1: %\documentclass[12pt,journal,draftclsnofoot,onecolumn]{IEEEtran}

2: %\documentclass[journal,draftcls,onecolumn]{IEEEtran}

3: \documentclass[final]{IEEEtran}

4: 

5: %\usepackage{cite}

6: \usepackage{graphicx}

7: %\usepackage{psfrag}

8: \usepackage{subfigure}

9: %\usepackage{url}

10: %\usepackage{stfloats}

11: 

12: \usepackage{amsmath}

13: \usepackage{amssymb}

14: \usepackage{algorithmic}

15: \usepackage{algorithm}

16: 

17: \hyphenation{op-tical net-works semi-conduc-tor}

18: 

19: \newtheorem{theorem}{Theorem}

20: \newtheorem{lemma}{Lemma}

21: \newtheorem{proposition}{Proposition}

22: \newtheorem{definition}{Definition}

23: \newtheorem{remark}{Remark}

24: 

25: \newcommand{\dotcup}{\ensuremath{\mathaccent\cdot\cup}}

26: 

27: \begin{document}

28: 

29: \title{On High Spatial Reuse Link Scheduling in STDMA Wireless Ad Hoc Networks}

30: 

31: \author{Ashutosh~Deepak~Gore\thanks{A.D. Gore and A. Karandikar are

32:     with the Information Networks Laboratory, Department of Electrical

33:     Engineering, Indian Institute of Technology - Bombay, Mumbai

34:     400076, India.  Email:

35:     \{adgore,karandi\}@ee.iitb.ac.in},~Srikanth~Jagabathula\thanks{S.

36:     Jagabathula is with the Laboratory for Information and Decision

37:     Systems, Massachusetts Institute of Technology, Cambridge, MA

38:     02139, USA.  Email: jskanth@mit.edu}~and~Abhay~Karandikar }

39: 

40: \maketitle

41: 

42: \begin{abstract}

43: Graph-based algorithms for point-to-point link scheduling in Spatial

44: reuse Time Division Multiple Access (STDMA) wireless ad hoc networks

45: often result in a significant number of transmissions having low

46: Signal to Interference and Noise density Ratio (SINR) at intended

47: receivers, leading to low throughput. To overcome this problem, we

48: propose a new algorithm for STDMA link scheduling based on a graph

49: model of the network as well as SINR computations.  The performance of

50: our algorithm is evaluated in terms of spatial reuse and computational

51: complexity.  Simulation results demonstrate that our algorithm

52: achieves better performance than existing algorithms.

53: \end{abstract}

54: 

55: \begin{keywords}

56: Wireless Ad hoc Networks, Spatial Time Division Multiple Access, Link

57: Scheduling, Physical Interference Model, Spatial Reuse.

58: \end{keywords}

59: 

60: \section{Introduction}

61: 

62: \PARstart{A}{} wireless ad hoc network consists of a finite number of

63: radio units (nodes) that are geographically distributed in a terrain

64: without any preplanned or fixed infrastructure. They communicate with

65: each other via the untethered and broadcast wireless medium.  In order

66: to use the scarce and expensive wireless spectrum efficiently, we need

67: to exploit channel spatial reuse, i.e., allow concurrent communication

68: between source-destination pairs which are ``reasonably" far from each

69: other using either the same time slot or frequency band.

70: 

71: A commonly used scheme for channel reuse is Spatial Time Division

72: Multiple Access (STDMA), in which time is divided into fixed-length

73: slots that are organized cyclically. An STDMA schedule describes the

74: transmission rights for each time slot in such a way that

75: communicating pairs assigned to the same slot do not collide.  STDMA

76: scheduling algorithms can be categorized into link scheduling

77:  and broadcast/node scheduling algorithms

78: \cite{ramanathan_lloyd__scheduling_algorithms}. In a wireless ad hoc

79: network, a link is an ordered pair of nodes $(t,r)$, where $t$ is a

80: transmitter and $r$ is a receiver. In link scheduling, the

81: transmission right in every slot is assigned to certain links. On the

82: other hand, in broadcast scheduling, the transmission right in every

83: slot is assigned to certain nodes. Thus, there is no apriori binding

84: of transmitter and receiver and the packet transmitted can be received

85: by every neighbor.  Link scheduling is suitable for unicast traffic,

86: while broadcast scheduling is suitable for broadcast traffic. In this

87: paper, we will concentrate on link scheduling for STDMA

88: networks. Specifically, we consider centralized link scheduling, i.e.,

89: the link transmission schedule is computed by a central

90: entity. Centralized scheduling is applicable for scenarios where the

91: time scale of topology change is much larger than the duration of the

92: schedule.

93: 

94: \subsection{Related Work}

95: 

96: The concept of STDMA for multihop wireless ad hoc networks was

97: formalized in \cite{nelson_kleinrock__spatial_tdma}. Centralized

98: algorithms \cite{hajek_sasaki__link_scheduling}

99: \cite{funabiki_takefuji__parallel_algorithm} as well as distributed

100: algorithms \cite{chlamtac_lerner__fair_algorithms}

101: \cite{chlamtac_pinter__distributed_nodes} have been proposed for

102: generating reuse schedules. The problem of determining an optimal

103: minimum-length STDMA schedule for a general multihop ad hoc network is

104: NP-complete for both link and broadcast scheduling

105: \cite{ramanathan_lloyd__scheduling_algorithms}.  In fact, this is

106: closely related to the problem of determining the minimum number of

107: colors to color all the edges (or vertices) of a graph under certain

108: adjacency constraints.  However, most wireless ad hoc networks can be

109: modeled by planar or close-to-planar graphs and thus near-optimal edge

110: coloring algorithms can be developed for these restricted classes of

111: graphs.

112: 

113: A significant work in STDMA link scheduling is reported in

114: \cite{ramanathan_lloyd__scheduling_algorithms}, in which the authors

115: show that tree networks can be scheduled optimally, oriented graphs

116: can be scheduled near-optimally and arbitrary networks can be

117: scheduled such that the schedule is bounded by a length proportional

118: to the graph thickness\footnote{The thickness of a graph is the

119: minimum number of planar graphs into which the given graph can be

120: partitioned.} times the optimum number of colors.

121: 

122: A probabilistic analysis of the throughput performance of graph-based

123: scheduling algorithms under the physical interference model is derived

124: in \cite{behzad_rubin__performance_graph}. The authors determine the

125: optimal number of simultaneous transmissions by maximizing a lower

126: bound on the physical throughput and subsequently propose a truncated

127: graph-based scheduling algorithm that provides probabilistic

128: guarantees for network throughput.

129: 

130: In \cite{guo_roy__spatial_reuse}, the authors present an analytical

131: framework to investigate co-channel spatial reuse in dense wireless ad

132: hoc networks based on path loss and log-normal shadowing models for

133: a 1-D infinite regular chain topology and a 2-D infinite

134: hexagonally-tessellated topology. They derive the minimum ratio of

135: inter-transmitter distance to transmitter-receiver distance, while

136: still maintaining desirable Signal to Interference and Noise

137: density Ratio (SINR) at the receivers. Their results

138: demonstrate that increasing transmission power improves spatial reuse

139: in ambient noise dominated environments. However, in co-channel

140: interference limited scenarios, increasing transmission power has

141: little effect on spatial reuse.

142: 

143: The performance of centralized graph-based and interference-based

144: STDMA scheduling via simulations is evaluated and compared in

145: \cite{gronkvist_hansson__comparison_between}.  To generate a

146: graph-based conflict-free schedule, the authors use a two-level graph

147: model with certain SINR threshold values chosen based on heuristics

148: and examples. To generate an interference-based conflict-free

149: schedule, the authors employ a method suggested in

150: \cite{somarriba__multihop_packet} which describes heuristics based on

151: two path loss models, namely terrain-data based ground wave propagation

152: model and Vogler's five knife-edge model.

153: 

154: In \cite{brar_blough_santi__computationally_efficient}, the authors

155: investigate throughput improvement in an 802.11-like wireless mesh

156: network with Carrier Sense Multiple Access with Collision Avoidance

157: (CSMA/CA) channel access scheme replaced by STDMA. For a successful

158: packet transmission, they mandate that two-way communication be

159: successful. Under this `extended physical interference model', they

160: present a greedy algorithm which computes the transmission schedule in

161: a centralized manner. Assuming a uniform random node distribution and

162: using results from occupancy theory, they derive an approximation

163: factor for the length of this schedule relative to the shortest

164: schedule.

165: 

166: Link scheduling for power-controlled STDMA

167: networks under the physical interference model is analyzed in

168: \cite{moscibroda_wattenhofer__complexity_connectivity}. The authors

169: define the scheduling complexity as the minimum number of time slots

170: required for strong connectivity of the network. They develop an

171: algorithm employing non-linear power assignment and show that its

172: scheduling complexity is polylogarithmic in the number of nodes.

173: 

174: In a related work

175: \cite{moscibroda_wattenhofer_zollinger__topology_control}, the authors

176: investigate the time complexity of scheduling a set of communication

177: requests in an arbitrary network.  They consider a `generalized

178: physical model' wherein the actual received power of a signal can

179: deviate from the theoretically received power by a

180: multiplicative factor. Their algorithm successfully schedules all

181: links in time proportional to the squared logarithm of the number of

182: nodes times the static interference measure

183: \cite{rickenbach_schmid__robust_interference}.

184: 

185: In \cite{kim_lim_hou__improving_spatial}, the authors investigate the

186: tradeoff between the average number of concurrent transmissions

187: (spatial reuse) and sustained data rate per node for an 802.11

188: wireless network.  Assuming that the channel data rate is given by the

189: Shannon capacity, they show that spatial reuse depends only on the

190: ratio of transmit power to carrier sense threshold. Keeping the

191: carrier sense threshold fixed, they propose a distributed power and

192: rate control algorithm based on interference measurement and evaluate

193: its performance via simulations.

194: 

195: In \cite{lim_lim_hou__coordinate_based}, the authors investigate

196: mitigation of inter-flow interference in an 802.11e wireless mesh

197: network from a temporal-spatial diversity perspective.  Measurements

198: of received signal strengths are used to construct a virtual

199: coordinate system to identify concurrent transmissions with minimum

200: inter-flow interference. Based on this new coordinate system, the

201: gateway node determines the scheduling order for downlink frames of

202: different connections. Through extensive simulation with real-life

203: measurement traces, the authors demonstrate throughput improvement

204: with their algorithms.

205: 

206: In \cite{kodialam_nandagopal__characterizing_achievable}, the authors

207: consider wireless mesh networks with half duplex and full duplex

208: orthogonal channels, wherein each node can transmit to at most one

209: node and/or receive from at most $k$ nodes ($k \geqslant 1$) during

210: any time slot. They investigate the joint problem of routing flows and

211: scheduling link transmissions to analyze the achievability of a given

212: rate vector between multiple source-destination pairs. The scheduling

213: problem is solved as an edge-coloring problem on a multi-graph and the

214: necessary conditions from scheduling problem lead to constraints on

215: the routing problem, which is then formulated as a linear optimization

216: problem. Correspondingly, the authors present a greedy coloring

217: algorithm to obtain a 2-approximate solution to the chromatic index

218: problem and describe a polynomial time approximation algorithm to

219: obtain an $\epsilon$-optimal solution of the routing problem using the

220: primal dual approach. Finally, they evaluate the performance of their

221: algorithms via simulations.

222: 

223: \subsection{Contributions of our Work}

224: 

225: In most STDMA algorithms, a link schedule is usually determined from a

226: graph model of the network

227: \cite{ramanathan_lloyd__scheduling_algorithms}

228: \cite{behzad_rubin__performance_graph}

229: \cite{gronkvist_hansson__comparison_between}. However, graph-based

230: scheduling algorithms assume a limited knowledge of the interference

231: and result in low network throughput. On the other hand, SINR-based

232: scheduling algorithms

233: \cite{brar_blough_santi__computationally_efficient}

234: \cite{moscibroda_wattenhofer__complexity_connectivity}

235: \cite{moscibroda_wattenhofer_zollinger__topology_control}

236: \cite{kim_lim_hou__improving_spatial} require a complete knowledge of

237: the interference and lead to higher throughput.  Existing literature

238: on SINR-based STDMA link scheduling consider system models which are

239: different from our system model. For example,

240: \cite{brar_blough_santi__computationally_efficient}

241: \cite{kim_lim_hou__improving_spatial} consider a variant of 802.11

242: wireless networks, \cite{lim_lim_hou__coordinate_based}

243: \cite{kodialam_nandagopal__characterizing_achievable} consider

244: wireless mesh networks and

245: \cite{moscibroda_wattenhofer__complexity_connectivity}

246: \cite{moscibroda_wattenhofer_zollinger__topology_control} assume

247: non-uniform transmit power at all nodes.

248: 

249: In this paper, we consider an STDMA wireless ad hoc network with

250: uniform transmit power at all nodes and propose a link scheduling

251: algorithm based on the graph model as well as SINR computations. We

252: introduce spatial reuse as an important performance metric and argue

253: that a high value of spatial reuse directly translates to high

254: long-term network throughput. We show that the proposed algorithm has

255: low computational complexity and high spatial reuse compared to

256: existing algorithms.

257: 

258: The rest of the paper is organized as follows. In Section \ref{xkybv},

259: we describe our system model along with the physical and protocol

260: interference models, discuss the limitations of graph-based scheduling

261: algorithms, formulate the problem and summarize the differences

262: between our work and existing work in SINR-based scheduling

263: algorithms. Section \ref{lxvym} describes the proposed link scheduling

264: algorithm. The performance of our algorithm is evaluated in Section

265: \ref{cymoe} and its computational complexity is derived in Section

266: \ref{mbxgt}. We conclude and suggest directions for future work in

267: Section \ref{eiovv}.

268: 

269: \section{System Model}

270: \label{xkybv}

271: 

272: Consider an STDMA wireless ad hoc network with $N$ static nodes

273: (wireless routers) in a two-dimensional plane.  During a time slot, a

274: node can either transmit, receive or remain idle. We assume

275: homogeneous and backlogged nodes. Let:

276: 

277: \begin{eqnarray*}

278: (x_j,y_j) &=& 

279: \mbox{Cartesian coordinates of $j^{th}$ node} \;=:\;{\mathbf r}_j\\

280: P &=& \mbox{transmission power of every node}\\

281: N_0 &=& \mbox{thermal noise density}\\

282: D(j,k) &=& \mbox{Euclidean distance between nodes $j$ and $k$}

283: \end{eqnarray*}

284: 

285: We do not consider fading and shadowing effects. The received signal

286: power at a distance $D$ from the transmitter is given by

287: $\frac{P}{D^{\alpha}}$, where $\alpha$ is the path loss factor.

288: 

289: A link schedule effectively assigns sets of links to time slots.

290: Specifically, a link schedule for the STDMA network is denoted by

291: $\Psi(C,{\mathcal S}_1,\cdots,{\mathcal S}_C)$, where

292: \begin{eqnarray*}

293: C &=& \mbox{number of slots in the link schedule}\\

294: {\mathcal S}_i 

295:  &=& \mbox{set of transmitter-receiver pairs which can}\\

296: && \mbox{communicate concurrently in the $i^{th}$ slot}\\

297:  &:=& \{t_{i,1}\rightarrow r_{i,1},\cdots,t_{i,M_i}\rightarrow r_{i,M_i}\}

298: \end{eqnarray*}

299: where $t_{i,j}\rightarrow r_{i,j}$ denotes a packet transmission from

300: node $t_{i,j}$ to node $r_{i,j}$ in the $i^{th}$ slot\footnote{A node

301: is generically denoted by $j$, $j=1,\ldots,N$. However, we have used

302: the notation $t_{i,j}$ to denote a node {\it transmitting} in the

303: $i^{th}$ slot. Similarly, a node {\it receiving} in the $i^{th}$ slot

304: is denoted by $r_{i,j}$.}. Note that $t_{i,j}, r_{i,j} \in

305: \{1,\ldots,N\}$ and $M_i = |{\mathcal S}_i|$.  The SINR at receiver

306: $r_{i,j}$ is given by

307: \begin{eqnarray}

308: {\mbox{SINR}}_{r_{i,j}} &=& 

309:  \frac{\frac{P}{D^{\alpha}(t_{i,j},r_{i,j})}}{N_0+\sum_{\stackrel{k=1}{k\neq j}}^{M_i}\frac{P}{D^{\alpha}(t_{i,k},r_{i,j})}}

310: \label{xvmyw}

311: \end{eqnarray}

312: We define the signal to noise ratio (SNR) at receiver $r_{i,j}$ by

313: \begin{eqnarray}

314: {\mbox{SNR}}_{r_{i,j}} &=& \frac{P}{N_0D^{\alpha}(t_{i,j},r_{i,j})}

315: \label{xklfy}

316: \end{eqnarray}

317: 

318: \subsection{Physical and Protocol Interference Models}

319: 

320: According to the {\it physical interference model}

321: \cite{gupta_kumar__capacity_wireless}, $t_{i,j}\rightarrow r_{i,j}$ is

322: successful if and only if (iff) the SINR at receiver $r_{i,j}$ is

323: greater than or equal to a certain threshold $\gamma_c$, termed as the

324: communication threshold.

325: \begin{eqnarray}

326: \frac{\frac{P}{D^\alpha(t_{i,j},r_{i,j})}}{N_0+\sum_{\stackrel{k=1}{k\neq j}}^{M_i} \frac{P}{D^\alpha(t_{i,k},r_{i,j})}} \geqslant \gamma_c

327: \label{mnmnx}

328: \end{eqnarray}

329: 

330: \noindent According to the {\it protocol interference model}

331: \cite{gupta_kumar__capacity_wireless},

332: $t_{i,j}\rightarrow r_{i,j}$ is successful if:

333: \begin{enumerate}

334: \item

335: the SNR at receiver $r_{i,j}$ is no less than the communication

336: threshold $\gamma_c$.  From (\ref{xklfy}), this translates to

337: \begin{eqnarray}

338: D(t_{i,j},r_{i,j}) &\leqslant& \left(\frac{P}{N_0 \gamma_c}\right)^\frac{1}{\alpha}

339:  \;\;=:\;\; R_c

340: \label{skbiq}

341: \end{eqnarray}

342: where $R_c$ is termed as communication range. 

343: 

344: \item

345: the signal from any unintended transmitter $t_{i,k}$ is received at

346: $r_{i,j}$ with an SNR less than a certain threshold $\gamma_i$, termed

347: as the interference threshold. This translates to

348: \begin{multline}

349: D(t_{i,k},r_{i,j}) 

350: \geqslant \left(\frac{P}{N_0 \gamma_i}\right)^\frac{1}{\alpha}

351:  =: R_i \\ \;\;\forall\;\; k=1,\ldots,M_i, \;k \neq j 

352: \label{djspd}

353: \end{multline}

354: where $R_i$ is termed as interference range. Note that $0 < \gamma_i <

355: \gamma_c$, thus $R_i > R_c$.

356: 

357: \end{enumerate}

358: The physical model of our system is denoted by $\Phi(N,({\mathbf

359: r}_1,\ldots,{\mathbf r}_N),P, \gamma_c,\gamma_i,\alpha,N_0)$.  

360: 

361: A schedule $\Psi(\cdot)$ is {\it feasible} if it satisfies the

362: following:

363: \begin{enumerate}

364: \item

365: Operational constraint: A node must not perform multiple operations in

366: a single time slot.

367: \begin{multline}

368: \{t_{i,j},r_{i,j}\} \cap \{t_{i,k},r_{i,k}\} = \phi \;\;\forall\;\;

369: i=1,\ldots,C \\\;\;\forall\;\; 1 \leqslant j < k \leqslant M_i

370: \label{xvnid}

371: \end{multline}

372: 

373: \item

374: Communication range constraint: Every receiver is within the

375: communication range of its intended transmitter.

376: \begin{multline}

377: D(t_{i,j},r_{i,j}) \leqslant R_c \;\forall i=1,\ldots,C \;\forall

378:  j=1,\ldots,M_i

379: \label{jdvot}

380: \end{multline}

381: 

382: \end{enumerate}

383: 

384: A schedule $\Psi(\cdot)$ is {\it exhaustive} if it satisfies the

385: following:

386: \begin{multline}

387: D(j,k) \leqslant R_c \Rightarrow j \rightarrow k \in \bigcup_{i=1}^C {\mathcal S}_i

388:  \;\;\mbox{and}\;\; k \rightarrow j \in \bigcup_{i=1}^C {\mathcal S}_i

389:  \\\;\; \forall \;\; 1 \leqslant j < k \leqslant N

390: \label{vuufl}

391: \end{multline}

392: 

393: A schedule $\Psi(\cdot)$ is {\it conflict-free} in terms of SINR, if

394: the SINR at every intended receiver does not drop below the

395: communication threshold.

396: \begin{eqnarray}

397: \mbox{SINR}_{r_{i,j}} \geqslant \gamma_c \;\;\forall\;\; i=1,\ldots,C, 

398: \;\;\forall\;\; j=1,\ldots,M_i

399: \label{pozxm}

400: \end{eqnarray}

401: 

402: \subsection{Graph-Based Scheduling}

403: 

404: The traditional approach in designing reuse schedules is to use a

405: graph model of the network and study the set of edges 

406: \cite{behzad_rubin__performance_graph}

407: \cite{gronkvist_hansson__comparison_between}.  The STDMA

408: network $\Phi(\cdot)$ is modeled by a directed graph ${\mathcal

409: G}({\mathcal V},{\mathcal E})$, where $\mathcal V$ is the set of

410: vertices and $\mathcal E$ is the set of edges. Let $\mathcal V =

411: \{v_1,v_2,\ldots,v_N\}$, where vertex $v_j$ represents the $j^{th}$

412: node in $\Phi(\cdot)$. In general, ${\mathcal E} = {\mathcal E}_c

413: \,\dotcup\, {\mathcal E}_i$, where ${\mathcal E}_c$ and ${\mathcal

414: E}_i$ denote the set of communication and interference edges

415: respectively. If node $k$ is within node $j$'s communication range,

416: then there is a communication edge from $v_j$ to $v_k$, denoted by

417: $v_j \stackrel{c}{\rightarrow} v_k$. If node $k$ is outside node $j$'s

418: communication range but within its interference range, then there is

419: an interference edge from $v_j$ to $v_k$, denoted by $v_j

420: \stackrel{i}{\rightarrow} v_k$. Thus, the mapping from $\Phi(\cdot)$

421: to $\mathcal G(\cdot)$ can be described as follows:

422: 

423: \begin{eqnarray*}

424: D(j,k) \leqslant R_c &\Rightarrow& v_j \stackrel{c}{\rightarrow} v_k \in 

425:  {\mathcal E}_c \;\;\mbox{and}\;\; v_k \stackrel{c}{\rightarrow} v_j \in 

426:  {\mathcal E}_c\\

427: R_c < D(j,k) \leqslant R_i &\Rightarrow& v_j \stackrel{i}{\rightarrow}  v_k \in 

428:  {\mathcal E}_i \;\;\mbox{and}\;\; v_k \stackrel{i}{\rightarrow}  v_j \in 

429:  {\mathcal E}_i

430: \end{eqnarray*}

431: 

432: A communication or an interference edge from $v_j$ to $v_k$ will be

433: denoted by $v_j \rightarrow v_k$.  The subgraph ${\mathcal

434: G}_c({\mathcal V},{\mathcal E}_c)$ consisting of communication edges

435: only is termed as the {\it communication graph}.

436: 

437: The schedule $\Psi(\cdot)$ is then designed from the graph $\mathcal

438: G(\cdot)$. Specifically, an STDMA link scheduling algorithm is

439: equivalent to assigning a unique color to every communication edge in

440: the graph, such that source-destination pairs corresponding to

441: communication edges with the same color transmit simultaneously in a

442: particular time slot. The traditional method for link assignment

443: requires that two communication edges $v_i \stackrel{c}{\rightarrow}

444: v_j$ and $v_k \stackrel{c}{\rightarrow} v_l$ can be colored the same

445: iff:

446: \begin{enumerate}

447: \renewcommand{\theenumi}{\roman{enumi}}

448: \item

449: vertices $v_i$, $v_j$, $v_k$, $v_l$ are all mutually distinct, i.e.,

450: there is no {\it primary edge conflict,} and

451: \label{lgcnn}

452: 

453: \item

454: $v_i \rightarrow v_l \not\in \mathcal G(\cdot)$ and $v_k \rightarrow v_j

455: \not\in \mathcal G(\cdot)$, i.e, there is no {\it secondary edge conflict}.

456: \label{plbbz}

457: \end{enumerate}

458: The first criterion is based on the operational constraint. The second

459: criterion states that a node cannot receive a packet while neighboring

460: nodes are transmitting.

461: 

462: Graph-Based scheduling algorithms utilize various graph coloring

463: methodologies to obtain a non-conflicting schedule, i.e., a schedule

464: devoid of primary and secondary edge conflicts.  To maximize the

465: throughput of an STDMA network, graph-based scheduling algorithms seek

466: to minimize the total number of colors used to color all the

467: communication edges of $\mathcal G(\cdot)$.

468: 

469: \subsection{Limitations of Graph-Based Scheduling Algorithms}

470: \label{bcprd}

471: 

472: Observe that Criteria \ref{lgcnn}) and \ref{plbbz}) are not sufficient

473: to guarantee that the resulting schedule $\Psi(\cdot)$ is

474: conflict-free. The link assignments that fulfill the above criteria do

475: not necessarily satisfy the SINR condition (\ref{pozxm}).

476: 

477: Importantly, graph-based scheduling algorithms do not maximize the

478: throughput of an STDMA network because:

479: \begin{enumerate}

480: 

481: \item

482: \label{mxfiw}

483: Due to hard-thresholding based on communication and interference

484: radii, graph-based scheduling algorithms can lead to high cumulative

485: interference at a receiver \cite{behzad_rubin__performance_graph}

486: \cite{gronkvist_hansson__comparison_between}.  This is because the

487: SINR at receiver $r_{i,j}$ decreases with an increase in the number of

488: concurrent transmissions $M_i$, while $R_c$ and $R_i$ have been

489: defined for a single transmission only.

490: \begin{figure}[thbp]

491: \centering

492: \includegraphics[width=3.4in]{reason1.eps}

493: \caption{Graph-Based algorithms can lead to high cumulative interference.}

494: \label{nvxye}

495: \end{figure}

496: For example, consider Figure \ref{nvxye} with six labeled nodes whose

497: coordinates are $1 \equiv (-360,0)$, $2 \equiv (-450,0)$, $3 \equiv

498: (90,0)$, $4 \equiv (0,0)$, $5 \equiv (360,0)$ and $6 \equiv (450,0)$.

499: The system parameters are $P=10$ mW, $\alpha = 4$, $N_0=-90$ dBm,

500: $\gamma_c = 20$ dB and $\gamma_i=10$ dB, which yields $R_c=100$ m and

501: $R_i=177.8$ m.  A graph-based scheduling algorithm will typically

502: schedule the transmissions $1 \rightarrow 2$, $3 \rightarrow 4$ and $5

503: \rightarrow 6$ in the same time slot, say the $i^{th}$ time slot, since

504: the resulting graph coloring is devoid of primary and secondary edge

505: conflicts. However, our computations show that the SINRs at receivers

506: $r_{i,1}$, $r_{i,2}$ and $r_{i,3}$ are $21.26$ dB, $18.42$ dB and

507: $19.74$ dB respectively.  From the physical interference model,

508: transmission $t_{i,1} \rightarrow r_{i,1}$ is successful, while

509: transmissions $t_{i,2} \rightarrow r_{i,2}$ and $t_{i,3} \rightarrow

510: r_{i,3}$ are unsuccessful. This leads to low throughput.

511: 

512: \item

513: On the other hand, graph-based scheduling algorithms can be extremely

514: conservative and result in a higher number of colors

515: \cite{moscibroda_wattenhofer_weber__protocol_design}.

516: \begin{figure}[thbp]

517: \centering

518: \includegraphics[width=2.4in]{reason2.eps}

519: \caption{Graph-Based algorithms can lead to higher number of colors.}

520: \label{hwerd}

521: \end{figure}

522: For example, with the same system parameters as in \ref{mxfiw}),

523: consider Figure \ref{hwerd} with four labeled nodes whose coordinates

524: are $1 \equiv (0,0)$, $2 \equiv (50,0)$, $3 \equiv (220,0)$ and $4

525: \equiv (170,0)$. Assume there are two transmission requests: $1

526: \rightarrow 2$ and $3 \rightarrow 4$.  If both the transmissions are

527: scheduled in the same slot, say the $i^{th}$ time slot, our

528: computations show that the SINRs at receivers $r_{i,1}$ and $r_{i,2}$

529: are both equal to $20.91$ dB. From the physical interference model,

530: both transmissions $t_{i,1} \rightarrow r_{i,1}$ and $t_{i,2}

531: \rightarrow r_{i,2}$ are successful, since signals levels are so high

532: at the receivers that strong interferences can be tolerated. However,

533: due to secondary edge conflicts, a graph-based scheduling algorithm

534: will schedule the above transmissions in different slots, thus

535: decreasing the throughput.

536: 

537: \item

538: Graph-based scheduling algorithms are not geography-aware, i.e., they

539: determine a schedule without being cognizant of the exact positions of

540: the transmitters and receivers.

541: 

542: \end{enumerate}

543: 

544: \subsection{Problem Formulation}

545: \label{pthme}

546: 

547: In STDMA, we construct a graph model ${\mathcal G}({\mathcal

548: V},{\mathcal E}_c \,\dotcup\, {\mathcal E}_i)$ of the physical network

549: $\Phi(\cdot)$.  The communication graph ${\mathcal G}_c({\mathcal

550: V},{\mathcal E}_c)$ is an approximation of $\Phi(\cdot)$, while the

551: two-tier graph ${\mathcal G}({\mathcal V},{\mathcal E}_c \,\dotcup\,

552: {\mathcal E}_i)$ is a better approximation of $\Phi(\cdot)$.  From

553: $\Phi(\cdot)$ and ${\mathcal G}_c(\cdot)$, one can exhaustively

554: determine the STDMA schedule which yields the highest throughput

555: according to the physical interference model. However, this is a

556: combinatorial optimization problem of prohibitive complexity

557: $(O(|{\mathcal E}_c|^{|{\mathcal E}_c|}))$ and is thus computationally

558: infeasible.

559: 

560: To overcome these problems, we propose a new suboptimal algorithm for

561: STDMA link scheduling based on a more realistic physical interference

562: model.  Our algorithm is based on the communication graph model

563: ${\mathcal G}_c({\mathcal V},{\mathcal E}_c)$ as well as SINR

564: computations.

565: 

566: To evaluate the performance of our algorithm and compare it with

567: existing suboptimal STDMA link scheduling algorithms, we define the

568: following metric: spatial reuse.  Consider the STDMA link schedule

569: $\Psi(\cdot)$ for the network $\Phi(\cdot)$. Under the physical

570: interference model, the transmission $t_{i,j} \rightarrow r_{i,j}$ is

571: successful iff (\ref{mnmnx}) is satisfied.  The {\it spatial reuse} of

572: the schedule $\Psi(\cdot)$ is defined as the average number of

573: successfully received packets per time slot in the STDMA

574: schedule. Thus

575: \begin{multline}

576: \mbox{Spatial Reuse $\;=\; \sigma$} =

577:  \frac{\sum_{i=1}^C\sum_{j=1}^{M_i}I({\mbox{SINR}}_{r_{i,j}}\geqslant\gamma_c)}{C} 

578: \label{qpjkk}

579: \end{multline}

580: where $I(A)$ denote the indicator function for event $A$, i.e.,

581: $I(A)=1$ if event $A$ occurs, $I(A)=0$ if event $A$ does not occur.

582: 

583: The essence of STDMA is to have a reasonably large number of

584: concurrent and successful transmissions.  For an STDMA network which

585: is operational for a long period of time, say $L$ slots, the total

586: number of successfully received packets is $L\sigma$. Thus, a high

587: value of spatial reuse\footnote{Note that spatial reuse in our

588:   system model is analogous to spectral efficiency in digital

589:   communication systems.} directly translates to higher long-term total

590: network throughput and the number of colors $C$ is relatively

591: unimportant.  Hence, spatial reuse turns out to be a crucial metric

592: for the comparison of different STDMA algorithms.

593: 

594: We seek low complexity conflict-free STDMA link scheduling algorithms

595: with high spatial reuse. We only consider STDMA schedules which are

596: feasible and exhaustive. Thus, our schedules satisfy (\ref{xvnid}),

597: (\ref{jdvot}), (\ref{vuufl}) and (\ref{pozxm}).

598: 

599: \subsection{Comparison with SINR-Based Scheduling Algorithms}

600: 

601: In cognizance of our system model and performance metric, our work is

602: reasonably different from existing SINR-based STDMA link scheduling

603: algorithms.

604: 

605: The works in \cite{brar_blough_santi__computationally_efficient}

606: \cite{moscibroda_wattenhofer__complexity_connectivity} focus on

607: minimizing the schedule length, which does not necessarily translate

608: to high network throughput. On the other hand, spatial reuse directly

609: corresponds to network throughput capacity

610: \cite{gupta_kumar__capacity_wireless}.  Power-controlled algorithms

611: can lead to excessively high transmit power (for example, Line 16 in

612: Algorithm 1 \cite{moscibroda_wattenhofer__complexity_connectivity}),

613: which is impractical since all wireless routers have constraints on

614: maximum transmit power. So, similar to

615: \cite{kodialam_nandagopal__characterizing_achievable}, we consider

616: uniform transmit power at all wireless routers.  Existing works on

617: SINR-based link scheduling, which are usually in the context of 802.11

618: wireless networks \cite{brar_blough_santi__computationally_efficient}

619: \cite{kim_lim_hou__improving_spatial} and wireless mesh networks

620: \cite{lim_lim_hou__coordinate_based}

621: \cite{kodialam_nandagopal__characterizing_achievable}, consider many

622: practical aspects of the underlying communication protocol and network

623: architecture. Consequently, their system models are quite different

624: from our pure STDMA network model.

625: 

626: To the best of our knowledge, this is the first attempt to develop a

627: centralized algorithm for SINR-based link scheduling in a pure STDMA

628: wireless ad hoc network with uniform power assignment.  Also, this

629: work is different from previous works due to the focus on spatial

630: reuse.

631: 

632: Hence, we compare the performance of our algorithm with existing

633: graph-based algorithms only.

634: 

635: \section{SINR-Based Link Scheduling Algorithm}

636: \label{lxvym}

637: 

638: \subsection{Structure}

639: 

640: We first describe the essential features of STDMA link scheduling

641: algorithms. The core of every link scheduling algorithm consists

642: of the following modules:

643: \begin{enumerate}

644: 

645: \item

646: An order in which communication edges are considered for coloring.

647: 

648: \item

649: A function which determines the set of all existing colors which can

650: be assigned to the edge under consideration without violating the

651: problem constraints.

652: 

653: \item

654: A {\it BestColor} rule to determine which conflict-free color to

655: assign to the edge under consideration.

656: \end{enumerate}

657: 

658: The second module considers only operational and communication range

659: constraints in graph-based scheduling algorithms. However, in

660: the SINR-based link scheduling algorithm that we propose, SINR

661: constraints are also taken into account. Note that this function is

662: completely described by the problem definition and does not vary from

663: algorithm to algorithm. The ordering of edges for coloring and the

664: BestColor rule play a significant role in determining the performance

665: and computational complexity of an STDMA scheduling algorithm.

666: 

667: \subsection{Motivation}

668: 

669: Recall that graph-based models are inadequate to design efficient link

670: schedules under the physical interference model and brute-force

671: computation of an optimal link schedule that maximizes spatial reuse

672: is prohibitively complex (see Sections \ref{bcprd} and \ref{pthme}).

673: Motivated by techniques from matroid theory

674: \cite{lawler__combinatorial_optimization}, we develop a

675: computationally feasible algorithm with demonstrably high spatial

676: reuse.  The essence of our algorithm is to partition the set of

677: communication edges into subsets (forests) and color the edges in each

678: subset sequentially. The edges in each forest are considered in a

679: random order for coloring, since randomized algorithms are known to

680: outperform deterministic algorithms, esp. when the characteristics of

681: the input are not known apriori

682: \cite{motwani_raghavan__randomized_algorithms}.

683: 

684: A similar matroid-based network partitioning technique is used in

685: \cite{brzezinski_zussman_modiano__enabling_distributed} to generate

686: high capacity subnetworks for a distributed throughput maximization

687: problem in wireless mesh networks.  Techniques from matroid theory

688: have also been employed to develop efficient heuristics for NP-hard

689: combinatorial optimization problems in fields such as distributed

690: computer systems \cite{ramalingom_thulasiraman_das__matroid_theoretic}

691: and linear network theory \cite{petersen__investigating_solvability}.

692: 

693: \subsection{ConflictFreeLinkSchedule Algorithm}

694: \label{xbpyf}

695: 

696: Our proposed SINR-based link scheduling algorithm is

697: ConflictFreeLinkSchedule, which considers the communication graph

698: ${\mathcal G}_c({\mathcal V},{\mathcal E}_c)$ and is described in

699: Algorithm \ref{hidwo}.

700: 

701: In Phase 1 (Line 3), we label all the vertices randomly. Specifically,

702: if ${\mathcal G}_c(\cdot)$ has $v$ vertices, we perform a random

703: permutation of the sequence $(1,2,\ldots,v)$ and assign these labels

704: to vertices with indices $1,2,\ldots,v$ respectively.

705: 

706: In Phase 2 (Line 4), the communication graph ${\mathcal G}_c(\cdot)$

707: is decomposed into what are called as out-oriented and in-oriented

708: graphs $T_1,T_2,\ldots,T_k$

709: \cite{ramanathan_lloyd__scheduling_algorithms}. Each $T_i$ is a forest

710: and every edge of ${\mathcal G}_c(\cdot)$ is in exactly one of the

711: $T_i$'s. This decomposition is achieved by partitioning graph

712: $G_c(\cdot)$, the undirected equivalent of ${\mathcal G}_c(\cdot)$,

713: into undirected forests. The number of forests can be minimized by

714: using techniques from Matroid theory ($k$-forest problem,

715: \cite{gabow_westermann__forests_frames}). However, this optimal

716: decomposition requires extensive computation. Hence, we adopt the

717: speedier albeit non-optimal approach of using successive breadth first

718: searches to decompose $G_c(\cdot)$ into undirected forests. Each

719: undirected forest is further mapped to two directed forests. In one

720: forest, the edges in every connected component point away from the

721: root and every vertex has at most one incoming edge, thus producing an

722: out-oriented graph. In the other forest, the edges in every connected

723: component point toward the root and every vertex has at most one

724: outgoing edge, thus producing an in-oriented graph\footnote{An

725:   in-oriented graph is also constructed by Algorithm 1 in

726:   \cite{moscibroda_wattenhofer__complexity_connectivity} to determine

727:   a link schedule in a power-controlled STDMA network.}.

728: 

729: In Phase 3 (Lines 5-14), the oriented graphs are considered

730: sequentially.  For each oriented graph, vertices are considered in

731: increasing order by label\footnote{In essence, the edges are scanned

732:   in a random order, since labeling is random.} and the unique edge

733: associated with each vertex is colored using the

734: FirstConflictFreeColor function.

735: 

736: \begin{algorithm}

737: \caption{ConflictFreeLinkSchedule}

738: \label{hidwo}

739: \begin{algorithmic}[1]

740: \STATE {\bf input:} Physical network

741: $\Phi(\cdot)$, communication graph ${\mathcal G}_c(\cdot)$

742: \STATE {\bf output:} A coloring $C: {\mathcal E}_c \rightarrow \{1,2,\ldots\}$

743: \STATE label the vertices of ${\mathcal G}_c$ randomly

744: \STATE use successive breadth first searches to partition ${\mathcal G}_c$ into oriented graphs

745: $T_i$, $1 \leqslant i \leqslant k$

746: \FOR{$i \leftarrow 1 \mbox{ to } k$}

747: \FOR{$j \leftarrow 1 \mbox{ to } n$}

748: \IF{$T_i$ is out-oriented}

749: \STATE let $x=(s,d)$ be such that $L(d)=j$

750: \ELSE

751: \STATE let $x=(s,d)$ be such that $L(s)=j$

752: \ENDIF

753: \STATE $C(x) \leftarrow \mbox{FirstConflictFreeColor}(x)$

754: \ENDFOR

755: \ENDFOR

756: \end{algorithmic}

757: \end{algorithm}

758: 

759: The FirstConflictFreeColor function is explained in Algorithm

760: \ref{pecdi}. For the edge under consideration $x$, it discards any

761: color that has an edge with a primary conflict with $x$. Now, we

762: consider the residual set of conflict-free colors sequentially. We

763: choose the first conflict-free color such that the resulting SINRs at

764: the receiver of $x$ and the receivers of all co-colored edges exceed

765: the communication threshold $\gamma_c$. If no such color is found, we

766: assign a new color to $x$. Thus, this function guarantees that the

767: ensuing schedule is conflict-free in terms of SINR. Since we choose

768: the first SINR-compliant color and not the `best' SINR-compliant color

769: according to some BestColor rule, the computational complexity of this

770: function is lower than that of any other function which checks the

771: SINRs at the receivers of all co-colored edges.

772: 

773: \begin{algorithm}

774: \caption{integer FirstConflictFreeColor($x$)}

775: \label{pecdi}

776: \begin{algorithmic}[1]

777: \STATE {\bf input:} Physical network $\Phi(\cdot)$, communication graph 

778: ${\mathcal G}_c(\cdot)$

779: \STATE {\bf output:} A conflict-free color

780: \STATE ${\mathcal C} \leftarrow \mbox{set of existing colors}$

781: \STATE ${\mathcal C}_c \leftarrow \{C(h):h \in {\mathcal E}_c$, $h$ is colored, $x$ and $h$ have a primary edge conflict$\}$

782: \STATE ${\mathcal C}_{cf} = {\mathcal C} \setminus {\mathcal C}_c$

783: \FOR{$i \leftarrow 1 \mbox{ to } |{\mathcal C}_{cf}|$}

784: \STATE $r \leftarrow i^{th} \mbox{ color in } {\mathcal C}_{cf}$

785: \STATE $E_r \leftarrow \{h:h \in {\mathcal E}_c, C(h)=r\}$

786: \STATE $C(x) \leftarrow r$

787: \IF{SINR at all receivers of $E_r \cup \{x\}$ exceed $\gamma_c$}

788: \STATE return $r$

789: \ENDIF

790: \ENDFOR

791: \STATE return $|\mathcal C|+1$

792: \end{algorithmic}

793: \end{algorithm}

794: 

795: \section{Performance Results}

796: \label{cymoe}

797: 

798: \subsection{Simulation Model}

799: \label{zxptb}

800: 

801: In our simulation experiments, the location of every node is generated

802: randomly, using a uniform distribution for its $X$ and $Y$

803: coordinates, in the deployment area. For a fair comparison of our

804: algorithm with the Truncated Graph-Based Scheduling Algorithm

805: \cite{behzad_rubin__performance_graph}, we assume that the deployment

806: region is a circular region of radius $R$. Thus, if $(X_j,Y_j)$ are

807: the Cartesian coordinates of the $j^{th}$ node, $j=1,\ldots,N$, then

808: $X_j \sim U[-R,R]$ and $Y_j \sim U[-R,R]$ subject to $X_j^2 + Y_j^2

809: \leqslant R^2$. Equivalently, if $(R_j,\Theta_j)$ are the polar

810: coordinates of the $j^{th}$ node, then $R_j^2 \sim U[0,R^2]$ and

811: $\Theta_j \sim U[0,2\pi]$.  After generating random positions for $N$

812: nodes, we have complete information of $\Phi(\cdot)$. Using

813: (\ref{skbiq}) and (\ref{djspd}), we compute the communication and

814: interference radii, and then map the network $\Phi(\cdot)$ to the

815: two-tier graph ${\mathcal G}({\mathcal V},{\mathcal E}_c \,\dotcup\,

816: {\mathcal E}_i)$.  Once the link schedule is computed by every

817: algorithm, the spatial reuse is computed using (\ref{qpjkk}).  We use

818: two sets of values for system parameters $P$, $\gamma_c$, $\gamma_i$,

819: $\alpha$, $N_0$, $N$ and $R$, which are prototypical values of system

820: parameters in wireless networks \cite{kim_lim_hou__improving_spatial}

821: and describe them in Section \ref{mclgs}.  For a given set of system

822: parameters, we calculate the spatial reuse by averaging this quantity

823: over one thousand randomly generated networks.  Keeping all other

824: parameters fixed, we observe the effect of increasing the number of

825: nodes $N$ on the spatial reuse $\sigma$.

826: 

827: In our experiments, we compare the performance of the following algorithms:

828: \begin{enumerate}

829: \item

830: ArboricalLinkSchedule \cite{ramanathan_lloyd__scheduling_algorithms}

831: (ALS)

832: 

833: \item 

834:   Truncated Graph-Based Scheduling Algorithm\footnote{In Truncated

835:     Graph-Based Scheduling Algorithm, for the computation of optimal

836:     number of transmissions $M^*$, we follow the exact method

837:     described in \cite{behzad_rubin__performance_graph}. Since $0 <

838:     \xi < \frac{N_0}{P}$, we assume that $\xi = 0.9999\frac{N_0}{P}$

839:     and compute successive Edmundson-Madansky (EM) upper bounds

840:     \cite{madansky__inequalities_stochastic}

841:     \cite{dokov_morton__higher_order} till the difference between

842:     successive EM bounds is less than $0.3\%$. We have experimentally

843:     verified that only high values of $\xi$ lead to reasonable values

844:     for $M^*$, whereas low values of $\xi$, say $\xi =

845:     0.1\frac{N_0}{P}$, lead to the extremely conservative value of

846:     $M^*=1$ in most cases.}  \cite{behzad_rubin__performance_graph}

847:   (TGSA)

848: 

849: \item

850: ConflictFreeLinkSchedule (CFLS)

851: 

852: \end{enumerate}

853: 

854: \subsection{Performance Comparison}

855: \label{mclgs}

856: 

857: In our first set of experiments (Experiment 1), we assume that $R=500$

858: m, $P=10$ mW, $\alpha = 4$, $N_0=-90$ dBm, $\gamma_c = 20$ dB and

859: $\gamma_i=10$ dB. Thus, $R_c=100$ m and $R_i=177.8$ m.  We vary the

860: number of nodes from 30 to 110 in steps of 5. Figure \ref{zowyx} plots

861: the spatial reuse vs. number of nodes for all the algorithms.

862: 

863: \begin{figure}[thbp]

864: \centering

865: \includegraphics[width=3.6in]{sr_expt1.eps}

866: \caption{Spatial reuse vs. number of nodes for Experiment 1.}

867: \label{zowyx}

868: \end{figure}

869: 

870: In our second set of experiments (Experiment 2), we assume that

871: $R=700$ m, $P=15$ mW, $\alpha = 4$, $N_0=-85$ dBm, $\gamma_c = 15$ dB

872: and $\gamma_i=7$ dB. Thus, $R_c=110.7$ m and $R_i=175.4$ m.  We vary

873: the number of nodes from 70 to 150 in steps of 5. Figure \ref{uusvr}

874: plots the spatial reuse vs. number of nodes for all the algorithms.

875: 

876: \begin{figure}[thbp]

877: \centering

878: \includegraphics[width=3.6in]{sr_expt2.eps}

879: \caption{Spatial reuse vs. number of nodes for Experiment 2.}

880: \label{uusvr}

881: \end{figure}

882: 

883: For the ALS algorithm, we observe that spatial reuse increases very

884: slowly with increasing number of nodes.

885: 

886: For the TGSA algorithm, we observe that spatial reuse is $18$-$27\%$

887: lower than that of ALS. A plausible explanation for this behavior is

888: as follows.  The basis for TGSA is the computation of $M^*$, the

889: optimal number of transmissions in every slot.  $M^*$ is determined by

890: maximizing a lower bound on the expected number of successful

891: transmissions in a time slot.  Since the partitioning of a maximal

892: independent set of communication arcs into subsets of cardinality at

893: most $M^*$ is arbitrary and not geography-based, there could be

894: scenarios where the transmissions scheduled in a subset are in the

895: vicinity of each other, resulting in moderate to high interference.

896: In essence, maximizing this lower bound does not necessarily translate

897: to maximizing the number of successful transmissions in a time slot.

898: Also, due to its design, the TGSA algorithm yields higher number of

899: colors compared to ALS.

900: 

901: For our CFLS algorithm, we observe that spatial reuse increases

902: steadily with increasing number of nodes and is consistently

903: $25$-$50\%$ higher than the spatial reuse of ALS and TGSA.

904: 

905: \subsection{Performance Comparison under Realistic Channel

906: Conditions}

907: 

908: In a realistic wireless environment, channel impairments like

909: multipath fading and shadowing affect the received SINR at a receiver

910: \cite{sklar__rayleigh_fading}.  In this section, we compare the

911: performance of the ALS, TGSA and CFLS algorithms in a wireless channel

912: which experiences Rayleigh fading and lognormal shadowing.

913: 

914: In the absence of fading and shadowing, the SINR at receiver $r_{i,j}$

915: is given by (\ref{xvmyw}). We assume that every algorithm (ALS, TGSA

916: and CFLS) considers only path loss in the channel prior to

917: constructing the two-tier graph ${\mathcal G}({\mathcal V},{\mathcal

918:   E}_c \,\dotcup\, {\mathcal E}_i)$ and computing the link schedule.

919: However, when evaluating the performance of each algorithm, we take

920: into account the fading and shadowing channel gains between every pair

921: of nodes. Specifically, for computing the spatial reuse using

922: (\ref{qpjkk}), we assume that the (actual) SINR at receiver $r_{i,j}$

923: is given by

924: \begin{multline*}

925: {\mbox{SINR}}_{r_{i,j}} = 

926:  \frac{\frac{P}{D^{\alpha}(t_{i,j},r_{i,j})}V(t_{i,j},r_{i,j})10^{W(t_{i,j},r_{i,j})}}{N_0+\sum_{\stackrel{k=1}{k\neq j}}^{M_i}\frac{P}{D^{\alpha}(t_{i,k},r_{i,j})}V(t_{i,k},r_{i,j})10^{W(t_{i,k},r_{i,j})}}

927: \end{multline*}

928: where random variables $V(\cdot)$ and $W(\cdot)$ correspond to channel

929: gains due to Rayleigh fading and lognormal shadowing respectively. We

930: assume that $\{V(k,l)|1\leqslant k,l \leqslant N, k \neq l\}$ are

931: independent and identically distributed (i.i.d.) random variables with

932: probability density function\footnote{$u(\cdot)$ is the unit step

933:   function.} (pdf)

934: \cite{tse_viswanath__fundamentals_wireless}

935: \begin{eqnarray*}  

936: f_V(v) &=& \frac{1}{\sigma_V^2}e^{\frac{-v}{\sigma_V^2}}u(v)

937: \end{eqnarray*}

938: and $\{W(k,l)|1\leqslant k,l \leqslant N, k\neq l\}$ are i.i.d. zero mean

939: Gaussian random variables with pdf

940: \cite{goldsmith__wireless_communications}

941: \begin{eqnarray*}

942: f_W(w) &=& \frac{1}{\sqrt{2\pi}\sigma_W}e^{\frac{-w^2}{2\sigma_W^2}}

943: \end{eqnarray*}

944: Random variables $V(\cdot)$ and $W(\cdot)$ are independent of each

945: other and also independent of the node locations.

946: 

947: Our simulation model and experiments are exactly as described in

948: Sections \ref{zxptb} and \ref{mclgs}. In our simulations, we assume

949: $\sigma_V^2 = \sigma_W^2 = 1$.  For Experiment 1, Figure \ref{cvcvb}

950: plots the spatial reuse vs. number of nodes for all the algorithms.

951: For Experiment 2, Figure \ref{vcvcb} plots spatial reuse vs. number of

952: nodes for all the algorithms.

953: 

954: \begin{figure}[thbp]

955: \centering

956: \includegraphics[width=3.6in]{f_sr_expt1.eps}

957: \caption{Spatial reuse vs. number of nodes for Experiment 1 under 

958: multipath fading and shadowing channel conditions.}

959: \label{cvcvb}

960: \end{figure}

961: 

962: \begin{figure}[thbp]

963: \centering

964: \includegraphics[width=3.6in]{f_sr_expt2.eps}

965: \caption{Spatial reuse vs. number of nodes for Experiment 2 under 

966: multipath fading and shadowing channel conditions.}

967: \label{vcvcb}

968: \end{figure}

969: 

970: From Figures \ref{zowyx}, \ref{uusvr}, \ref{cvcvb} and \ref{vcvcb}, we

971: observe that spatial reuse decreases by $20$-$40\%$ in a channel

972: experiencing multipath fading and shadowing effects. A plausible

973: explanation for this observation is as follows. Since the channel

974: gains between every pair of nodes are independent of each other, it is

975: reasonable to assume that the interference power at a typical receiver

976: remains almost the same as in the non-fading case. This is because,

977: even if the power received from few unintended transmitters is low,

978: the power received from other unintended transmitters will be high (on

979: an average); thus the interference power remains constant.

980: Consequently, the change in SINR is determined by the change in

981: received signal power only. If the received signal power is higher

982: compared to the non-fading case, the transmission is anyway successful

983: and spatial reuse remains unchanged (see (\ref{qpjkk})). However, if

984: the received signal power is lower, the transmission is now

985: unsuccessful and spatial reuse decreases. Hence, on an average, the

986: spatial reuse decreases.

987: 

988: Finally, from Figures \ref{cvcvb} and \ref{vcvcb}, we observe that our CFLS

989: algorithm achieves $5$-$17\%$ higher spatial reuse than the ALS

990: algorithm and $40$-$80\%$ higher spatial reuse than the TGSA

991: algorithm, under realistic channel conditions.

992: 

993: \section{Analytical Results}

994: \label{mbxgt}

995: 

996: In this section, we derive upper bounds on the running time complexity

997: (computational complexity) of the ConflictFreeLinkSchedule algorithm.

998: We will use the following notation with respect to the communication

999: graph ${\mathcal G}_c({\mathcal V},{\mathcal E}_c)$:

1000: \begin{eqnarray*}

1001: e &=& \mbox{number of communication edges}\\

1002: v &=& \mbox{number of vertices}\\

1003: \theta &=& \mbox{thickness of the graph}\\

1004: &:=& \mbox{minimum number of graphs into which the}\\

1005: && \mbox{undirected equivalent of ${\mathcal G}_c(\cdot)$ can be partitioned}

1006: \end{eqnarray*}

1007: 

1008: \begin{figure}

1009: \centering

1010: \subfigure[Experiment 1]

1011: {

1012: \label{ywhlv}

1013: \includegraphics[width=3.35in]{theta_expt1_1.eps}

1014: }

1015: \hspace{0.2in}

1016: \subfigure[Experiment 2]

1017: {

1018: \label{drjnb}

1019: \includegraphics[width=3.35in]{theta_expt2.eps}

1020: }

1021: \caption{Comparison of thickness and number of edges with number of vertices.}

1022: \label{wvlpt}

1023: \end{figure}

1024: 

1025: Before we prove our results, it is instructive to observe Figure

1026: \ref{wvlpt}, which shows the variation of $\theta$ and $e$ with $v$

1027: for the two experiments described in Section \ref{mclgs}.  Since

1028: determining the thickness of a graph is NP-hard

1029: \cite{mutzel_odenthal__thickness_graphs}, each value of $\theta$ in

1030: Figure \ref{wvlpt} is an upper bound on the actual thickness based on

1031: the number of forests into which the undirected equivalent of the

1032: communication graph has been decomposed using successive breadth first

1033: searches.  We observe that the graph thickness increases very slowly

1034: with the number of vertices, while the number of edges increases

1035: super-linearly with the number of vertices.

1036: 

1037: \begin{lemma}

1038: \label{apmxr}

1039: An oriented graph $T$ can be colored using no more than $O(v)$ colors

1040: using ConflictFreeLinkSchedule.

1041: \end{lemma}

1042: \begin{proof}

1043: Since an oriented graph with $v$ vertices has at most $v$ edges, the

1044: edges of $T$ can be colored with at most $v$ colors.

1045: \end{proof}

1046: 

1047: \vspace{1.5ex}

1048: 

1049: \begin{remark}

1050: The number of colors obtained by our algorithm depends not only on the

1051: graph representation of the STDMA network, but also on the positions

1052: of the nodes and the values of $P$, $\gamma_c$, $\gamma_i$, $\alpha$

1053: and $N_0$. Since our algorithm is both graph-based and SINR-based

1054: (hybrid), it is not possible to obtain a tighter upper bound in Lemma

1055: \ref{apmxr}.

1056: \end{remark}

1057: 

1058: \vspace{1.5ex}

1059: 

1060: \begin{lemma}

1061: For an oriented graph $T$, the running time of

1062: ConflictFreeLinkSchedule is $O(v^2)$.

1063: \label{dolfb}

1064: \end{lemma}

1065: \begin{proof}

1066: Assuming that an element can be chosen randomly and uniformly from a

1067: finite set in unit time (Chapter 1,

1068: \cite{motwani_raghavan__randomized_algorithms}), the running time of

1069: Phase 1 can be shown to be $O(v)$.  Since there is only one oriented

1070: graph, Phase 2 runs in time $O(1)$. In Phase 3, the unique edge

1071: associated with the vertex under consideration is assigned a color

1072: using FirstConflictFreeColor. From Lemma \ref{apmxr}, the size of the

1073: set of colors to be examined $|{\mathcal C}_c \cup {\mathcal C}_{cf}|$

1074: is $O(v)$.  In FirstConflictFreeColor, the SINR is checked only once

1075: for every colored edge in the set $\bigcup_{i=1}^{|{\mathcal C}_{cf}|}

1076: E_i$ and at most $v$ times for the edge under consideration $x$. With

1077: a careful implementation, FirstConflictFreeColor runs in time $O(v)$.

1078: So, the running time of Phase 3 is $O(v^2)$.  Thus, the total running

1079: time is $O(v^2)$.

1080: \end{proof}

1081: 

1082: \vspace{3ex}

1083: 

1084: \begin{theorem}

1085: For an arbitrary graph $\mathcal G$, the running time of

1086: ConflictFreeLinkSchedule is $O(ev\log v + ev\theta)$.

1087: \end{theorem}

1088: \begin{proof}

1089: Assuming that an element can be chosen randomly and uniformly from a

1090: finite set in unit time

1091: \cite{motwani_raghavan__randomized_algorithms}, the running time of

1092: Phase 1 can be shown to be $O(v)$.  For Phase 2, the optimal

1093: partitioning technique of \cite{gabow_westermann__forests_frames}

1094: based on Matroids can be used to partition the communication graph

1095: ${\mathcal G}_c$ into at most $6\theta$ oriented graphs in time

1096: $O(ev\log v)$.  Thus, $k \leqslant 6\theta$ holds for Phase 3.  From

1097: Lemma \ref{dolfb}, it follows that the first oriented graph $T_1$ can

1098: be colored in time $O(v^2)$. However, consider the coloring of the

1099: $j^{th}$ oriented graph $T_j$, where $2 \leqslant j \leqslant k$.

1100: When coloring edge $x$ from $T_j$ using FirstConflictFreeColor,

1101: conflicts can occur not only with the colored edges of $T_j$, but also

1102: with the edges of the previously colored oriented graphs

1103: $T_1,T_2,\ldots,T_{j-1}$.  This fact is exemplified in Appendix

1104: \ref{jdvow}.  Hence, the worst-case size of the set of colors to be

1105: examined $|{\mathcal C}_c \cup {\mathcal C}_{cf}|$ is $O(e)$. Note

1106: that in FirstConflictFreeColor, the SINR is checked only once for

1107: every colored edge in the set $\bigcup_{i=1}^{|{\mathcal C}_{cf}|}

1108: E_i$ and at most $e$ times for the edge under consideration $x$. With

1109: a careful implementation, FirstConflictFreeColor runs in time $O(e)$.

1110: Hence, any subsequent oriented graph $T_j$ can be colored in time

1111: $O(ev)$. Thus, the running time of Phase 3 is $O(ev\theta)$.

1112: Therefore, the overall running time of ConflictFreeLinkSchedule is

1113: $O(ev\log v + ev\theta)$.

1114: \end{proof}

1115: 

1116: \section{Discussion}

1117: \label{eiovv}

1118: 

1119: In this paper, we have developed ConflictFreeLinkSchedule, an

1120: SINR-based link scheduling algorithm for STDMA multihop wireless ad

1121: hoc networks under the physical interference model.  The performance

1122: of our algorithm is superior to existing link scheduling algorithms

1123: for STDMA networks with uniform power assignment.  A practical

1124: experimental modeling shows that, on an average, our algorithm

1125: achieves $40\%$ higher spatial reuse than the ArboricalLinkSchedule

1126: \cite{ramanathan_lloyd__scheduling_algorithms} and Truncated

1127: Graph-Based Scheduling \cite{behzad_rubin__performance_graph}

1128: algorithms.  Since schedules are constructed offline only once and

1129: then used by the network for a long period of time, these improvements

1130: in performance directly translate to higher long-term network

1131: throughput.

1132: 

1133: The computational complexity of ConflictFreeLinkSchedule is comparable

1134: to the computational complexity of ArboricalLinkSchedule and is much

1135: lower than the computational complexity of Truncated Graph-Based

1136: Scheduling Algorithm. Thus, in cognizance of spatial reuse as well as

1137: computational complexity, ConflictFreeLinkSchedule is a good candidate

1138: for efficient SINR-based STDMA link scheduling algorithms.

1139: 

1140: We have recently developed computationally efficient algorithms for

1141: STDMA broadcast scheduling under the physical interference model. It

1142: would be interesting to apply techniques like simulated annealing,

1143: genetic algorithms and neural networks to compute high spatial reuse

1144: conflict-free STDMA link schedules.

1145: 

1146: \appendices

1147: \section{Example of Primary Edge Conflicts with Previously Colored Oriented Graphs}

1148: \label{jdvow}

1149: 

1150: \begin{figure}[thbp]

1151: \centering \includegraphics[width=1.35in]{Gc.eps}

1152: \caption{A six-node STDMA wireless ad hoc network, its communication graph and node labels.}

1153: \label{hxcvm}

1154: \end{figure}

1155: 

1156: \begin{figure}[thbp]

1157: \centering

1158: \includegraphics[width=3.5in]{T1_4.eps}

1159: \caption{Decomposition of Fig. \ref{hxcvm} into two out-oriented graphs $(T_1,T_3)$ and two in-oriented graphs $(T_2,T_4)$.}

1160: \label{pndiz}

1161: \end{figure}

1162: 

1163: Consider the six-node STDMA wireless ad hoc network shown in Figure

1164:  \ref{hxcvm}, along with its associated communication graph ${\mathcal

1165:  G}_c(\cdot)$ and node labels.  Using successive breadth first

1166:  searches, ${\mathcal G}_c(\cdot)$ is partitioned into four oriented

1167:  graphs $T_1$, $T_2$, $T_3$ and $T_4$, as shown in Figure \ref{pndiz}.

1168:  A conflict-free coloring of the first oriented graph $T_1$ is shown

1169:  in Table \ref{kdhfo}.  Now, when we color an edge from any other

1170:  oriented graph, we must take into account the colors of the edges in

1171:  $T_1$. For example:

1172: \begin{enumerate}

1173: \item

1174: In $T_2$, Edge $6 \rightarrow 3$ cannot be assigned Color $2$ due to a

1175: primary edge conflict with Edge $1 \rightarrow 3$ of $T_1$.

1176: 

1177: \item

1178: In $T_3$, Edge $5 \rightarrow 6$ cannot be assigned Color $1$ due to a

1179: primary edge conflict with Edge $2 \rightarrow 5$ of $T_1$.

1180: 

1181: \item

1182: In $T_4$, Edge $3 \rightarrow 2$ cannot be assigned Color $3$ due to a

1183: primary edge conflict with Edge $1 \rightarrow 2$ of $T_1$.

1184: 

1185: \end{enumerate}

1186: 

1187: \begin{table}[hbtp]

1188: \begin{tabular}{|c|c|} \hline

1189: Edge of $T_1$ & Color \\ \hline

1190: $3 \rightarrow 6$ & 1 \\ \hline

1191: $2 \rightarrow 5$ & 1 \\ \hline

1192: $1 \rightarrow 3$ & 2 \\ \hline

1193: $1 \rightarrow 2$ & 3 \\ \hline

1194: $2 \rightarrow 4$ & 4 \\ \hline

1195: \end{tabular}

1196: \caption{Conflict-free coloring of first oriented graph $T_1$.}

1197: \label{kdhfo}

1198: \end{table}

1199: 

1200: \bibliographystyle{IEEEtran.bst}

1201: 

1202: \bibliography{s}

1203: 

1204: \end{document}

1205: