of hops of each other, with high probability. If the source knows the current and future position of all nodes, it will be able to estimate exactly which hops to traverse. Consequently, around n1 simultaneous transmission are possible, and all packets need a nite number of hops to reach their destination, so that the capacity remains on the order of n1 . Establishing a bound for the required number of hops is equivalent to establishing how fast the message will spread in the node population if every node forwards it to all its neighbors. Thus, the problem is intimately related to the spread of rumors and the spread of epidemics. Both of these phenomena have been studied extensively [66]. Finally, it was assumed throughout this work that nodes are allowed to coordinate locally. Specically, nodes within the same cell are allowed to schedule consecutive transmissions with no communication overhead. We conjecture that the need for local coordination can be substituted with a random access scheme, for example Aloha, with a reduction in throughput not exceeding a factor of n , for any > 0.
Appendix
Lemma 5 (On fading distributions) Equations (5.3) and (5.4) hold for Nakagami, Ricean, and Rayleigh fading. Proof: For all of these distributions there is clearly an fm > 0 such that (5.4) holds, so we concentrate on establishing (5.3). Let fij be a fading coefcient, and let R2 = fij , where R is the normalized amplitude of the received signal.
108
Nakagami fading: By assumption, the probability density function of R is given as pR (r) = 2 m m 2m1 mr2 ( ) r e (m)
1 2
where = E(R2 ), () is the well-known Gamma function, and m [67]. We then have:
+ +
is the fading gure
F c (x) P [fij > x] = P [R >
x] =
pR (r) dr
m m m r2 re 2 dr = e 2 x .
The inequality will hold for x > x1 , where x1 is sufciently large, since the exponential term dominates the behavior of the integrand. So (5.3) holds with q = large x1 . Ricean fading: The probability density function of R is given as pR (r) = where s2 is proportional to the power of the specular, non-fading component, 2 is proportional to the power of the fading component, and I0 () is the zeroth order modied Bessel function of the rst kind, for which the following holds [68]: ex x > 5. I0 (x) x So we have: F c (x) P [fij > x] = P [R >
+ x
r +s r 2 2 e 2 2 2 2
m 2
and a sufciently
rs I0 ( 2 )
(5.36)
x] =
pR (r) dr
r r2 +s2 + rs e 22 2 2
2 dr rs
+ x
x r r22 e 4 dr e 42 . 2 2
The rst inequality comes from (5.36) and will hold for q=
1 . 42
x > 5. The second inequality
will only hold for sufciently large x. So (5.3) holds for an appropriate choice of x1 , and Rayleigh fading: This is a special case of Nakagami fading, with m = 1.
Chapter 6 Conclusions and Future Work
HIS WORK
establishes theoretical bounds on the performance of wireless ad hoc
networks, i.e., collections of hierarchically equivalent nodes communicating over a
wireless channel, with arbitrary topology and arbitrary trafc requirements. This process of establishing theoretical bounds provides us with excellent intuition on the fundamental limitations of such networks, and highlights good design principles. Therefore, as a logical next step, we also apply the experience we gain in a more practical setting, i.e., the crosslayer design of protocols for such networks. We start by dening and studying capacity regions. These regions describe the set of achievable rate combinations between all source-destination pairs in the network under various transmission strategies, such as variable-rate transmission, single hop or multihop routing, power control and successive interference cancellation. The introduced formulation is very versatile: multihop cellular networks and networks with energy constraints (and possibly delay constraints) can be treated as special cases. In addition, with minor modications, the formulation can handle node mobility, time-varying at fading channels, and various restrictions on the routing of packets. Using this formulation, we study various network topologies to show that multihop routing, spatial reuse and successive interference cancellation all lead to signicant gains. However, gains from power control are signicant only if very limited or no rate adaptation is used. We also determine that fading and node mobility can actually improve the network capacity. At the heart of this formulation lies the concept of the rate matrix. The rate matrix 109
110
CHAPTER 6. CONCLUSIONS AND FUTURE WORK
is an extremely succinct description of what goes on in the network at a given time. In particular, we believe that the rate matrix captures exactly the information required for the evaluation of the network capacity, although this fact does not seem obvious at rst. Having dened the concept of the rate matrix, what separates us from the capacity region is plainly an enumeration of all possible transmission schemes. This aspect of our formulation is less appealing, as the number of different schemes grows exponentially with the number of nodes. However, the problem is fundamentally of exponential complexity. In other words, the exponential complexity is intrinsic to the problem and not to our proposed solution. If we are to make specic assumptions about the channel gain matrix and the receiver function, part of the complexity washes away. (As an extreme case, calculating the capacity region of a network is a trivial task, irrespective of the number of nodes, when (i) the nodes are collocated, (ii) all nodes use the same transmission rate, and (iii) there can be only one transmission at any given time.) Moreover, instead of considering all transmission schemes, we may alternatively consider a typical subset that will capture most of the capacity region. Alternatively, genetic algorithms can be used. Instead of computing explicitly all possible rate matrices, we can start by computing only a few simple ones (for example those corresponding to only a few simultaneous transmissions) and then start combining them, keeping at each time those new rate matrices that have a set of nice properties, for example large entries. This is the subject of future work. Finally, it should be noted that the Shannon capacity region of wireless ad hoc networks remains to be found; in strict information theoretic terms our capacity regions only dene the maximum achievable rates under specic transmission protocols, which may be suboptimal. However, our formulation captures most of the capabilities that a network is expected to have, in the present and also the foreseeable future: multihop multi-path routing, spatial reuse, successive interference cancellation, power control, and variable-rate transmission. In addition, our work may be used as the basis of a more information theoretic approach. Having developed a formulation for calculating the capacity regions of ad hoc networks, it is only natural to ponder how much of the capacity region of a network is actually achieved by practical algorithms. In the second part of the thesis, it is shown that various decisions at different layers of the protocol stack can have a very detrimental effect on the performance. As a result, the gap between what is theoretically possible and what is
111
actually achieved can be very large. Leveraging our intuition from our study of capacity regions, we conclude the second part of this thesis with the introduction of two medium access control protocols, PBOA and PRUA. The rst one also performs power control. Each is integrated with its own queuing discipline, and both are very robust with respect to the choice of the routing protocol. PBOA and PRUA can achieve higher throughputs than CSMA/CA. Equally importantly, they are superior in terms of energy efciency. However, their distributed nature introduces a signicant gap between their performance and the actual capacity of the network. As discussed, our work on capacity regions is necessarily limited to the study of networks with a small number of nodes, because of the numerical complexity of our formulation. To overcome this limitation, it is necessary that concessions be made. Therefore, for the third part of the thesis, we adopt the approach of Gupta and Kumar [36]. The critical idea is to consider random topologies and trafc patterns, to allow the number of nodes to go to innity, and to compute the performance asymptotically. In this manner, statistical averaging is introduced. All results will only hold with high probability, i.e., with probability approaching unity as the number of nodes approaches innity. However this is a small concession to make, given the wealth of results that can be obtained in this manner. Although we borrow the basic idea of Gupta and Kumar, we apply it in a much simplied framework, which allows us to incorporate more detail in our models, and still keep our proofs tractable. As an example, we calculate lower bounds on the achievable throughput of ad hoc networks, as Gupta and Kumar do, but with more general trafc models and possibly in the presence of fading. In addition, our framework easily extends to incorporate node mobility, and in this manner we arrive at a fundamental trade-off between the achievable throughput and packet delays. It is a truism that, in our eld, the ultimate usefulness of a theoretical work is directly related to the effect it has on the design and performance of practical systems. Clearly, knowing the capacity of ad hoc networks allows us to identify how far our protocols are from the optimum performance. Furthermore, since all our derivation of capacity bounds are constructive, the process of determining the bounds provides us with intuition on the fundamental limitations of networks, and highlights good design principles. For example, the theoretical work on capacity regions of Chapter 3 motivated the cross-layer design of
112
CHAPTER 6. CONCLUSIONS AND FUTURE WORK
Chapter 4. This process can be duplicated with the theoretical results for very large networks of Chapter 5. In particular, our work on the tradeoff between packet delay and overall throughput can easily be cast into a practical protocol that performs medium access, routing and queuing, while trading off throughput with delay. The performance of such a protocol would be affected by many practical considerations, such as buffer sizes, the parameters of the mobility pattern, and so on. It would be interesting to investigate how clearly the fundamental throughput-delay tradeoff would manifest itself in practical network designs. This is the subject of future work.
Bibliography
[1] R. E. Kahn, S. A. Gronemeyer, J. Burchel, and R. C. Kunzelman, Advances in packet radio technology, Proc. IEEE, vol. 66, no. 11, pp. 14681496, Nov. 1978. [2] B. Leiner, D. Nielson, and F. Tobagi, Eds., Proc. IEEE, Special Issue on Packet Radio Networks, vol. 75, no. 1, Jan. 1987. [3] C.-K. Toh, Ad Hoc Mobile Wireless Networks: Protocols and Systems, 1st ed. glewood Cliffs, NJ: Prentice Hall, Dec. 2001. [4] B. Bing, R. V. Nee, and V. Hayes, Eds., IEEE Commun. Mag., Special Issue on Wireless Local Area and Home Networks, vol. 39, no. 11, Nov. 2001. [5] C. Perkins, Ed., Ad Hoc Networking, 1st ed. 2000. [6] M. Naghshineh, Ed., IEEE Personal Commun. Mag., Special Issue on Advances in Mobile Ad Hoc Networks, vol. 8, no. 1, Feb. 2001. [7] G. J. Pottie, Wireless sensor networks, in Proc. IEEE Information Theory Workshop, vol. 1, Killarney, Kerry, Ireland, June 1998, pp. 139140. [8] A. Michail and A. Ephremides, Energy efcient routing for connection-oriented trafc in ad-hoc wireless networks, in Proc. 11th IEEE PIMRC, vol. 2, London, UK, Sep. 2000, pp. 762766. [9] J. G. Proakis, E. M. Sozer, J. A. Rice, and M. Stojanovic, Shallow water acoustic networks, IEEE Commun. Mag., vol. 39, no. 11, pp. 15611571, Nov. 2001. 113 Boston, MA: Addison-Wesley, Dec. En-
114
BIBLIOGRAPHY
[10] H. Gharavi, Ed., Proc. IEEE, Special Issue on Sensor Networks and Applications, vol. 91, no. 10, Oct. 1993. [11] Z. J. Haas, M. Gerla, D. B. Johnson, C. E. Perkins, M. B. Pursley, M. Steenstrup, and C.-K. Toh, Eds., IEEE J. Select. Areas Commun., Special Issue on Wireless Ad Hoc Networks, vol. 17, no. 8, Aug. 1999. [12] D. P. Bertsekas and R. Gallager, Data Networks, 2nd ed. Prentice Hall, Dec. 1991. [13] N. Abramson, The ALOHA systemAnother alternative for computer communications, in Proc. AFIPS Fall Joint Computer Conference, vol. 37, Montvale, NJ, Nov. 1970, pp. 281285. [14] L. Kleinrock and F. A. Tobagi, Packet switching in radio channels: Part ICarrier Sense Multiple Access models and their throughput-delay characteristics, IEEE Trans. Commun., vol. 23, no. 12, pp. 14001416, Dec. 1975. [15] F. A. Tobagi and L. Kleinrock, Packet switching in radio channels: Part IIthe hidden terminal problem in carrier sense multiple-access and the busy tone solution. IEEE Trans. Commun., vol. 23, no. 12, pp. 14171433, Dec. 1975. [16] P. Karn, MACA: A new channel access method for packet radio, in Proc. CNC, vol. 1, London, Ontario, Canada, Sep. 1990, pp. 134140. [17] V. Bharghavan, A. Demers, S. Shenkar, and L. Zhang, MACAW: A media access protocol for wireless LANs, in Proc. ACM SIGCOMM, London, UK, Aug. 1994, pp. 212225. [18] C. L. Fullmer and J. J. Garcia-Luna-Aceves, Floor acquisition multiple access (FAMA) for packet-radio networks, in Proc. ACM SIGCOMM, Cambridge, MA, Aug. 1995, pp. 262273. [19] , Solutions to hidden terminal problems in wireless networks, in Proc. ACM SIGCOMM, Cannes, France, Sep. 1997, pp. 3949. Englewood Cliffs, NJ:
BIBLIOGRAPHY
115
[20] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specication, IEEE Std. 802.11. [21] B. P. Crow, I. Widjaja, J. G. Kim, and P. T. Sakai, IEEE 802.11 Wireless Local Area Networks, IEEE Commun. Mag., vol. 35, no. 9, pp. 116126, Sep. 1997. [22] I. Cidon and M. Sidi, Distributed assignment algorithms for multihop packet radio networks, IEEE Trans. Comput., vol. 38, no. 10, pp. 13531361, Oct. 1989. [23] A. Ephremides and T. Truong, Scheduling broadcasts in multihop radio networks, IEEE Trans. Commun., vol. 38, no. 4, pp. 456460, Apr. 1990. [24] C. Zhu and M. Corson, A ve phase reservation protocol (FPRP) for mobile ad hoc networks, in Proc. IEEE INFOCOM, vol. 1, San Fransisco, CA, Apr. 1998, pp. 322 331. [25] M. K. Marina, G. D. Kondylis, and U. C. Kozat, Rbrp: A robust broadcast reservation protocol for mobile ad hoc networks, in Proc. IEEE ICC, vol. 3, Helsinki, Finland, June 2001, pp. 878885. [26] I. Chlamtac, A. Farag` , A. D. Myers, V. R. Syrotiuk, and G. Z ruba, ADAPT: A o a dynamically self-adjusting media access control protocol for ad hoc networks, in Proc. IEEE GLOBECOM, vol. 1, Rio De Janeiro, Brazil, Dec. 1999, pp. 1115. [27] Z. Tang and J. J. Garcia-Luna-Aceves, A protocol for topology-dependent transmission scheduling in wireless networks, in Proc. IEEE WCNC, vol. 3, New Orleans, LA, Sep. 1999, pp. 13331337. [28] S. S. Lam, Packet broadcast networksa performance analysis of the R-ALOHA protocol, IEEE Trans. Comput., vol. 29, no. 7, pp. 596603, July 1980. [29] D. J. Goodman and S. X. Wei, Efciency of Packet Reservation Multiple Access, IEEE Trans. Veh. Technol., vol. 40, no. 1, pp. 170176, 1991. [30] C. E. Perkins and P. Bhagwat, Highly dynamic destination-sequenced distancevector routing (dsdv) for mobile computers, in Proc. ACM SIGCOMM, vol. 1, London, UK, Aug. 1994, pp. 234244.
116
BIBLIOGRAPHY
[31] S. Ramanathan and M. Steenstrup, A survey of routing techniques for mobile communication networks, ACM/Baltzer Mobile Networks and Appl., vol. 1, no. 2, pp. 89104, Oct. 1996. [32] C. E. Perkins and E. M. Royer, Ad hoc on-demand distance vector routing, in Proc. IEEE WMCSA, vol. 1, New Orleans, LA, Feb. 1999, pp. 90100. [33] E. M. Royer and C.-K. Toh, A review of current routing protocols for ad-hoc mobile wireless networks, IEEE Personal Commun. Mag., vol. 6, no. 2, pp. 4655, Apr. 1999. [34] L. Kleinrock and J. Silvester, Spatial reuse in multihop packet radio networks, Proc. IEEE, vol. 75, no. 1, pp. 156167, Jan. 1987. [35] T. J. Shepard, A channel access scheme for large dense packet radio networks, in Proc. ACM SIGCOMM, Stanford, CA, Aug. 1996, pp. 219230. [36] P. Gupta and P. R. Kumar, The capacity of wireless networks, IEEE Trans. Inform. Theory, vol. 46, no. 2, pp. 388404, Mar. 2000. [37] , Internets in the sky: The capacity of three dimensional wireless networks, Communications in Information and Systems, vol. 1, no. 1, pp. 3355, Jan. 2001. [38] M. Grossglauser and D. Tse, Mobility increases the capacity of ad-hoc wireless networks, in Proc. IEEE INFOCOM, vol. 3, Anchorage, AL, Apr. 2001, pp. 13601369. [39] DIMACS Workshop on Network Information Theory, Piscataway, NJ, Mar. 2003. [40] G. J. Foschini and Z. Miljanic, A simple distributed autonomous power control algorithm and its convergence, IEEE Trans. Veh. Tech., vol. 42, no. 4, pp. 641646, Apr. 1993. [41] D. Mitra, An asynchronous distributed algorithm for power control in cellular radio systems, in Proc. 5th WINLAB Workshop, vol. 1, Rutgers University, New Branswick, NJ, Oct. 1993, pp. 249257.
BIBLIOGRAPHY
117
[42] N. Bambos, Toward power-sensitive network architectures in wireless communications: Concepts, issues, and design aspects, IEEE Personal Commun. Mag., vol. 5, no. 3, pp. 5059, June 1998. [43] N. Bambos, S. C. Chen, and G. J. Pottie, Channel access algorithms with active link protection for wireless communication networks with power control, IEEE/ACM Trans. Networking, vol. 46, no. 2, pp. 388404, Mar. 2000. [44] M. Andersin, Z. Rosberg, and J. Zander, Soft and safe admission control in cellular networks, IEEE/ACM Trans. Networking, vol. 5, no. 2, pp. 255265, Apr. 1997. [45] R. Diestel, Graph Theory, 2nd ed. New York, NY: Springer Verlag, Jan. 2000.
[46] J. L. Sobrinho and A. S. Krishnakumar, Quality-of-Service in ad hoc carrier sense multiple access wireless networks, IEEE J. Select. Areas Commun., vol. 17, no. 8, pp. 13531368, Aug. 1999. [47] S. Narayanaswamy, V. Kawadia, R. S. Sreenivas, and P. R. Kumar, Power control in ad hoc networks: Theory, architecture, algorithm and implementation of the compow protocol, in Proc. European Wireless, Florence, Italy, Feb. 2002, pp. 156162. [48] A. J. Goldsmith and S.-G. Chua, Variable-rate variable-power MQAM for fading channels, IEEE Trans. Commun., vol. 45, no. 10, pp. 12181230, Oct. 1997. [49] M. V. Eyobo lu and J. G. D. Forney, Trellis precoding: Combined coding, precodg ing and shaping for intersymbol interference channels, IEEE Trans. Inform. Theory, vol. 38, no. 2, pp. 301314, 1992. [50] R. Gallager, Energy limited channels: Coding, multiaccess and spread spectrum, in Proc. Conf. Inform. Sci. Syst., Princeton, NJ, Mar. 1988, p. 372. [51] S. Verd , On channel capacity per unit cost, IEEE Trans. Inform. Theory, vol. 36, u no. 5, pp. 10191030, Sep. 1990. [52] J. P. Monks, V. Bharghavan, and W. W. Hwu, A power controlled multiple access protocol for wireless packet networks, in Proc. IEEE INFOCOM, vol. 1, Anchorage, AL, Apr. 2001, pp. 111.
118
BIBLIOGRAPHY
[53] J. Broch, D. B. Johnson, and D. A. Maltz, The dynamic source routing protocol for mobile ad hoc networks, Internet Draft draft-ietf-manet-dsr-01.txt, Dec. 1998. [54] S.-J. Lee and M. Gerla, Dynamic load-aware routing in ad hoc networks, in Proc. IEEE ICC, vol. 10, Helsinki, Finland, June 2001, pp. 32063210. [55] M. R. Pearlman, Z. J. Haas, P. Sholander, and S. S. Tabrizi, On the impact of alternate path routing for load balancing in mobile ad hoc networks, in Proc. ACM MobiHOC, vol. 1, Boston, MA, Aug. 2000, pp. 310. [56] L. Kleinrock, Queuing Systems, Volume I: Theory, 1st ed. and Sons, Jan. 1975. [57] M. Shreedhar and G. Varghese, Efcient fair queuing using decit round robin, IEEE/ACM Trans. Networking, vol. 4, no. 3, pp. 375385, June 1996. [58] L. Tassiulas and A. Ephremides, Jointly optimal routing and scheduling in packet radio networks, IEEE Trans. Inform. Theory, vol. 38, no. 1, pp. 165168, 1992. [59] T. Ozugur, M. Naghshineh, P. Kermani, and J. A. Copeland, Fair media access for wireless LANs, in Proc. IEEE GLOBECOM, vol. 1B, Rio De Janeiro, Brazil, Dec. 1999, pp. 570579. [60] T. ElBatt and A. Ephremides, Joint scheduling and power control for wireless ad hoc networks, in Proc. IEEE INFOCOM, vol. 2, New York, NY, June 2002, pp. 976985. [61] Z. J. Haas, J. Deng, and S. Tabrizi, Collision-free medium access control scheme for ad-hoc networks, in Proc. IEEE MILCOM, vol. 1, Atlantic City, NJ, Nov. 1999, pp. 276280. [62] R. Motwani and P. Raghavan, Randomized Algorithms, 1st ed. Cambridge University Press, 1995. [63] W. Feller, An Introduction to Probability Theory and its Applications, 3rd ed. York, NY: John Wiley and Sons, 1975, vol. I. [64] N. Bansal and Z. Liu, Capacity, delay and mobility in wireless ad-hoc networks, in Proc. IEEE INFOCOM, San Francisco, CA, April 2003. New Cambridge, UK: New York: John Wiley
BIBLIOGRAPHY
119
[65] D. Shah, J. Mammen, A. E. Gamal, and B. Prabhakar, Throughput-delay trade-off in wireless networks, submitted. [66] D. J. Daley and J. Gani, Epidemic Modeling: An Introduction, 1st ed. University Press, May 2001. [67] J. G. Proakis, Digital Communications, 3rd ed. New York, NY: McGraw Hill, 1995. [68] M. Abramowitz and I. A. Stegun, Eds., Handbook of Mathematical Functions, 1st ed. New York, NY: Dover Publications, 1965. [69] S. Toumpis and A. J. Goldsmith, Capacity regions for wireless ad hoc networks, in Proc. IEEE ICC, vol. 5, New York, NY, Apr. 2002, pp. 31683173. [70] , Ad hoc network capacity, in Proc. Asilomar Conference on Signals, Systems and Computers, vol. 2, Pacic Grove, CA, Oct. 2000, pp. 12651269. [71] , Some capacity results for ad-hoc networks, in Proc. Annual Allerton Conf. on Communications, Control, and Computing, Allerton, IL, Oct. 2000, pp. 775784. [72] , Performance, optimization, and cross-layer design of media access control protocols for wireless ad hoc networks, in Proc. IEEE ICC, Anchorage, AK, May 2003. [73] , Capacity regions for ad hoc networks, to appear in IEEE Trans. Wireless Commun. Cambridge

10 1 1 81

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

10 1 1 81

Hochgeladen von

Copyright:

Verfügbare Formate

CAPACITY AND CROSS-LAYER DESIGN OF WIRELESS AD HOC NETWORKS

Stavros Toumpis July 2003

c Copyright by Stavros Toumpis 2003 All Rights Reserved

Andrea J. Goldsmith (Principal Adviser)

Approved for the University Committee on Graduate Studies.

of this thesis is on the establishment of theoretical lower and upper

bounds on the trafc-carrying capabilities of wireless ad hoc networks. This theo-

(log n) 2 , provided that nodes are

S TANFORD I had the opportunity to associate with a number

of truly remarkable people.

For Ivan Tsikouras, 1975-1993

Uniform Capacity of Canonical Topologies . . . . . . . . . . . . . . . . . 27 Discrete-Rate Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . 28 xi

3.6 3.7 3.8

Energy Constrained Networks . . . . . . . . . . . . . . . . . . . . . . . . 40 3.9.1 3.9.2 3.9.3

Progressive Ramp Up Algorithm (PRUA) . . . . . . . . . . . . . . . . . . 65 4.8.1 4.8.2 4.8.3

Fundamental Limits of Distributed Protocols . . . . . . . . . . . . . . . . . 68

4.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5 Large Wireless Networks 5.1 72

Immobile Nodes with no Fading . . . . . . . . . . . . . . . . . . . . . . . 80 5.3.1 5.3.2 5.3.3 5.3.4

A Polynomially Increasing Delay Constraint . . . . . . . . . . . . . . . . . 100 5.6.1 5.6.2 5.6.3

Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 105 109 113

6 Conclusions and Future Work Bibliography

the example ad hoc network of Section 3.3, with = 0 dB.

IRELESS AD HOC NETWORKS

consist of mobile nodes communicating over a

Figure 1.1: The ad hoc topology.

1.2. CURRENT RESEARCH ON WIRELESS AD HOC NETWORKS

Internet, PSTN, etc.

1.2 Current Research on Wireless Ad Hoc Networks

Modern communication networks are systems of enormous com-

1.2. CURRENT RESEARCH ON WIRELESS AD HOC NETWORKS

TRANSPORT NETWORK PROTOCOLS NETWORK

DATA LINK CONTROL

1.2. CURRENT RESEARCH ON WIRELESS AD HOC NETWORKS

Chapter 2 System Model

ESEARCH IN WIRELINE NETWORKS

has been facilitated by the existence of a

Throughout the paper, terms being dened are set in boldface.

CHAPTER 2. SYSTEM MODEL

Chapter 3 Capacity Regions for Wireless Ad Hoc Networks

we dene and study capacity regions for wireless ad hoc networks

with an arbitrary number of nodes and an arbitrary topology. These multidimensional

3.2. RATE MATRICES AND CAPACITY REGIONS

3.2 Rate Matrices and Capacity Regions

3.2.1 Transmission Schemes and Time Division Schedules

CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS

ai = 1. We say that the network is using

ai Si , and we refer to the fractions ai , i = 1, . . . , k

3.2. RATE MATRICES AND CAPACITY REGIONS

3.2.2 Rate Matrices

CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS

3.2.3 Ad Hoc Network Capacity Regions and Uniform Capacity

3.2. RATE MATRICES AND CAPACITY REGIONS

CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS

3.3. CAPACITY REGIONS FOR VARIOUS TRANSMISSION PROTOCOLS

3.3 Capacity Regions for Various Transmission Protocols

CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS

3.3.1 Single Hop Routing, No Spatial Reuse

3.3. CAPACITY REGIONS FOR VARIOUS TRANSMISSION PROTOCOLS

1.5 r12 (Mbps)

3.3.2 Multihop Routing, No Spatial Reuse

b Ri , i = 1, . . . , N b . Determining these matrices is straightforward by applying (2.1) and

CHAPTER 3. CAPACITY REGIONS FOR WIRELESS AD HOC NETWORKS