ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS

No. of Pages 15, Model 3+

Ad Hoc Networks xxx (2006) xxxxxx www.elsevier.com/locate/adhoc

2 3 4 5 6 7

Rajarshi Gupta *, John Musacchio, Jean Walrand

Department of Electrical Engineering and Computer Science, University of California, 620 Country Brook Loop, San Ramon, CA 94583, United States Received 30 December 2004; received in revised form 26 January 2006; accepted 28 January 2006

8 Abstract 9 10 11 12 13 14 15 16 17

18 Keywords: Ad-hoc networks; Graph theory; QoS 19

UN CO RR

EC

The capacity of an arbitrary ad-hoc network is dicult to estimate due to interference between the links. We use a conict graph that models this interference relationship to determine if a set of ow rates can be accommodated. Using the cliques (complete subgraphs) of the conict graph, we derive constraints that are sucient for a set of ow rates to be feasible, yet are guaranteed to be within a constant bound of the optimal. We also compute an alternate set of sucient constraints that can be easily derived from the rows of the matrix representation of the conict graph. These two sets of constraints are particularly useful because their construction and verication may be distributed across the nodes of a network. We also extend the ad-hoc network model to incorporate variations in the interference range, and obstructions in the network. 2006 Published by Elsevier B.V.

TE

20 1. Introduction 21 22 23 24 25 26 27 28

Ad-hoc networks are becoming commonplace in todays world as wireless enabled devices proliferate. This leads to newer applications based on these networks, and hence the need to support quality of service. For a broader discussion and motivating examples, see Chapter 1 in [1]. A vital problem faced by researchers in this eld is to determine the capacity of an arbitrary ad-hoc
This work was supported by the Defense Advanced Research Project Agency under Grant N66001-00-C-8062. * Corresponding author. Tel.: +1 925 327 0507. E-mail addresses: [email protected] (R. Gupta), [email protected] (J. Musacchio), [email protected] (J. Walrand). 1570-8705/$ - see front matter 2006 Published by Elsevier B.V. doi:10.1016/j.adhoc.2006.01.003
q

network. This is dicult because neighboring links using the same channel interfere, and the interference relationships between all of the links in a network can be quite complex. Several researchers interested in the capacity of ad-hoc networks have modelled the ad-hoc network using randomized models, and evaluated asymptotic bounds on the capacity. Other work has addressed the question of whether a given ow vector is feasible on a particular ad-hoc network, where feasible means that a global scheduler with access to all the information in the network could nd a link scheduling policy that would achieve the desired rates. In this work, we are also interested in methods for determining whether a ow vector is feasible, but we are particularly interested in methods that

PR

OO F

Sucient rate constraints for QoS ows in ad-hoc networks

29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44

ADHOC 158 18 February 2006; Disk Used 2

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

2 B

UN CO RR

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81

are suitable for distributed control in an ad-hoc network. As in [2,3], we make use of a conict graph that models the interference relationships between the dierent links of a network. Every link in the connectivity graph G = (V, E) is represented by a node in the conict graph CG = (VC, EC). Two nodes in CG are connected by an edge if the nodes correspond to links in G that interfere. In Fig. 1, we show an example of a connectivity graph in which the interference between links is marked using dotted lines. The corresponding conict graph is shown on the right. The authors of [2,3] show that a set of necessary and sucient conditions to whether a set of ows is feasible is found by a computationally expensive process of nding all of the independent sets of the conict graph, and then writing constraints in terms of the independent sets. (We review the details of the independent set method, as we call it, in Section 3.2.) Because the independent set constraints are computationally expensive and require global information, they are not suitable to be used in a distributed scheme. We therefore look to nd a different set of conditions that can be computed in a distributed way and that are at least sucient, though perhaps not necessary, for a ow vector to be feasible. One such set of constraints we refer to are the row constraints, because they are derived by using the rows of the matrix representation of the conict graph. While they are sucient conditions that are relatively easy to compute, they are much more restrictive than is necessary in many cases. This motivates us to develop a dierent set of sufcient conditions using cliques (complete subgraphs) of the conict graph. While previous work

interference

5 E Connectivity Graph

Conflict Graph

Fig. 1. Example of a connectivity graph with its conict graph.

[3] has used cliques to nd necessary conditions for a set of ow rates to be feasible, we nd sucient conditions. We refer to our sucient clique constraints as scaled clique constraints because they are constructed by modifying the necessary clique constraints by a constant scaling factor. Unlike the row constraints, the scaled clique constraints are within a constant bounded factor of being necessary; a factor that is independent of the structure of the network, or the ows on it. One potential drawback of a scheme based on cliques is that the number of cliques in a graph can grow exponentially with the number of nodes. However, we discuss a technique for identifying approximate cliques that can be implemented distributedly, and that grows only polynomially with the number of nodes. We discuss how the suciency of the clique constraints is maintained when using this approximation. At rst, we model the nodes having a constant interference range, and utilize the resulting unit disk structure (dened in Section 5.4) of the graph. However, to consider more realistic ad-hoc network scenarios, we augment the network model to allow for variances in the interference range, and evaluate its eect on the scaling factor. We also incorporate obstructions in the network, and extend the proof of suciency to include this. Finally, we attempt to validate our theoretical constraint-based approach by simulating various ad-hoc networks, and comparing the achieved rates in the simulation to the theoretical limits predicted by our model. The results are encouraging at rst sight, as the networks appear to satisfy our constraint-based approach. However, a deeper study reveals that in some cases, large ineciencies in the underlying MAC protocol may be the primary cause for the limited throughput achieved; our constraints are satised as a corollary of this eect. We conclude that the constraint-based framework presented here becomes universally practicable only when used in conjunction with an ecient MAC protocol. This paper does not try to present a new architecture to achieve QoS in ad-hoc networks. What we propose is a theoretical framework that allows us to answer questions about the feasibility of a specic set of ows on an arbitrary ad-hoc network. However, this framework may indeed be used as tools to develop new and improved protocols for QoS routing in ad-hoc networks, as we discuss in Section 2.1.

82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133

EC

TE

PR

OO F

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

145 2. Related work 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 Many researchers have looked at modelling the capacity of ad-hoc networks. In [5], Gupta and Kumar study how the capacity of an ad-hoc network scales asymptotically with the number of nodes n in the network. The authors study the problem under two dierent models of interference: the protocol model and the physical model. In the protocol model, a receiver can successfully receive a senders transmission if the sender is geographically closer, by some margin, than any other node that is actively transmitting. In contrast, the physical model models the transmissions of other nodes as noise, and assumes a receiver successfully receives a senders transmission if the signal to interference (SIR) power ratio is above a threshold. The authors show that under both the protocol and p physical models, the maximum capacity is H1= n per node, assuming optimal node placement in a disk of unit area. They also show that under random node placement, and under the protocol p interference model, the capacity is H1= n log n. In [6], the authors extend this result to show that randomly scattered p nodes can in fact achieve the same H1= n per-node transmission rate of arbitrarily located nodes. Other researchers have also extended the work of [5] by considering the eects of node mobility [7], and throughput-delay trade-os [8]. Li et al. analyze the capacity of specic network topologies, and run packet level simulations of both the specic topologies and of random graphs [9]. With a packet level simulator, the authors show that the maximum throughput achieved in the simulation, where the link schedule is determined by a distributed MAC protocol (802.11), is somewhat less than that predicted by the analytical capacity model which assumes an ideal schedule. For example, for a network consisting of a chain of nodes relaying a single ow, analysis suggests that a rate of up to 1 4

TE

2.1. Extensions The theoretical framework presented in this paper has already proved to be vital in our study of practical algorithms for quality of service in adhoc networks. Our understanding of cliques, and the crucial role they play in ad-hoc networks, has been central to the development of a suite of algorithms for the routing and MAC layers. We have utilized the clique-based framework to propose realistic admission control and QoS routing mechanisms for ad-hoc networks. In [11], we proposed a distributed ad-hoc shortest widest path (ASWP) routing algorithm, which transforms the well-known shortest widest path paradigm to the ad-hoc domain by taking interference into account. We further proposed IQRouting [12] a sourcebased heuristic mechanism that is able to select QoS routed paths in a dynamic manner, using only localized state information. Each of these algorithms relies on the computation of available bandwidth (Section 7.2), and utilizes cliques as the central unit of QoS. Even after proving the existence of a feasible schedule, a more important practical problem is to nd a mechanism that can achieve such a schedule. While a distributed solution may be too dicult to achieve, it is fair to hope for approximate mechanisms which are able to approach the feasible network throughput to within a bounded factor of

PR

OO F

134 135 136 137 138 139 140 141 142 143 144

The rest of the paper is organized as follows. In Section 2, we describe the related work in the eld, while Section 3 summarizes the network model that we base our results upon. Sections 4 and 5 are used to present the row and clique constraints, and prove suciency and bounds on these. In Section 6, we compare these two sets of constraints. The application of these constraints in distributed algorithms are described in Section 7. Simulation results are presented in Section 8, before we conclude the paper.

the channel capacity should be feasible, but with nodes using the 802.11 MAC, a rate of only 1 is 7 achieved. Luo et al. [2] as well as Jain et al. [3] study the problem of nding rate constraints on a set of ows in an ad-hoc network, modelling what would be possible if there were a global scheduler. Both works model the interference between links in an ad-hoc network using conict graphs and nd capacity constraints by nding the independent sets of the conict graph. The concepts of conict graph and independent sets are discussed in more detail in the next section. In [10], Kodialam and Nandagopal model the routing and scheduling of ows in an adhoc network as a graph edge coloring problem, and nd necessary and sucient conditions for the achievability of a rate vector. The caveat is that this model only considers conicts between links incident at the same node, and does not take into account interference due to all other neighboring links.

184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233

UN CO RR

EC

ADHOC 158 18 February 2006; Disk Used 4

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

234 235 236 237

the optimal. We suggested a novel backo mechanism in [13] that takes a step in this direction, achieving a fair distribution of resources between wireless nodes sharing the medium.

238 3. Modelling the ad-hoc network 239 3.1. Determining conict graph 240 241 242 243 244 245 246 247 248 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 We consider a wireless ad-hoc network with N stations. Each station is equipped with a radio having communications range q, and a potentially larger interference radius x. Our model of interference is similar to that of the protocol model introduced in [5]. A transmission from station i to station j is successful if both of the following conditions are satised: d ij < q and d kj > x 1

for every other station k that is simultaneously transmitting. Here dij denotes the distance between i and j. The connectivity graph G is a directed graph whose vertices correspond to wireless stations and the edges correspond to the wireless links. There is a directed link from vertex i to vertex j i dij < q. The nodes of the conict graph represent links in the connectivity graph. A pair of nodes, lij and lkl in the conict graph are connected by an edge if they cannot have simultaneous transmissions according to the protocol interference model. This is similar to the model of the conict graph as presented in [2,3]. To avoid confusion in the rest of the paper, we adopt the convention of using the prex CG (e.g. CG-node, CG-edge) when referring to the conict graph. Additionally, a wireless device participating in the ad-hoc protocol is sometimes referred to as an ad-hoc station. We envision the following technique for computing the conict graph. We assume each station knows its position (e.g. using GPS) and disseminates its position information to other stations in the local neighborhood. Each station then geometrically computes which stations are within an interference radius; we call such stations interfering neighbors. However, the transmission and interference footprints are not perfect circles in reality, due to factors such as obstacles, multi-path fading, etc. A better scheme would use measurements, as well as any available position information from GPS, to deter-

UN CO RR

mine which stations interfere. In Sections 5.5 and 5.6, we attempt to augment our model to incorporate more realistic interference patterns. In cases where the networks MAC protocol uses RTS/CTS (request to send/clear to send) or acknowledgements, such as 802.11 [14], one might choose to use stricter rules for identifying conicting links than the conditions presented in (1). For a successful transmission to occur, these stricter conditions require that dkj > x and dki > x, for every other station k that is simultaneously transmitting or receiving. This is because a receiving station will be sending acknowledgements or will have sent a CTS message that silences other stations in the vicinity. An alternate approach to constructing a conict graph, is to compare the geometric centers of each link, i.e. the mid-point of a line segment connecting receiver and transmitter. This is in contrast to comparing distances between the stations themselves [15]. A sucient condition for a pair of links, lij and lkl not to conict would be jcij cklj > (x + q), where cij is the position of the geometric center of link lij. Note that this is sucient whether one uses the conditions (1) or the stricter RTS/CTS rules, for a link transmission to be successful. One may then construct a conict graph by assuming that pairs of links that do not satisfy these link-center conditions conict. However, this approximation would lead to some pairs of links being modelled as conicting, even though they do not conict in reality, and thus would result in a conservative view of network capacity. It is useful to note that the discussions about row and clique constraints presented in this paper are not dependent on the way the CG is calculated. A more accurate CG representation is still amenable to the same analysis as presented here. 3.2. Independent set constraints One can nd a set of necessary and sucient conditions for a proposed set of ow rates to be feasible, by looking at the independent sets in CG [2,3]. An independent set in the CG is a set of CG-nodes that have no edges between them. The idea is to identify all of the maximal independent sets of CG, calling them I1 ; I2 ; . . . ; Iz . Then, the independent set constraints say that a set of ow rates F1 ; F2 ; . . . ; Fn is feasible i there exists non-negative weights k1, k2, . . ., kz such that Pz i1 ki 6 1 and

284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333

EC

TE

PR

OO F

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

335 336 337 338 339 340 341 342 343 344

Fj 6 C

X
i:j2Ii

ki

8j 2; f1; . . . ; ng

345 4. Row constraints 346 347 348 349 350 351 352 We describe here one set of sucient conditions for a set of ow rates to be feasible. We represent a set of ow rates as the vector F of size n 1, where n is the number of links in the network and Fi is simply the ow rate assigned to link i. We also make use of a conict graph incidence matrix M dened as follows: 8 > < 1 if links j and k are connected by Mjk > : 0 an edge, otherwise.

355 354 356

TE

The row constraints may be evaluated in a local- 404 ized and distributed manner by evaluating 405 Mi Fi 6 Ci
i

PR

OO F
8i;
i
A D X C

where C is the capacity of the channel. Unfortunately, computing all of the independent sets in a conict graph is exponential in the number of CG-nodes [3], so using this method in a large CG is not practical. Furthermore, the method requires global information about the entire graph. For these reasons, we look for methods that require less computation, and that can be done with local information at each CG-node.

replacing i has a unique color. Further, if CG-nodes i and j were adjacent in CG, all the components of their replacement cliques are adjacent in CGf, and hence have unique colors. Consequently, by letting each color correspond to a slot, we can ensure that CG-nodes i and j ate scheduled for appropriate durations, which are disjoint. Next we observe that MF 6 C implies MK < T, where T is a n 1 vector, with all entries equal to T. This in turn implies that each CGf-node has strictly less than T neighbors. We may now color the graph with the greedy coloring algorithm that follows. Label the CGf-nodes with indices {1, . . ., N} and begin coloring the nodes in increasing order of index, by assigning each CGfnode a color with index in {1, . . ., T}. For CGf-node i, we assign it the lowest indexed color not already assigned to a neighbor. We can always nd such a color with index in {1, . . ., T}, simply because CGfnode i has less than T neighbors. Thus we have a valid coloring for CGf, which implies the existence of a feasible schedule for the ow vector F. h

382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403

4 407 408 409 410 411 412 413 414 415 416 417 418 419 420

357 Theorem 1. A set of ow rate assignments F has a 358 feasible schedule if 359 361 MF 6 C. 3

367 368 369 370 371 372 373 374 375 376 377 378 379 380 381

Proof. We assume that the entries of the ow rate vector F are rational multiples of each other. Let T be the smallest integer such that ow rates Fi are integer multiples of C for all i 2 {1, . . ., n}, and T let K be the vector of integers such that Fi Ki C . We will construct a periodic schedule T with period T. We begin by transforming the conict graph CG by replacing a CG-node i by a clique consisting of Ki nodes. Each of the new nodes in the replacement clique needs to be connected to every neighbor of the replaced CG-node. Let this transformed graph be called CGf. Observe that a coloring of CGf implies a schedule for CG. This is because each node in the clique

UN CO RR

362 363 364 365 366

where C is a n 1 vector, with all entries equal to the channel capacity C. We refer to expression (3) as the row constraints because each row of the conict graph matrix M is used to construct a linear inequality on the values of F1 ; F2 ; . . . ; Fn .

where F and C are the ow vector and capacity vector representing only the neighbors of link i. Since all the non-zero elements of the ith row of M lies in the interference neighborhood of CG-node i, we only need to consider those relevant entries of F and C. Each CG-node only needs local information from all its neighbors to check the validity of these constraints. While the row constraints are sucient, they are not necessary constraints. Indeed, in some examples they can be much more restrictive than is necessary. Fig. 2 shows a conict graph that illustrates such an example. The row constraints imply that

EC

Fig. 2. Row constraints star graph.

ADHOC 158 18 February 2006; Disk Used 6

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

421 423 FA FB FC FD FX 6 C . 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438

QF 6 C. 439 5. Clique constraints 440 5.1. Cliques: denition and background 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 We begin with a few denitions well-known in graph theory. Consider a bi-directional graph with nodes and edges. An induced subgraph is a subset of the nodes together with any edges whose endpoints are both in this subset. An induced subgraph that is a complete graph is called a clique. A maximal clique of a graph is a clique such that it is not contained in any other clique. In the conict graph in Fig. 1, ABC, ACD and ADE are all maximal cliques. A clique in a conict graph is closely related to the capacity of ad-hoc networks. CG-nodes that form a clique are all connected to each other consequently only one CG-node in a clique may be active at once. There are two main reasons to utilize cliques in place of independent sets. First, cliques in a CG are inherently local structures and therefore amenable to localized algorithms. Second, as we describe in Section 7.1, we can approximate the maximal cliques around a link in a computationally simple and distributed fashion.

TE

Consider the conict graph as shown in Fig. 1. Let the allocated ow on each CG-node be denoted by FA ; FB etc. Then, the clique constraints QF 6 C are: FA FB FC 6 C ; FA FC FD 6 C ; FA FD F E 6 C .

PR

However, expression (5) is much stronger than what is necessary to be feasible. For example, one could achieve the rates FA FB FC FD C if CG-node X is switched o. Yet, if we used expression (5) as our guide (with FX 0), we would have to set all the rates to C/4. Thus, in this case, obeying the row constraints would lead us to rates that are only 1/4 of actually feasible rates. In theory, the row constrained solution could be arbitrarily far from optimal, as in the case if the star had many rays instead of 4. Because the row constraints could possibly be overly conservative, we are motivated to develop a dierent set of constraints using cliques. We describe the method in the next section.

order q n. Here, q is the number of maximal cliques that this link i belongs to, and n is the total number of links. In this matrix, & 1; if links i and k 2 clique j; Qijk 0; if links i or k 62 clique j. Note that this matrix Q does not include information about the entire network it covers only the interference neighborhood of link i. The union of the clique matrices across all the links gives the global clique matrix Q. Since a network must satisfy the capacity constraints for all cliques, we can write the clique constraints in a matrix form. As in Section 4, we denote the ow vector F and the capacity vector C. Hence we have,
i

468 469 470 472

473 474 475 476 477 478 479 480 481 482 483 6 485 486 487 488 489

OO F

491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512

EC

5.3. Insuciency of clique constraints The clique constraints provide a necessary condition for a realizable schedule to exist, since there cannot be a feasible schedule over links that form a violated clique constraint. One might hope that these constraints would also be sucient conditions for a realizable schedule. Unfortunately, that is only true for a special sub-class of graphs called Perfect Graphs [16]. Perfect graphs are those whose chromatic number (least number of colors required to color the graph, such that every adjacent node has a separate color) and clique number (size of the largest clique) are equal for all induced subgraphs. As noted in [3], the simplest example of an imperfect graph where the clique constraints are insucient is the conict graph shaped like a pentagon, shown in Fig. 3. Although the clique constraints suggest a valid ow allocation of 0.5C on each link, in reality only 0.4C on each link is achievable since at most two out of the ve CG-nodes may be active simultaneously.

463 5.2. Specifying clique constraints

464 Assume that each CG-node (i.e., link in the con465 nectivity graph) is aware of all the maximal cliques 466 that it belongs to. This information may be 467 described by an incidence matrix Qi , which is of

UN CO RR

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

Conflict Graph

In [4], the authors dene the imperfection ratio imp(G) of a transformed weighted graph (e.g. CGf) as the supremum of the ratio between its chromatic number and its clique number. They further show bounds on imp(G) if the graph belongs to the class of UDG. For a UDG,

550 551 552 553 554 555 556 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575

Fig. 3. A pentagon conict graph.

513 5.4. Suciency using scaled clique constraints 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 In Section 4, we proved the suciency of row constraints for a schedule to exist. In this section, we demonstrate another sucient condition based on clique constraints, scaled by 0.46. When modelling a link by its mid-point (Section 3.1), the resulting CG has an unit disk graph structure. A graph is said to be a unit disk graph (UDG) [17] when there is an edge between two nodes if and only if their Euclidean distance is at most 1 (or a constant value x). We then use properties of UDG to prove the following theorem. Theorem 2. When the conict graph is modelled as a UDG, a set of ow rate assignments F has a feasible schedule if QF 6 C 0:46. 7

EC

TE

Consider the scaled clique constraints of expression (7): In addition to being sucient, they are within a bounded factor of the necessary conditions of expression (6). Consequently, we are assured that ow vectors satisfying the scaled clique constraints are no further than a factor of 0.46 from an optimally feasible ow vector. It is also worth noting that the clique constraints may be evaluated in a distributed fashion by checking Qi Fi 6 Ci 0:46 8i;
i i

PR

We sketch the proof of this result in Section 5.6. To complete our proof, we observe that QF 6 C 0:46 implies that QK 6 0:46T. This in turn implies that j(CGf) 6 0.46T, since the clique number j(CGf) is simply the largest element of QK. Now applying expression (8), we have vCGf 6 1 0:46 jCGf 6 T . Thus we have a sufcient condition for the existence of a feasible schedule. h

OO F
f

Assume all link capacity = C

impCG sup

vCGf 1 . 6 2:155 % jCGf 0:46

9 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598

Proof. Following in the lines of the proof of Theorem 1, we impose an integer ow rate vector K by assuming a T-periodic slotted time schedule, where Fi Ki C . Next, we transform the conict graph T CG by replacing a CG-node i by a clique consisting of Ki nodes. Let this transformed graph be called CGf. As we observed in the proof of Theorem 1, a coloring of CGf implies a schedule for CG. Denote v(CGf) as the chromatic number of CGf, and let j(CGf) be the clique number of the graph. Then it is well known that v(CGf) P j(CGf), since we at least need to use a different color for every member of the largest clique. We would have a feasible schedule for CG if we could color CGf using at most T colors (i.e., v(CGf) 6 T). This would ensure that all CG-nodes in a clique are scheduled for disjoint slots, yet the number of available slots in the periodic schedule is not exceeded.

UN CO RR

where F and C are the ow vector and capacity vector as known to link i. The only non-zero elements of the clique matrix Qi lie in the interference neighborhood of CG-node i, and so the only aected elements of Fi and Ci are the corresponding ones. 5.5. Obstructions in ad-hoc network Typically, interference regions in a real ad-hoc network are not shaped like perfect disks, due to the presence of obstructions. In such situations, the underlying CG may not be a UDG. Thus, condition (7) may not be sucient to guarantee the existence of a feasible schedule, as the proof of Theorem 2 depends on the UDG property. One solution to this problem is to construct what we call a virtual conict graph, where links that lie within an interference range of each other are always modelled as conicting, even if in reality an obstacle prevents the links from interfering each other. Note that the virtual-CG is a UDG by construction, even if the underlying CG is not. Using the virtual-GG we may state and prove the following.

ADHOC 158 18 February 2006; Disk Used 8

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

599 Theorem 3. A set of ow rate assignments F has a 600 feasible schedule if 601 603 QV F 6 C 0:46; 10 604 where QV is the clique incidence matrix of the virtual605 CG, as dened above. This is valid even if the true 606 CG is not UDG. 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 Proof. Note that the virtual-CG has the same set of CG-nodes (corresponding to links in the connectivity graph) as the true-CG, while the set of CG-edges of the virtual-CG are a superset of the edges of the true-CG. We may dene integer ow rates Ki where Ki Fi C , a virtual-CGf, and a true-CGf in the T same way that we did in the proof of Theorem 1. By Theorem 2, condition (10) implies that there exists a coloring of the virtual-CGf with at most T colors. However, a valid coloring of the virtualCGf is a valid coloring of the true-CGf as well, because each node in the true-CGf has a subset of the neighbors it has in the virtual-CGf. Thus, there exists a valid coloring of the true-CGf, which in turn implies the existence of a feasible schedule to achieve the ow rate vector F. h

B
1/2 1

1/2

OO F
sqrt(3)/2 1 1/2 1

1/2

TE

p stripes of width 3=2 separated by a distance of 1. Since each stripe is perfect, and their separation is greater than 1, the entire stripe-graph is perfect. Then, the probability p that any node is covered p 3 2 . by a particular stripe-graph is given by p p 3
2

PR

Fig. 4. Connectivity band of a node in a stripe.

646 647 648 649 650

623 5.6. Variance in interference range 624 625 626 627 628 629 630 631 632 633 634 635 636 637

UN CO RR

Often, the interference region may not be shaped like a perfect disk, but be uneven near the edges, due to fading eects. Such an interference region may be modelled as being bounded between two disks. Two CG-nodes cannot interfere if their distance >1, and two CG-nodes will always interfere if their distance 6x 6 1. But if two CG-nodes are separated by a distance between x and 1, they may or may not interfere. We would like to take this variance into account while employing our constraint-based approach. In order to present the proof of our extension, it is useful to sketch the proof of the original imperfection theorem, as given in [4].

EC

Extending a probability argument, the authors show 651 that the imperfection of the graph is bounded 652 . h above, by the reciprocal, i.e., impG 6 1 653 p 654 655 656 657 658 659 660 661 662 663 664 665 666

We dene the connectivity band of a node in a stripe in the CG. In Fig. 4, it follows from geometry that a node A will necessarily be connected to all other nodes in the stripe which lie within a connectivity band of height 1/2 in either direction (shaded region in the gure). So if two nodes are not connected to each other, their Y coordinates must be separated by at least 1/2. The following theorem now follows from Theorem 4. Theorem 5. When the interference range x in a CG varies between 1 2 6 x 6 w 6 1, the imperfection ratio is bounded by, s  2 1z 1 2 ; where z x . 12 impCG 6 z 2 Proof. In [17], the authors show that the Stripe Lemma holds if for all nodes located within the stripe, two nodes on either side of a third node, but not adjacent to the third node (e.g. nodes B and C in Fig. 4) are never connected themselves. This is a corollary of having a connectivity band

638 Theorem 4 [4]. For a UDG, p 3 1 impG 6 2 p % 2:155. 3 640 2 641 642 643 644 645

668 669 670 671 672 673 674

11

Proof. The proof uses the Stripe Lemma from p[17] which implies that in a UDG, stripes of width 3=2 are perfect. The technique used in the proof is to cover the UDG with a large number of randomly positioned stripe-graphs. A stripe-graph consists of

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

675 of width 1/2 on either side of a node B and C on 676 either side of A are necessarily separated by a dis677 tance greater than 1. 678 679 680 681 682 683 684 685 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 When the interference range varies between x and 1, a modication of the Stripe Lemma using triangle geometry ensures the stripe is always 2 that perfect provided z2 1 6 x2 , where z is the width 2 of the stripe. Consider Fig. 4 again: We need to ensure that the band of height 1/2 on either direction of A is within its range of connectivity, i.e. the diagonal of the connectivity band 6x. At the q

6. Row vs. clique constraints We have presented two sets of constraints, both of which may be evaluated in a distributed fashion. In this section, we discuss the ecacy of using one set of constraints versus the other.

715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746

686 limit, we get z

UN CO RR

Hence, existence of a feasible schedule is guaranz teed provided QF 6 C 1 . z Table 1 tabulates the values of the bound on imperfection ratio as a function of the unevenness in the interference region. As seen, the imperfection can grow as the interference range varies more. This extension can also account for the approximation error introduced by representing a link by its mid-point (Section 3.1). To account for this, we can model the interference range as lying between two discs of radius x and x + q. Once again it is important to realize that the clique constraints applied to a virtual conict graph akin to the one described in Section 5.5 will also imply a feasible schedule. In this case, the virtual CG corresponds to the maximum interference range (i.e., modelling the variance in the interference range by its upper bound). A feasible schedule in the virtual CG ensures a feasible schedule in the true CG as well. Depending on the value of the imperfection bound as given in Table 1, we may choose to use the clique constraints on the virtual CGQV F 6 C 0:46 or the constraints on the true CGQF 6 z C 1 . z

EC

TE

PR
Scaled Clique

x2 1=22 . Note that since p x 6 1, we have z 6 3=2. Following the proof technique of Theorem 4 [4], the imperfection ratio in this case is bounded by z impCG 6 1 z . h

We have seen that the row constraints are sucient, but may be arbitrarily far from being necessary, as we observed in Section 4. On the other hand, the scaled clique constraints are not only sufcient, but also within a bounded factor of the necessary constraints (Section 5.4). We present a schematic of the relationship between the various sets of constraints discussed earlier. Each of the constraint sets denes a polytope in Rn , describing the feasible range of values for the ow vector F. We show a two dimensional representation of these regions in Fig. 5. The independent set (IS) polytope corresponds to the necessary and sucient condition for a feasible schedule to exist, and is therefore the benchmark. The outer clique polytope is derived from necessary conditions and so contains the IS polytope. Scaling this by a factor of 0.46 gives us the scaled clique polytope. This corresponds to sucient conditions for feasibility, and is hence entirely contained within the IS polytope. The row constraints also describe sucient conditions, and so the row polytope lies within the IS polytope as well. The row and the scaled clique constrained polytopes certainly overlap (e.g. at the point when the ow vector is 0), but do not contain one another.

Table 1 Imperfection ratio bound as a function of interference unevenness x 1 0.9 0.8 0.7 0.6 z 0.866 0.75 0.62 0.49 0.33 imp 2.155 2.33 2.60 3.04 4 1/imp 0.46 0.43 0.38 0.33 0.25

Fig. 5. Polytopes describing feasible regions for the ow rate vector, as dened by the row, clique and independent set constraints.

OO F
Clique Independent Set

6.1. Relating row and clique constraints

Row

ADHOC 158 18 February 2006; Disk Used 10

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

784 783 785 786 787 788 789 790 791 792 793 794 795

MB F 6 C .

Proof. As in the proof of Theorem 1 we dene integer ow rates Ki where Ki Fi C , and we dene T the graph CGf by replacing each node i of CG with a clique of Ki children nodes. Recall that a coloring of graph CGf using at most T colors implies the existence of a feasible schedule. Let CGA f denote the subgraph of CGf restricted to CGf-nodes whose parents are in A. Theorem 2 implies the existence of a A coloring for CGA f . We color the nodes of CGf using such a coloring, and now we seek to color the remaining nodes of CGf.

UN CO RR

774 775 776 777 778 779 780 781

Theorem 6. Suppose the CG-nodes are partitioned into two disjoint sets A and B. Let QA denote a reduced clique matrix, containing only the rows of Q that correspond to cliques containing one or more CGnodes in A. Similarly let MB denote a reduced conict graph incidence matrix, containing only the rows of M that describe the neighbors of CG-nodes in B. The ow rate vector F is feasible if A Q F 6 0:46 C ;

EC

761 762 763 764 765 766 767 768 769 770 771 772 773

Because the row constraints may be less restrictive in some circumstances than the clique constraints, we would like to give each station in the network the exibility to choose to use either the row constraints or the scaled clique constraints, depending on which are less restrictive for the situation. However, we need to be sure that in a network where some stations use clique constraints, and others use row constraints, the union of the constraints being checked across the network constitute sucient conditions for a ow rate vector to be feasible. We show that this is indeed the case in the theorem that follows.

7.1. Computing cliques For our purposes, we would like to compute maximal cliques in a computationally simple, distributed and localized manner. General algorithms to generate cliques (e.g. [1820]) are centralized in nature. Also, these are exponential algorithms since the number of maximal cliques in a graph (even in a UDG) is exponential. So we aim for a polynomial approximation algorithm. The essence of the approximation is to use slightly super-maximal cliques. When the number of cliques grows large, several nearby cliques will have a signicant intersection, i.e., their membership will dier only at a few nodes. In this case, the approximation algorithm generates the union of these as the super-maximal clique. The exact set of maximal cliques generated depends upon the location of the nodes. By using approximate cliques, the constraints are only strengthened further as multiple individual constraints are replaced by their union. Thus, the suciency of the approximated clique constraints implies the suciency of the actual set of clique constraints. The heuristic algorithm presented by authors of this paper in [21] distributedly approximates all maximal cliques that a CG-node belongs to. The algorithm in fact works on any UDG, a more general class of graphs than conict graphs alone. It makes use of certain key geographic structures of these graphs. For each edge, we limit the set of vertices that may form cliques with this as the longest edge. We then consider several characteristic shapes determined by that edge, and prove that all cliques

PR

760 6.2. Using both row and clique constraints

OO F

747 748 749 750 751 752 753 754 755 756 757 758 759

Indeed, there are notable cases (typically simple networks, and/or few ows) where the row constraints are less pessimistic than scaled clique constraints. An example is the pentagon CG (Fig. 3), where row constraints allow a ow rate of C/3 = 0.33C is possible on each link, while scaled clique constraints only allow a ow rate of 0.46 C/2 = 0.23C. However, as the number of ows increases, we are likely to get situations where a CG-node has many neighbors, which do not all form a single clique. Such topologies make the row constraints too restrictive, as is often the case in realistic adhoc networks.

We observe that MB F 6 C implies MB K 6 T which in turn implies that each node in CGB f has less than T neighbors. It is worth emphasizing that each node in CGB f has less than T neighbors in total, A including neighbors in CGB f and neighbors in CGf . We may now color the remaining nodes of CGf by using the same greedy coloring algorithm as in Theorem 1. Because the remaining nodes each have less than T neighbors, we can always nd a color with index in {1, . . ., T}. Thus we have found a coloring for all of CGf, and therefore there exists a feasible schedule to accommodate ow rate vector F. h 7. Distributed algorithms

796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843

TE

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

11

855 7.2. Capacity estimation 856 857 858 859 860 861 862 863 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 We can use the clique constraints to estimate the capacity of an ad-hoc network, in a localized and distributed way. Assume that each CG-node is aware of all its interference neighbors, and their allocated ows. Using this localized information, each CG-node can estimate its available capacity Ci as Ci minfCi 0:46 Qi Fi g.
i

PR

OO F

844 845 846 847 848 849 850 851 852 853 854

involving this edge are included in one of the sets of vertices contained in these shapes. The evaluation of the vertices located within these shapes may be done in polynomial time, enabling us to limit the running time of the algorithm. The algorithm works in O(mD2) time and generates O(mD) cliques, where m is the number of edges in the graph and D is its maximum degree. We also provide a modied version of the algorithm which improves the performance in many cases, albeit without aecting the worst case running time.

The admission control algorithm is eected when a new ow request {src, dest, path, bw} is received by the network. The new ow description is sent out along the path of the ow. Every station that receives the ow request updates its ow vector Fi with the new ow parameters. And then it recomputes Ci minfCi 0:46 Qi Fi g for all its links. If Ci < 0 for any link i, an admission denied message is generated. We assume that the route for the ow is known ahead of time. This allows us to accept or reject ow requests based on the ad-hoc bandwidth available in the network. We can further incorporate routing techniques that make use of interference knowledge, as in [12]. 8. Simulation results We present simulation results to test the various ideas discussed earlier in the paper. We perform simulations using OPNET [22], which implements detailed packet level simulation models of channels, interference, as well as the 802.11 MAC and ad-hoc routing protocols. 8.1. Feasible vs. actual schedule In Sections 4 and 5, we have shown that a feasible schedule exists when the row constraints or the scaled clique constraints are satised. However, it is important to note that the mere existence of a feasible schedule does not imply our being able to nd it, let alone impose it on all the ad-hoc stations. The task of nding a distributed scheduling mechanism that achieves the feasible schedule is a well-known open question, and is beyond the scope of this paper. Instead, we compare our theoretical models with a practical MAC protocol the default 802.11b. We make no changes to the existing 802.11b, and simply use it to check against the capacity limits predicted by our model. 8.2. Row constraints First, we evaluate the row constraints presented in Section 4. The conict graph evaluated here is shaped like a star, as shown earlier in Fig. 2. All the links A, B, C, and D interfere with link X, but none of these interfere with each other. Assuming the eective capacity of the channel to be 5 Mbps,

893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937

13

883 7.3. Admission control 884 885 886 887 888 889 890 891 892

A distributed admission control scheme may now be overlaid on the capacity estimation framework. We assume as above that the CG-nodes in the conict graph are aware of all their interference neighbors, and their allocated ows. Using the approximation algorithm described in [21], each CG-node (in reality, the node at one end of this link) keeps track of the maximal cliques it is part of, and the available bandwidth on them.

UN CO RR

C is the available bandwidth on link i, taking into account ows allocated on i, as well as interference from neighboring links. We consider all maximal cliques that i belongs to, and take the worst case available capacity over all the cliques. This ensures that the available capacity satises the conditions for suciency, as described earlier in the paper. The value Ci is a vital commodity for algorithms involving quality or rate guarantees in a network. In a wired network, the width or available bandwidth of a path is determined by the remaining bandwidth on the bottleneck link in the path. In the case of an ad-hoc network, it is the bandwidth available on the bottleneck clique that oers a parallel metric of comparison. We therefore expect our methods of computing Ci to be utilized frequently, in distributed algorithms for admission control and QoS routing.

EC

TE

ADHOC 158 18 February 2006; Disk Used 12 Table 2 Simulation results star conict graph Comment One outer link Two outer link Three outer link Four outer link Two outer + center Four outer + center Limit outer links Limit outer links Generated 5 25 35 45 25+5 45+5 4 2.5 + 2 4 0.5 + 4

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

Table 3 Simulation results pentagon conict graph Received 5 2 4.5 3 4.5 4 4.5 2 4.5 + 0.25 4 4.5 + 0.01 4 2.5 + 0.85 4 0.5 + 3.5 Comment Unsaturated Unsaturated Reaching saturation Saturated Over-saturated Generated 51 5 1.5 52 5 2.5 55 Received 51 5 1.5 52 5 1.8 5 1.8

941 FA FB FC FD FX 6 5. 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 In fact, each of the link A, B, C, and D should be able to achieve close to the 5 Mbps capacity. The results of this simulation are presented in Table 2. All rates are in Mbps, and the notation 3 p + q in the table implies that three of the outer links are all generating/receiving p Mbps, while link X at the center gets q Mbps. As seen from the rst set of rows of the table, all four of the outer links can indeed achieve 4.5 Mbps each so these rates exceed the row constrained rates by a large margin. This corroborates the fact that the row constraints may be overly pessimistic, as discussed in Section 4. The conict graph shown in Fig. 2 is perfect, and so the unscaled clique constraints of expression (6) should be sucient [3]. The clique constraints for this graph look like FA FX 6 5; FB FX 6 5, etc. Indeed, as seen from the received rates in Table 2, the clique constraints are always satised. We also observe the unfair nature of the sharing of links. The second set of rows show that link X is starved when multiple of its neighbors are transmitting simultaneously. Only by rate limiting the neighboring trac (third set of rows) can link X hope to achieve its share.

TE

8.4. Randomly generated network

PR

938 satisfying the row constraints (at CG-node X) 939 requires

although analysis conrms that each link is limited to 2 Mbps at most (Section 5.3). The simulation results presented in Table 3 support the analysis quite well. When the trac on each link is shaped to the predicted limit of 2 Mbps, all links are able to achieve 2 Mbps each. Increasing the oered rates further, to 2.5 Mbps on each link, only makes matters worse the links get saturated and the throughput is reduced to only 1.8 Mbps per link. The row constraints would only allow 5/ 3 = 1.67 Mbps per link, and so they are again too restrictive.

OO F

974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013

967 8.3. Clique constraints and scaling 968 969 970 971 972 973

In order to observe the insuciency of the unscaled clique constraints, we need to simulate an imperfect CG the pentagon in Fig. 3. The unscaled clique constraints on this conict graph suggests that a rate of 2.5 Mbps should be achievable on each link (assuming a 5 Mbps channel capacity);

We generate a random ad-hoc network consisting of 50 ad-hoc stations in a 2.5 km 2.5 km area. The locations of the stations are fed into OPNET. Transmission range is set to 500 m, and the interference range is 1 km. These numbers roughly correspond to a battalion of tanks in a battleeld, with powerful radios. We pick ve pairs of stations at random and set up video ows between them (running for 5 min), and alter these rates in order to change the load on the network. We use the DSR routing protocol [23] to determine the routes, and measure the amount of trac received at each of the receiving stations in the OPNET simulation. In parallel, we feed the routes generated into the theoretical model implemented in MATLAB [24]. Using Eq. (13), we calculate the minimum spare capacity Ci on each link i, to compare against the actual rates received by the ows. The minimum spare capacity in the network is calculated by taking the minimum of Ci over all the links. We run simulations involving three, four and ve ows. In many of the simulations, the calculated spare capacity on the network is negative, which predicts that some of the ows may be losing trac. Each simulation run yields a single point on the plot, with the rates and the routes of all the ows

UN CO RR

EC

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

13

600

Average Rate Received per Flow: 50 nodes in 2.5 km X 2.5 km

500

400

300

200

100

100

200

300

400

500

600

Fig. 6. Comparing received rates with theoretical model.

UN CO RR

1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031

together determining the Ci values for each link, and thereby for the network. A summary of the simulation results is shown in Fig. 6. For an individual simulation run, we assign the same video rate to each of the three, four or ve ows. On the X-axis, we plot the rate of video trac sent out on each of these ows; while the Y-axis plots the average rate received over all the ows. By looking at the routes, and interpolating between the various video rates, we can also determine the exact transmitted rate at which = 0, as predicted by our model. We determine this limit for each of the three, four and ve ow cases, and plot these using the dotted vertical lines. As seen from the gure, the ows appear to receive almost all their trac until the predicted limit. In each case, the ows experience a sharp loss of quality soon after the theoretical limit is crossed.

EC

TE

PR

Average rate sent (in kbps)

ble to analytically determine that imperfection vCGf for particular ow ratio. At best, we can nd j CGf vectors. On the other hand, there are several ineciencies in the distributed nature of the MAC protocol (e.g. [9,25]). When graph imperfection is signicantly less than 2.155, our observation that the throughput falls o soon after the oered load exceeds the scaled clique constraints, must be because MAC ineciency in addition to any graph imperfection is limiting the achievable rates. In fact, given an ideal scheduler, the unscaled clique constraints ought to be necessary and sucient in Fig. 6, since vCGf we nd that j 1 in this specic case. Thus, in CGf this example, the limits on the throughput is primarily due to the ineciency of the MAC. We are not violating the clique constraints here they are merely being superseded by the ineciencies in the MAC protocol. The scaling factor of 0.46 is required, to account for the worst case imperfection in a CG, since we cannot predict the ow vectors ahead of time. In simple networks and/or few ows (e.g. Section 8.3), the eect of the graph imperfection is easily visible. In some other situations however, the eects of MAC ineciency, rather than graph imperfection, might be the dominating cause for the gap between a ow vector satisfying the unscaled clique constraints and one that is actually achievable with a distributed MAC. Thus, the constraint-based framework presented here becomes universally practicable only when used in conjunction with an ecient MAC protocol.

1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094

Average rate received (in kbps)

9. Conclusions This paper presents a theoretical model to predict the capacity of an arbitrary ad-hoc network, with a given set of desired ows. We model the ad-hoc network, and the underlying interference between links, as a conict graph. We then propose two sets of constraints the row constraints and the scaled clique constraints to determine if a ow vector is feasible on this network. Our main contribution is to prove that each of the above constraints are sucient for the existence a feasible schedule. We also discuss the tightness of these constraints under adhoc network conditions. Our second contribution is to expand the model of the network to incorporate variations in the interference range, and consider obstructions (like build-

1032 8.5. MAC ineciencies and graph imperfection 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 Our simulation results indicate that the achieved throughput in an ad-hoc network is often limited to a point close to the scaled clique limits. However, the bound of 0:1 2:155 on the imperfection ratio 46 may be signicantly larger than the typical case imperfection ratio. In fact, [4] states a conjecture that the imperfection ratio is in fact bounded above by 3 . 2 Recall that the imperfection ratio is the ratio between the chromatic number v(CGf) and the largest clique size j(CGf), maximized across all possible ow vectors. Consequently, for the random networks generated in our simulations, it is not possi-

OO F

ADHOC 158 18 February 2006; Disk Used 14

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

1111 Acknowledgement 1112 Special thanks to Dr. Stephanie Gerke for her 1113 advice and comments. We are also grateful to Bill 1114 Hodge for his help in running many of the laborious 1115 OPNET simulations. 1116 References

1117 [1] R. Gupta, Quality of Service in Ad-Hoc Networks, PhD 1118 Thesis, University of California, Berkeley, May 2005. 1119 [2] H. Luo, S. Lu, V. Bhargavan, A new model for packet 1120 scheduling in multihop wireless networks, in: Proceedings 1121 ACM Mobicom, 2000, pp. 7686. 1122 [3] K. Jain, J. Padhye, V.N. Padmanabhan, L. Qiu, Impact of 1123 interference on multi-hop wireless network performance, in: 1124 Proceedings ACM Mobicom 2003, San Diego, CA, Septem1125 ber 2003. 1126 [4] S. Gerke, C. McDiarmid, Graph imperfection I, Journal of 1127 Combinatorial Theory, Series B 83 (2001) 5878. 1128 [5] P. Gupta, P.R. Kumar, The capacity of wireless networks, 1129 IEEE Transactions on Information Theory 34 (5) (2000) 1130 910917. 1131 [6] M. Franceschetti, O. Dousse, D. Tse, P. Thiran, On the 1132 throughput capacity of random wireless networks, in 1133 preparation. 1134 [7] M. Grossglauser, D. Tse, Mobility increases the capacity of 1135 ad-hoc wireless networks, IEEE/ACM Transactions on 1136 Networking 10 (4) (2002) 477486. 1137 [8] A. El Gamal, J. Mammen, B. Prabhakar, D. Shah, 1138 Throughput-delay trade-o in wireless networks, in: Pro1139 ceedings IEEE INFOCOM, Hong Kong, March 2004. 1140 [9] J. Li, C. Blake, D.S.J. De Couto, H.I. Lee, R. Morris, 1141 Capacity of ad hoc wireless networks, in: Proceedings ACM 1142 Mobicom 2001, Rome, Italy, July 2001. 1143 [10] M. Kodialam, T. Nandagopal, Characterizing the achievable 1144 rates in multihop wireless networks, in: Proceedings ACM 1145 Mobicom 2003, San Diego, CA, September 2003.

UN CO RR

EC

TE

PR

1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110

ings or hills) in the network. We extend the above proofs of suciency to incorporate these changes. An important motivation to utilize the row and clique constraints is that they are localized in nature and amenable to distributed approximation algorithms. Our third contribution is to propose distributed algorithms for capacity estimation that utilize these constraints. The estimated capacity may then be utilized by other algorithms to implement admission control and QoS routing schemes in ad-hoc networks. While this paper does not claim to propose an implementable protocol for distributed ad-hoc QoS routing, it provides the underlying theoretical frame work that makes it possible to develop such schemes.

[11] Z. Jia, R. Gupta, J. Walrand, P. Varaiya, Bandwidth guaranteed routing for ad-hoc networks with interference consideration, in: Proceedings ISCC 2005, Cartagena, Spain, June 2005. [12] R. Gupta, Z. Jia, T. Tung, J. Walrand, Interference-aware QoS routing (IQRouting) for ad-hoc networks, in: Proceedings Globecom 2005, St. Louis, Missouri, November 2005. [13] R. Gupta, J. Walrand, Achieving fairness in a distributed adhoc MAC, in: E.H. Abed (Ed.), Advances in Control, Communication Networks, and Transportation Systems, Systems and Control: Foundations and Applications Series, Springer-Birkhauser, Boston, 2005 (Book Chapter). [14] M. Gast, 802.11 Wireless Networks: The Denitive Guide, Oreilly and Associates, 2002. [15] R. Negi, A. Rajeswaran, Physical layer eect on MAC performance in ad-hoc wireless networks, in: Proceedings Communications, Internet and Information Technology CIIT, 2003. [16] L. Lovasz, A characterization of perfect graphs, Journal of Combinatorial Theory, Series B 13 (1972) 9598. [17] A. Graf, M. Stumpf, G. Weisenfels, On coloring unit disk graphs, Algorithmica 20 (1998) 277293. [18] F. Harary, I.C. Ross, A procedure for clique detection using the group matrix, Sociometry 20 (1957) 205215. [19] J.G. Augustson, J. Minker, An analysis of some graph theoretical cluster techniques, Journal of the ACM (JACM) 17 (4) (1970) 571588. [20] C. Bron, J. Kerbosch, Finding all cliques in an undirected graph, Communications of the ACM 16 (1973) 575577. [21] R. Gupta, J. Walrand, O. Goldschmidt, Maximal cliques in unit disk graphs: polynomial approximation, in: Proceedings INOC 2005, Lisbon, Portugal, March 2005. [22] OPNET Modeller, OPNET Technologies Inc. Available from: <http://www.opnet.com>. [23] D.B. Johnson, D.A. Maltz, Dynamic source routing in adhoc wireless networks, in: T. Imielinski, H. Korth (Eds.), Mobile Computing, Kluwer, 1996, pp. 153181. [24] Matlab Simulation Environment, The Mathworks Inc. Available from: <http://www.mathworks.com>. [25] G. Bianchi, Performance analysis of the IEEE 802.11 distributed coordination function, IEEE Journal on Selected Areas in Communications 18 (3) (2000).

1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1190 1203 1204 1205

Rajarshi Gupta is currently a Senior Engineer with Qualcomm Inc, having completed his Ph.D. in EECS at UC Berkeley. Prior to this, he completed his MS also at Berkeley, and his BS at the University of Maryland. From 1999 to 2003, he worked with Extreme Networks as a Senior Designer, where he has been the author of 10 patents. He is interested in algorithms to ensure quality in networks both wired and ad-hoc. This includes: analysis of network capacity; switching and scheduling mechanisms for ecient utilization of resources; and, routing algorithms to guarantee quality of service.

OO F

ADHOC 158 18 February 2006; Disk Used

ARTICLE IN PRESS
R. Gupta et al. / Ad Hoc Networks xxx (2006) xxxxxx

No. of Pages 15, Model 3+

15

UN CO RR

EC

TE

PR

1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1207 1220 1221 1222 1223 1224 1225 1226

OO F

John Musacchio received the BS degree in electrical engineering from the Ohio State University in 1996 and the MS degree in electrical engineering from the University of California, Berkeley in 1998. From 2000 to 2003, he was with the Silicon Valley start-up Terablaze, where he helped architect a high-speed switch fabric chip-set. Since 2003 he has been at the University of California, Berkeley where he expects to complete his Ph.D. in electrical engineering by January 2005. In January 2005 he will join the Department of Information Systems and Technology Management of the University of California, Santa Cruz as an Acting Assistant Professor. His current research interests include stochastic modelling and control of queuing networks, applications of game theory in network pricing, and ad-hoc networks.

Jean Walrand received the Ph.D. degree from the Department of Electrical Engineering and Computer Sciences of the University of California at Berkeley where he is now Professor. His research interests include decision theory, stochastic processes, and communication networks. He is the author of An Introduction to Queueing Networks (Prentice Hall, 1988) and of Communication Networks: A First Course (2nd ed. McGraw-Hill,1998) and co-author of High-Performance Communication Networks (2nd ed, Morgan Kaufman, 2000). He is a Fellow of the Belgian American Education Foundation and of the IEEE and a recipient of the Lanchester Prize and of the Stephen O. Rice Prize.

1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1228 1241 1242 1243 1244 1245