1: \documentclass[10pt,a4,twocolumn]{IEEEtran}

2: \usepackage{epsfig,fullpage,times,url,graphicx}

3: 

4: \DeclareGraphicsRule{.JPG}{eps}{*}{`jpeg2ps #1} 

5: 

6: \author{SeonYeong Han and Nael B. Abu-Ghazaleh\\

7: Computer Science Dept.\\

8: State University of New York at Binghamton\\

9: \url{{shan6@,nael@cs.}binghamton.edu}}

10: 

11: \title{On the Effect of Fading on Ad-hoc Networks}

12: %

13: \begin{document}

14: \pagestyle{empty}

15: 				  

16: \maketitle

17: 

18: \begin{abstract}~\label{abstract}

19: 

20: 

21: Most MANET (Mobile Ad hoc NETwork) research assumes idealized

22: propagation models.  Experimental results have shown significant

23: divergence from simulation results due to the effect of signal fading

24: in realistic wireless communication channels.  In this paper, we

25: characterize the impact of fading on protocol performance. We first

26: study the effect of fading on MAC performance and show that its effect

27: can be dominating.  One of our important conclusions is that

28: eliminating RTS/CTS packets results in more effective operation under

29: fading.  We also identify an unfairness problem that arises due to

30: backoffs in the presence of fading.  Moreover, fading results in

31: several subtle interactions between the MAC and routing layers.  We

32: identify several of these problems and make observations about

33: effective approaches for addressing them.  For example, the criteria

34: for determining the best path should not only consider the link status

35: but also the link order. In addition, because routing protocols rely

36: on MAC level transmission failure (when the retry limit is exceeded),

37: route failure errors are often generated unnecessarily. Finally,

38: because MAC level broadcasts are unreliable, they are especially

39: vulnerable to fading. We analyze these effects and outline preliminary

40: solutions to them.

41: \end{abstract}

42: 

43: \section{Introduction}

44: 

45: %

46: %

47: %

48: %

49: %

50: Mobile Ad hoc NETworks (MANETs) are networks made of mobile nodes

51: that self-configure and collaborate to forward packets among each other

52: without the benefit of an access point.  These networks are especially

53: important when infrastructure is unavailable (e.g., unplanned

54: networks, in remote areas, or after a disaster), or expensive.  In

55: such networks, each node must play the role of a router as well as a

56: station.

57: 

58: %

59: %

60: %

61: %

62: %

63: %

64: %

65: 

66:  Most existing MANET research assumes idealized wireless propagation:

67: nodes have a fixed transmission range and all receivers within this

68: range receive a transmission correctly (assuming no collision occurs).

69: However, due to fading and multipath effects this assumption

70: deviates significantly from reality: wireless transmission can suffer

71: deep fading (drop in power level) with very small changes in location,

72: or in time.  Fades of 20dB (the signal dropping to 1\% of its ideal

73: value) are not rare~\cite{Rappaport}.  These properties have profound

74: implications on protocol performance and designs that make many of the

75: decisions taken under idealized assumptions invalid.  Wireless

76: transmission success ratio drops with the distance between sender and

77: receiver.  This behavior leads to several interactions within the MAC

78: layer and across layers, some of which are subtle.

79: 

80: The effect of fading on MANET behavior has been observed and

81: studied empirically;  Decouto et al show that,

82: contrary to idealized models, fading has considerable effect on

83: link state -- using shortest path as a measure of

84: path quality can therefore be misleading~\cite{decouto-03a}.  Moreover, the packet loss

85: rate varies from link to link (as a function of the distance and the

86: surrounding environment); the quality of a hop must be made visible to

87: the routing protocol to enable effective route selection.  To capture

88: this effect, they propose ETX: a link cost metric that is a function

89: of the forward and backward delivery ratio on a

90: link~\cite{decouto-03}.  Other link cost metrics have also been

91: proposed to allow fading sensitive route selection~\cite{adya-04}.

92: Draves et al conducted an experimental analysis of these metrics and

93: found that ETX performs best~\cite{draves-04}.  Integrating these

94: effects into a routing architecture has been studied as well.  Woo et

95: al explored a routing architecture for managing wireless propagation

96: vagaries in a sensor network environment~\cite{woo-03}.  One of their

97: main conclusions is that using ETX as a route metric provides stable

98: routing performance.  However, the static sensor network environment

99: allows specialized routing and link estimation solutions that are

100: difficult to generalize to a MANET environment.  Draves et al propose

101: Link Quality Source Routing (LQSR), an extension of the Dynamic Source

102: Routing Protocol to address the problems that arise due to

103: fading~\cite{draves-04}.  In addition to allowing path selection

104: based on link quality, LQSR has several additional interesting

105: features such as continuous monitoring of path quality and higher

106: route stability.

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: %

134: 

135: Other effects of fading have also been observed and are reviewed

136: throughout this paper.  The majority of the work is experimental.

137: More specifically, problems are observed in real testbeds and

138: solutions are proposed to address them.  As a result most of the work

139: is focused on routing implications.  In this paper, we take a

140: complementary bottom up approach to this problem.  We start with an

141: analysis of the problem at the MAC layer.  As a result, we capture

142: some problems that occur in the MAC layer itself as well as problems

143: that affect the upper layers.  The bottom up approach provides a more

144: systemic and comprehensive evaluation of the effect of

145: fading on protocol performance than experimental testbed analysis.

146: The contributions of the paper also include the identification of

147: new problems that arise due to fading and the development

148: of initial solutions for them.

149: 

150: We first characterize the effect of fading on the performance MANET

151: protocols analytically and using simulation using a slow fading

152: propagation model.  One of our contributions is to show that Collision Avoidance (Request to Send (RTS)/ and Clear to Send (CTS) packets of IEEE 802.11)

153: is harmful to performance under fading and not that beneficial for

154: collision avoidance.  Moreover, we identify an unfairness problem that

155: occurs due to unbalanced backoffs that occur due to fading losses.

156: In addition, we outline the implications of this behavior on the

157: routing layer and propose preliminary solutions to the problems. 

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: %

184: %

185: %

186: %

187: %

188: %

189: 

190: The remainder of this paper is organized as follows.

191: Section~\ref{background} explains some background

192: information. Section~\ref{sec:mac} analyzes the basic effect of

193: fading on MAC layer performance. Section~\ref{sec:noRTS} makes the case for

194: eliminating Collision Avoidance from MAC protocols.  Section~\ref{sec:backoff}

195: outlines the unfairness and inefficiency problems occurring due to backoff under fading.  In Section~\ref{sec:routing} we present high level descriptions of additional problems that occur

196: at upper layers.  Finally Section~\ref{sec:conclude} presents some concluding

197: remarks.

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: %

234: %

235: %

236: %

237:  

238: %

239: %

240: %

241: %

242: %

243: %

244: 

245: %

246: %

247: %

248: %

249: %

250: %

251: 

252: 

253: %

254: %

255: %

256: %

257: %

258: %

259: %

260: %

261: %

262: %

263: %

264:  

265: 

266: \section{Background}~\label{background}

267: 

268: 

269: %

270: %

271: %

272: %

273: %

274: 

275: %

276: %

277: %

278: %

279: %

280: %

281: %

282: %

283: %

284: %

285: %

286: %

287: %

288: %

289: 

290: %

291: %

292: %

293: %

294: %

295: %

296: 

297: \subsection{Wireless Propagation} 

298: 

299: %

300: %

301: %

302: %

303: %

304: 

305: %

306: %

307: %

308: %

309: %

310: %

311: %

312: %

313: %

314: 

315: In a typical terrestrial environment, the transmitted signal reflects

316: off of surrounding objects, refracts when travelling through obstacles

317: and suffers diffraction due to sharp edges.  Moreover, doppler shifts

318: occur due to moving objects.  As a result, many echos of the

319: transmitted signal are received with different delays and power-levels

320: depending on the path they took.  Together, these result in large

321: transient flucuations in the power level: this is known as

322: fading~\cite{Rappaport}.  

323: 

324: 

325: The details of wireless propagation are beyond the scope of this

326: paper.  Briefly, fading of a wave can be explained by slow fading

327: (e.g., due to dominant shadowing objects) and fast fading (due

328: to numerous smaller objects)~\cite{Sklar-97}.  Slow fading occurs over

329: time periods generally longer than a packet length; thus, we may

330: consider a single transmission power value per packet.  Alternatively,

331: small scale fading, or fast fading, occurs within the packet.  Fast

332: fading model considers the effect of scattered wave (also called by

333: multi-path reception model).  The Rayleigh or Ricean statistical

334: distributions have been shown to capture this effect well.  Ricean is

335: used when there is a line of sight path between sender and receiver

336: and Rayleigh is used when there is none.

337: 

338: With advanced modulation technologies (specifically, spread

339: spectrum/CDMA) and specialized antennas (such as the RAKE receiver),

340: the effect of small scale fading can be almost eliminated in the RF

341: frequency bands used for wireless communication.  Thus, we focus only

342: on slow fading.

343: 

344: 

345: %

346: %

347: %

348: %

349: %

350:  

351: %

352: %

353: %

354: %

355: 

356: %

357: %

358: %

359: %

360: %

361: %

362: 

363: 

364: \subsection{Propagation Models}

365: 

366: There are three kinds of propagation models typically used in MANET

367: protocol simulation and analysis.  The simplest model is Free Space

368: model.  The energy is in inverse proportion of square of distance.

369: This model is too simple to apply to realistic terrestrial setting.

370: Therefore, a more realistic model (called Two Ray Ground), which

371: considers the reflection of signal against the ground as well as

372: directly propagated signal, is used.  At short distance, only the

373: directly propagated signal matters.  Thus, the path loss exponent

374: ($\beta$), which determines the (exponentiated) rate of attenuation of

375: the signal with distance, is 2; i.e., the power drops with the square

376: of distance.  However, after a cross-over distance which depends on

377: the height of the antennas, both the direct component and the ground

378: reflected component combine to create a higher path loss factor

379: (typically assumed to be between 3 and 4).  The Two-Ray Ground Model

380: is also idealized; it does not consider the fading effects

381: described above.

382: 

383: 

384: %

385: %

386: %

387: %

388: %

389: %

390: %

391: %

392: %

393: %

394: %

395: %

396: 

397: %

398: %

399: %

400: %

401: %

402: %

403: %

404: %

405: %

406: %

407: %

408: %

409: %

410: 

411: %

412: %

413: %

414: %

415: %

416: %

417: %

418: %

419: %

420: %

421: %

422: 

423: We use a fading propagation model in this paper that statistically

424: models slow fading.  Although this model is known (e.g., it is

425: available in the NS-2~\cite{ns-2} and Qualnet

426: simulators~\cite{qualnet}), it is not typically used.  In this model,

427: the path loss is a random variable that has a log-normal distribution,

428: with a mean equal to the expected two-ray ground path loss.  More

429: specifically, the signal power consists of two parts: the mean power

430: and the fading effect. The mean received power at a communication

431: distance is the idealized power as a function of that distance is

432: calculated as follows,

433: 

434: 

435: \begin{equation}

436: [\frac{\overline{P_r(d)}}{P_r(d_0)}]_{dB} =  -10\beta log(\frac{d}{d_0})          

437: \label{eq4}

438: \end{equation}

439: where $d0$ is a reference distance that is a function of the antenna

440: height.  The path loss equation in formula dB is eq~\ref{eq4}.

441: $\beta$ is the ideal path loss exponent (i.e., without considering

442: fading).

443: 

444: 

445: The second component models the transient fading effect.  Received

446: power is adjusted by a log-normal random variable $X_{dB}=

447: (N(0,\sigma^2))$.  The fading is modeled as a gaussian distribution

448: with average 0, and standard deviation $sigma$.  The overall received

449: power ius expressed as,

450: \begin{equation}

451: [\frac{\overline{P_r(d)}}{P_r(d_0)}]_{dB} =  -10\beta log(\frac{d}{d_0})  + X_{dB}.

452: \label{eq5}

453: \end{equation}

454: Because random variable $X_{dB}$, the range no longer represents a

455: discrete threshold with ``in-range'' or ``out-of-range'' nodes.

456: Rather, it is now continuous: packets in the ideal range may be lost

457: and packets outside the ideal range may be received.  The probability

458: of correct reception decreases with distance accoding to

459: Eq.~\ref{eq5}.  While this model is available in simulators such as

460: NS-2 and QualNet, it is almost never used

461: in MANET network level research.  We use it as the basis for our

462: analysis.

463: 

464: 

465: Rather than focusing on the impact of $\sigma$ and $\beta$ on protocol

466: behavior, we pick representative values and use them to evaluate the

467: impact of fading on MANET protocols.  Clearly, the impact can be

468: amplified or lessened with different $\sigma$ and $\beta$ values.

469: 

470: \subsection{Fading Model Limitations}

471: 

472: The model we use is statistical and has several limitations.  Recent

473: observations have shown that transmission losses at the link layer are

474: not generally independent~\cite{aguayo-04}.  Moreover, a limitation of

475: the model we consider is that the PDF of the received power is

476: strictly a function of distance, making links symmetric.  However,

477: fading is known to cause asymmetric link qualities.  Since IEEE

478: 802.11 requires symmetric communication, it is likely to suffer as a

479: result of asymetric link qualities.  The effect of Asymmetry on

480: routing behavior has been considered but we do not know of any

481: analysis of the microeffects that arise due to

482: asymmetry~\cite{marina-02}; this is a topic of future work.  There is

483: a need for developing more accurate propagation models that capture

484: these effects.  However, we believe the simple fading models we use

485: are sufficient to identify the effects and problems we discuss in this

486: paper: more refined models will certainly enable more accurate

487: characterization of their effect.

488: 

489: Another limitation of the model is that we do not consider the effect

490: of multi-rate MAC protocols.  Recent versions of IEEE 802.11 recognize

491: the effect of fading on link quality~\cite{ieee-802.11std}.  To combat

492: this effect, the use multi-rate transmission where the modulation

493: scheme used is adapted dynamically and at the packet level to match

494: the link quality (high rate when link quality is high and lower rate

495: when it is low)~\cite{sadeghi-02}.  While this mitigates the effect of

496: fading, it comes at a high cost because the interframe spacing,

497: backoff periods and the rate negotiation headers have to be

498: transmitted at the lowest rate~\cite{pagtzis-01}.  We focus on single

499: transmission rate MAC protocols. While multi-rate MACs help to

500: mitigate the effect of fading, low-cost, low-power radios continue

501: to be single rate~\cite{bluetooth,ieee802.15std,woo-01}.  These radios are

502: the most likely candidate for use in embedded and low power devices.

503: 

504: 

505: 

506: %

507: 

508: %

509: %

510: %

511: %

512: %

513: 

514: 

515: %

516: %

517: %

518: %

519: %

520: %

521: %

522: %

523: %

524: %

525: %

526: %

527: %

528: %

529: %

530: %

531: %

532: %

533: %

534: %

535: %

536: %

537: %

538: %

539: 

540: 

541: %

542: %

543: %

544: %

545: %

546: %

547: %

548: %

549: %

550: %

551: %

552: %

553:  

554: 

555: %

556: %

557: %

558: %

559: %

560: %

561: %

562: %

563: %

564: %

565: %

566: %

567: %

568: %

569: %

570: 

571: %

572: %

573: %

574: %

575: %

576: 

577: 

578: %

579: %

580: %

581: %

582: %

583: %

584: %

585: %

586: %

587: %

588: %

589: %

590: %

591: %

592: %

593: %

594: %

595: %

596: %

597: %

598: %

599: %

600: 

601: %

602: %

603: %

604: %

605: %

606: %

607: 

608: %

609: %

610: %

611: %

612: %

613: %

614: %

615: %

616: 

617: %

618: %

619: %

620: %

621: %

622: %

623: %

624: %

625: %

626: %

627: %

628: %

629: %

630: %

631: 

632: %

633: %

634: %

635: 

636: %

637: %

638: %

639: %

640: %

641: %

642: %

643: %

644: %

645: 

646: %

647: %

648: %

649: %

650: %

651: %

652: %

653: %

654: %

655: %

656: %

657: %

658: %

659: %

660: %

661: 

662: %

663: %

664: %

665: %

666: %

667: %

668: %

669: %

670: %

671: %

672: %

673: %

674: %

675: %

676: %

677: %

678: %

679: %

680: %

681: %

682: %

683: %

684: %

685: 

686: %

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: \section{Basic Effect of Fading on the MAC Protocol}~\label{sec:mac}

713: 

714: In this section, we analyze the effect of fading on the MAC layer.

715: The effects identified in this section are not intended to be

716: comprehensive.  The fading model used is idealized and the

717: scenarios are simple.  Other intricate interactions may arise in more

718: realistic environments.  Our aim is to emphasize that fading must

719: be treated as a first class problem in MANET MAC protocol design.

720: Furthermore, in the next Section, we argue that it must also be

721: carefully designed for in higher level protocols.

722: 

723: The NS-2 network simulator is used for all simulation experiments in

724: this paper~\cite{ns-2}.  We use a path loss component of 3 (typical

725: range is 2--6 in real environments).  For the log-normal fading

726: model we set $\sigma=4dB$ (recall that 20dB fades are not uncommon; a

727: standard deviation of 4dB is rather conservative).  The transmission

728: power is set such that the ideal range is 250 meters.

729: 

730: 

731: \begin{figure}

732: \centerline{\includegraphics[width=0.85\linewidth]{power.eps}}

733: \caption{Unstable receiving power} 

734: \label{fig-recvpower220}

735: \end{figure}

736: 

737:  Figure~\ref{fig-recvpower220} shows a simulation trace of the

738: received power at distance 220 meters.  Note that when the receiving

739: power drops below the receiver sensitivity threshold, the transmission

740: is lost. Because we use a slow fading model, the same receive power is

741: assumed per packet.  

742: 

743: \subsection{Preliminaries}

744: 

745: In fading model, the received power depends on path loss $\beta$

746: and the log-normal component $X_{dB}$.  Eq (~\ref{eq5}) can be

747: simplified as the following: $\psi (p) = \tau(d) + X_{dB}$, where

748: $\psi(p)=[\frac{\overline{P_r(d)}}{P_r(d_0)}]_{dB} $ and

749: $\tau(d)=-10\beta log(\frac{d}{d_0}) $. Recall that $\psi(p)$ has

750: normal distribution $N( \tau(d) ,\sigma^2 )$.  Ignoring interference,

751: the probability of correct transmission is the probability of the

752: received power being higher than the receiver sensitivity threshold

753: ($P_{th}$, which is a radio constant).  This can be directly computed

754: as P($\psi(p) \ge P_{th}$).  

755: 

756: \begin{figure}

757: \centerline{\includegraphics[width=0.85\linewidth]{delivery-ratio.eps}}

758: \caption{Delivery Ratio With Distance} 

759: \label{fig-deliveryratio}

760: \end{figure}

761: 

762: Figure ~\ref{fig-deliveryratio} shows the computed delivery ratio

763: vs. that of the idealized two ray ground model.  Clearly fading has

764: a large effect on the packet delivery ratio as the distance between

765: the sender and the receiver increases.  Although we show our results

766: as function of distance, they should be considered of in terms of

767: delivery ratio (which in our model has a one to one mapping per

768: Figure~\ref{fig-deliveryratio}).  In real testbeds, the packet

769: reception probability is influenced by the surrounding environment and

770: its PDF is not only a function of distance.

771: 

772: This observation confirms the need for routing protocols to be aware

773: of the link quality. The figure also shows the delivery ratio obtained

774: from simulation which not surprisingly, is very close to the

775: calculated ones.  Since the link level transmission ratio in the

776: simulation is being generated with the same distribution used in the

777: analysis, the simulation represents a Monte Carlo solution with

778: packets dropped according to the propagation model equation.

779: 

780: \subsection{Effect of Packet Retransmission}

781: 

782: Packet retransmission is used to increase reliability and to recover

783: both from transmission errors and collisions.  In this section we

784: analyze the effect of retransmission on packet delivery ratio and show

785: that under low delivery retransmission is in fact counterproductive.

786: Moreover, we analyze the effect of retransmission on packet delay.  We

787: do this using the retransmission algorithm of IEEE 802.11.

788: 

789: In IEEE 802.11, there is a Short Retry Limit (SRL) and Long Retry

790: Limit (LRL).  A transmission is classified as long or short based on

791: its length relative to a fixed threshold.  A transmission below the

792: threshold is counted against the SRL, while a transmission above it is

793: counted against both SRL and LRL.  SRL and LRL are 7 and 4

794: respectively.  

795: 

796: \subsubsection{Effect on Packet Delivery Ratio}

797: 

798: Successful delivery of a data packet requires successful delivery of

799: sequence of RTS, CTS, DATA and ACK packets.  If the data packet is

800: shorter than the threshold, 7 retransmissions are tried regardless of

801: where the failure is.  If it is longer than the threshold, failures

802: that occur in the DATA or ACK are counted against both limits (since a

803: long transmission of the data has occurred), but failures in RTS or

804: CTS count against SRL only.

805: 

806: The effect of retransmission on overall packet delivery in the

807: presence of fading (but ignoring collisions) can be derived as

808: follows.  In order for a transmission to succeed, RTS, CTS, DATA and

809: ACK should not fail.  Thus packet success ($p_s$)and failure ($p_f$)

810: probabilities on a given try are given by $p_s=p^4$ and $p_f=(1-p^4)$,

811: where p is the link level delivery ratio (Eq~\ref{eq5}).  If there is

812: a failure in the first try at either of these four transmissions, a

813: second retry occurs, and so on.  The success in the first four tries

814: can be computed (binomial experiment) as:

815: \begin{eqnarray}

816: P_{1to4}= p_s+p_f*p_s+p_f^2*p_s+p_f^3*p_s \nonumber \\ 

817: p_s(1+p_f+p_f^2+p_f^3)

818: \end{eqnarray}

819: A fifth transmission will only occur if at least one of the failures

820: did not count against LRL.  The probability of a short failure

821: (failure in RTS or CTS) can be computed as

822: \begin{equation}

823: p_{sf} = 1-p^2

824: \end{equation}

825: Similarly, the probability of a long retry failure is

826: \begin{equation}

827: p_{lf} = p^2(1-p^2)

828: \end{equation}

829: The probability of a success in the fifth transmission can be computed

830: as the probability of failure in all 4 first transmissions, with one

831: or more of the failures being short, and a success in the fifth try.

832: Specifically, 

833: \begin{equation}

834: P_5=(p_f^4-p_{lf}^4)*p_s

835: \end{equation}

836: Similar reasoning can be applied to compute $P_6$ and $P_7$ to give a

837: total packet delivery probability as follows.

838: \begin{eqnarray}

839: P_{packet}=p_s(1+p_f+p_f^2+p_f^3+ p_f^4-p_{lf}^4 \nonumber \\

840: +((4*p_{sf}^2*p_{lf}^3) + (p_f^4-p_{lf}^4  \nonumber\\

841: -4 p_{sf}*p_{lf}^3)*p_f)) + (16*(p_{sf}^3*p_{lf}^3))\nonumber \\

842: +(4*p_{sf}^3*p_{lf}+p_{sf}^4)*p_f^2))~\label{eqcbrpckcase}

843: \end{eqnarray}

844: \noindent

845:  For short length data packet, the packet delivery ratio can be

846: derived as the probability of success in the first 7 trials of a

847: binomial experiment with $p_s$ outcome probability.

848: 

849: 

850: \begin{figure}

851: \centerline{\includegraphics[width=0.85\linewidth]{packet-delivery-ratio.eps}}

852: \caption{Packet Delivery Ratio} 

853: \label{fig-cbrdeliveryratiobycalc}

854: \end{figure}

855: Figure ~\ref{fig-cbrdeliveryratiobycalc} shows the probability of data

856: packet delivery.  Perhaps surprisingly, the long length data delivery

857: ratio is almost identical to the short length one; this indicates that

858: the high number of transmissions may not be needed (we note that this

859: conclusion may change when collisions are considered).  When the

860: delivery ratio is above 0.6, CBR packet delivery ratio is better than

861: link level delivery ratio.  However, {\bf below 0.6, the packet delivery

862: ratio is worse than the link delivery ratio despite retransmission.}

863: This effect occurs because correct retransmission requires correct

864: reception of 4 packets (RTS/CTS/DATA/ACK), while link level success

865: requires correct reception of only 1.  The retransmissions are not

866: sufficient to overcome this disadvantage at low link level delivery

867: ratios.

868: 

869: 

870: 

871: \subsection{Effect on Delay}

872: 

873: %

874: %

875: %

876: %

877: %

878: 

879: Fading also affects the delay of packet delivery; multiple retries

880: are needed, with an exponentially increasing backoff between them.  We

881: now develop a simple analytical model for the delay.  Developing a

882: closed form solution for packet delay in the presence of the long and

883: short retry limits requires enumerating possible sequences of failures

884: (a few hundred cases).  Instead, we develop a solution for the

885: expected value of a backoff period as a function of the link level

886: delivery ratio. 

887: 

888: 

889: We note that the backoff is reset to the minimum value (31 slots)

890: whenever a CTS or ACK packet is received, and doubled whenever they

891: fail to be received (with a cap of 1023 slots).  The backoff period is

892: selected randomly between 0 and the current window size.  A CTS (ACK)

893: is received only if the RTS (Data) packet is received.  The

894: probability of two consecutive correct transmissions (ignoring

895: interference) is $p^2$.  The expected value of a backoff period can be

896: computed as follows

897: \begin{eqnarray}

898: E(Backoff) = \frac{p^2}{2} \cdot ( 31 + 63 *(1-p^2)  \nonumber  \\

899: + 127*(1-p^2)^2 + 255 \cdot (1-p^2)^3  \nonumber  \\

900: + 511 \cdot (1-p^2)^4) + \frac{1023}{2} \cdot (1-p^2(1  \nonumber  \\

901: + (1-p^2) + (1-p^2)^2 +(1-p^2)^3 +(1-p^2)^4))

902: \end{eqnarray}

903: \noindent

904: 

905: 

906: \begin{figure}

907: \centerline{\includegraphics[width=0.85\linewidth]{expected-backoff.eps}}

908: \caption{Expected Value of the Backoff} 

909: \label{Ebackoff}

910: \end{figure}

911: 

912: 

913: %

914: %

915: %

916: %

917: %

918: 

919: \begin{figure}[ht]

920: \centerline{\includegraphics[width=0.85\linewidth]{delay1hop.eps}}

921: \caption{One-hop Packet Delay}

922: \label{delay1hop}

923: \end{figure}

924: 

925: This expected value is shown in Figure~\ref{Ebackoff}.  The average

926: backoff period is significantly higher than minimum even at reasonably

927: stable links.  The backoff period increases exponentially but is

928: capped at the maximum period resulting in the behavior shown.  Note

929: that the average packet delay will increase faster than this ratio

930: since the number of required retransmissions will increase as the

931: delivery ratio drops (incurring multiple backoff periods in addition

932: to the retransmissions).  The one-hop packet delay obtained by

933: simulation is shown in Figure~\ref{delay1hop}.  The delay increases

934: quickly with the as the quality of the link drops.  

935: 

936: In the following sections, we discuss some important implications that

937: follow from this analysis.  First, we make the case that RTS/CTS is

938: harmful to performance under fading.  Second, we identify an

939: unfairness problem that arises due to fading.  Finally, we discuss

940: additional effects that arise due to fading.

941: 

942: %

943: %

944: %

945: %

946: 

947: \section{RTS/CTS Considered Harmful}~\label{sec:noRTS}

948: 

949: As was observed in Figure~\ref{fig-cbrdeliveryratiobycalc}, the packet

950: delivery ratio drops below the transmission delivery ratio when the

951: link state is poor (below 60\% transmission delivery ratio).  The

952: result is explained by the fact that 4 transmissions must succeed for

953: a packet to be delivered (RTS,CTS, Data and ACK).  At low transmission

954: success probability it becomes highly improbable for 4 consecutive

955: transmissions to succeed.

956: 

957: An alternative approach (which is supported by IEEE 802.11 for short

958: packets) is to rely on just CSMA without RTS/CTS and use

959: acknowledgments to recover from errors and collisions.  We first

960: analyze this approach assuming no collisions and then revisit to

961: discuss the effect of collisions.  In this approach, only two

962: consecutive transmissions (Data and ACK) have to be received

963: correctly; this has a much higher probability of success than 4

964: consecutive transmissions.  In this case, only two transmissions need

965: be received correctly (DATA and ACK).  In this section, we make the

966: argument that removing RTS/CTS is beneficial for performance.  We make

967: this argument in two parts: first we show that removing RTS/CTS

968: significantly improves packet delivery in the presence of losses, and

969: then show that the effect of eliminating RTS/CTS on reducing

970: collisions is not large.

971: 

972: \subsection{MAC-layer Analysis without RTS/CTS}

973: 

974:   The probability of correct reception with a retransmit limit of 4

975: when RTS/CTS is not used can be obtained as a 4 step binomial

976: experiment as follows:

977: \begin{equation}

978: P_{packet}=p^2(1+(1-p^2)+(1-p^2)^2+(1-p^2)^3)

979: \end{equation}

980: 

981: 

982: \begin{figure}

983: \centerline{\includegraphics[width=0.85\linewidth]{norts-delivery.eps}}

984: \caption{Packet Delivery with/without RTS/CTS} 

985: \label{no-rts-delivery}

986: \end{figure}

987: 

988: \begin{figure}

989: \centerline{\includegraphics[width=0.85\linewidth]{capacity.eps}}

990: \caption{Link Capacity} 

991: \label{capacity}

992: \end{figure}

993: 

994: 

995: Packet delivery ratio is significantly improved when the RTS/CTS

996: handshake is omitted.  This effect can be seen in

997: Figure~\ref{no-rts-delivery}.  The average packet delay is reduced if

998: RTS/CTS are omitted.  This can be proven using the observation that

999: the expected delay of a transmission assuming all RTS/CTS are

1000: successful and ignoring the cost of RTS/CTS is equivalent to the

1001: expected delay of the approach that does not use RTS/CTS.  Thus, using

1002: RTS/CTS, we have to add the cost of the RTS/CTS as well as failed

1003: RTS/CTS and the associated increase in backoffs.

1004: Figure~\ref{capacity} shows the obtained capacity with and without

1005: RTS/CTS for a single hop -- clearly, there is a large advantage for

1006: the case without RTS/CTS.  Similar positive results hold for the delay

1007: (not shown).

1008: 

1009: 

1010: %

1011: %

1012: %

1013: %

1014: %

1015: %

1016: %

1017: %

1018: %

1019: %

1020: 

1021: \subsection{Effect on Collisions}

1022: 

1023: A potential drawback of this approach is that the advantages of

1024: collision avoidance (RTS and CTS) in reducing collisions and collision

1025: cost are lost. We argue that RTS/CTS is only of limited success in eliminating

1026: collisions and that the complimentary Carrier Sense Multiple Access

1027: (CSMA) is more effective in reducing collisions.  Some collisions

1028: occur when two senders sense the medium to be idle and transmit

1029: concurrently; such collisions cannot be prevented by either approach.

1030: However, the vulnerability period to such collisions is low and their

1031: effect is not likely to dominate.  One advantage that RTS/CTS offer to

1032: such collisions is that contention is carried out with small control

1033: packets, whereas a full data packet is lost if only CSMA is used.

1034: 

1035: 

1036: \begin{figure}

1037: \centerline{\epsfig{file=csma.eps,width=0.7\linewidth,angle=270,silent=}}

1038: \caption{CSMA vs. Collision Avoidance} 

1039: \label{csma}

1040: \end{figure}

1041: 

1042: Collisions occur in a wireless environment if the ratio of the

1043: received packet power to the interfering power (other transmissions

1044: and noise) fall below a threshold called the {\em capture threshold}.

1045: Its well known that collision avoidance is not sufficient to prevent

1046: collisions: a node that is close enough to the receiver to interfere

1047: but not close enough to receive the CTS packet can possibly cause a

1048: collision.  In Figure~\ref{csma}, we consider a situation where

1049: a source (S) is communicating with a receiver (R) in the presence of

1050: an interfering node (I).  The x-axis shows the distance between the S

1051: and R, while the y-axis shows that between I and R.  Points on the

1052: figure above the capture lines indicate that capture does happen (the

1053: interferer is too far) while those below indicate that a collision

1054: happens.  The capture value represents the capture threshold, while

1055: $n$ represents the path loss factor (the signal decays with

1056: $\frac{1}{r}^n$ where $r$ is the distance from the transmitter).  The

1057: capture line plots the interferer distance that would

1058: cause $\frac{P_{sender}}{P_{interferer}}=Capture$.  Note

1059: that this approximate analysis assumes only a single interferer,

1060: ignores the effect of noise, and assumes ideal propagation.

1061: 

1062: Essentially to control collisions all potential interferers (below the

1063: capture line on Figure~\ref{csma}) should be blocked.  Collision

1064: Avoidance (RTS/CTS) can only block interferers in reception range of

1065: the receiver (to receive the CTS).  Thus, many potential interferers

1066: are not blocked with the RTS/CTS mechanism.  This is the area under

1067: the capture line but above the 250 meter line.

1068: 

1069: The main mechanism for avoiding collisions in commercial wireless

1070: cards is an aggressive Carrier Sense (with low sense threshold).  This

1071: aggressive threshold is used both to attempt for interferers out of

1072: receivers range and for the fact that carrier sense occurs at the

1073: sender, but collisions occur at the receiver.  For a WaveLAN

1074: card~\cite{wavelan-manual} with a nominal transmission range of 250

1075: meters, the carrier sense threshold is set such that transmitters in a

1076: circle of radius 2.2 times the transmission range around S are forced

1077: to stay idle (assuming a path loss factor of 4).  Having this

1078: aggressive threshold reduces collisions (it does not completely

1079: eliminate them), but increases the exposed terminal problem.

1080: 

1081: CSMA's effectiveness in blocking interferers is a superset of that of

1082: Collision Avoidance: all collisions that can be eliminated by

1083: collision avoidance can also be eliminated by CSMA.  Furthermore, CSMA

1084: is able to prevent many more collisions not preventable by RTS/CTS.

1085: Since CSMA is applied at the sender but the interference happens at

1086: the receiver, its effectiveness depends on the location of the

1087: interferer from the source.  More specifically, for carrier sense to

1088: occur successfully, the signal power of the interferer at the sender

1089: should be above the carrier sense threshold.  In the average case

1090: (when the interferer to source distance is equal to the receiver to

1091: source distance; that is, the interferer location is not biased either

1092: closer to the sender or further away from it), CSMA is able to prevent

1093: all possible collisions (the average-case line on Figure~\ref{csma}).

1094: The worst case for CSMA occurs is when the interferer is on the side

1095: of the receiver away from the sender (this is the worst-case line on

1096: Figure~\ref{csma}).  In this case, the interferer's distance to the

1097: sender is maximized for a given interferer-receiver distance.  Even in

1098: the worst case, CSMA prevents all collisions catchable by CA -- recall

1099: that all the cases under the line are prevented for each mechanism.

1100: Note that in the worst case, some collisions are not preventable.

1101: 

1102: Thus, we perceive no benefit for RTS/CTS in reducing collisions.

1103: Their only benefit is in reducing the cost of collisions for

1104: collisions that occur due to concurrent sensing of an idle channel; in

1105: this case, collisions occur on the small RTS/CTS control packets

1106: instead of full length data packets.  This only works for interferers

1107: in range to receive a CTS -- this is a relatively small area of

1108: possible interferers as seen on the Figure.  Moreover, in the case of multiple

1109: interferers, their combined interference power may cause a collision.  This

1110: helps CSMA and hurts CA: CA requires that each of the intereferers be in

1111: reception range, while CSMA naturally takes their effect in since carrier

1112: sense measures their combined effect.  Thus, given the

1113: dramatic degradation CA causes on packet delivery ratio in the presence of

1114: fading and the small benefit it provides to collisions 

1115: it is beneficial to eliminate them.  

1116: 

1117: We note that this conclusion is only dependent on the carrier

1118: sense threshold.  If the path loss component is different, or if the

1119: send power is different, this affects both the carrier sense and the

1120: collision avoidance.  Furthermore, if the capture threshold is

1121: different, this has the same effect on both CSMA and CA (the slope of

1122: the capture line changes, but not the other lines).

1123: 

1124: %

1125:   %

1126: 

1127: %

1128: %

1129: %

1130: %

1131: %

1132: 

1133: 

1134: %

1135: %

1136: %

1137: %

1138: %

1139: %

1140: %

1141: %

1142: %

1143: %

1144: %

1145: %

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: \section{Backoff Implications -- Inefficiency and Unfairness}~\label{sec:backoff}

1182: 

1183: The second major implication from the basic MAC analysis in

1184: Section~\ref{sec:mac} concerns the backoff mechanism.  The backoff

1185: mechanism in contention based MAC protocols (such as 802.11 DCF mode

1186: used for MANETs) is intended to regulate the offered load to the

1187: shared medium.  The underlying assumption is that all packet losses

1188: are due to collisions.  While this assumption is true in wired shared

1189: media where errors are exceptionally rare, it is not true in wireless

1190: environments.  As a result of losses due to fading, the backoff

1191: timer is increased.  This leads to two important side effects: (1)

1192: Inefficiency in using the medium: backoffs occur even without

1193: collisions, leading to nodes backing off excessively

1194: causing unnecessary channel idle time;

1195: and (2) unfairness: because the expected backoff period increases with

1196: the of transmission losses, links that experience losses have a larger

1197: average backoff than those that do.  IEEE 802.11 is known to be

1198: susceptible to short-term transient unfairness even under idealized

1199: propagation assumptions~\cite{vaidya-00}.  Under fading

1200: we show that steady-state unfairness can occur.

1201: 

1202: The inefficiency issue was already alluded to in the discussion of the

1203: delay and effective throughput.  Backoff increasingly contributes to

1204: the delay as the transmission delivery ratio drops: as a larger number

1205: of retries is needed, we have the multiplicative effect of a larger

1206: number of backoff periods and longer average backoffs.  The problem

1207: occurs due to the implicit assumption that losses are due to

1208: contention.  In the presence of fading, this is often not the case

1209: causing inappropriate backoff.  Finer discrimination between

1210: contention losses (collisions) and transmission losses (fading) are

1211: needed.

1212: 

1213: The unfairness problem occurs due to the imbalance in backoff

1214: durations.  On average, links with lower transmission success ratios

1215: will have a higher backoff period (Figure~\ref{Ebackoff}).  As a

1216: result, competition for the link is no longer fair -- links with

1217: higher transmission delivery ratio have a higher probability of

1218: capturing the link.  IEEE 802.11 already suffers from short term

1219: unfairness~\cite{vaidya-00}.  However, the unfairness problem

1220: identified here is sustained.  Addressing this problem also requires

1221: backoff algorithms that can discriminate between collisions and

1222: transmission losses.

1223: 

1224: %

1225: \begin{figure}

1226: \centerline{\includegraphics[width=0.85\linewidth]{throughput.eps}}

1227: \caption{Unfair Throughput} 

1228: \label{unfair3}

1229: \end{figure}

1230: 

1231: 

1232: To illustrate this problem, we simulate two single hop connections

1233: whose sources are in range with each other, but whose receivers are

1234: not (to isolate the effect of a single contention point).  The

1235: distance (and therefore the transmission delivery ratio) of one

1236: connection was fixed at 150 meters (Connection 1) and the other varied

1237: (Connection 2).  Each connection generates CBR traffic at a rate that

1238: would use all the available bandwidth if there is no contention.

1239: Figure~\ref{unfair3} shows the raw throughput obtained by the two

1240: connections with and without RTS/CTS.  The first observation is that

1241: the no RTS/CTS version is able to

1242: obtain higher throughput even in the presence of contention.  The

1243: unfairness problem can be seen on the diagram as the first connection

1244: gets an increasingly higher portion of the available bandwidth at the

1245: expense of the weaker connection.

1246: 

1247: %

1248: \begin{figure}

1249: \centerline{\includegraphics[width=0.85\linewidth]{unfairness.eps}}

1250: \caption{Unfairness with Packet size 500 bytes} 

1251: \label{unfair1}

1252: \end{figure}

1253: 

1254: \begin{figure}

1255: \centerline{\includegraphics[width=0.85\linewidth]{unfairness1500.eps}}

1256: \caption{Unfairness with Packet size 1500 bytes} 

1257: \label{unfair2}

1258: \end{figure}

1259: 

1260: 

1261: 

1262: Its difficult to assess the degree of unfairness since the expected

1263: throughput of the hop goes down with the increased hop delay as the

1264: connection becomes weaker.  To normalize this effect, we use a metric

1265: of the percentage of delivered packets as a ratio of the maximum

1266: deliverable packets on the connection if no interference was present.

1267: In a fair implementation, the two connections would be able to get an

1268: equivalent percentage of their maximum throughput.

1269: Figures~\ref{unfair1} and ~\ref{unfair2} illustrate the problem for 500

1270: byte and 1500 byte packets respectively.  Clearly, the strong

1271: connection dominates for both scenarios.  When the two distances are

1272: equal, the connections share the bandwidth almost equally since the

1273: backoffs are not biased in favor of either connection.  The unfairness

1274: problem is clear as the connection using the strong link quickly

1275: dominates the weaker one.  This is true for both packet sizes studied

1276: and with and without RTS/CTS.

1277: 

1278: \begin{figure}

1279: \centerline{\includegraphics[width=0.85\linewidth]{unfairnessCW.eps}}

1280: \caption{Backoff Disabled} 

1281: \label{fair}

1282: \end{figure}

1283: %

1284: %

1285: 

1286: To verify that the unfairness is due to backoff, we disabled backoff

1287: in the above scenario since the vast majority of losses were due to

1288: fading.  In this case, the two connections were able to get fair

1289: access to the medium across all distances (Figure~\ref{fair}).  This

1290: naive algorithm assumes that all losses are due to fading.

1291: Clearly, this is not a feasible solution, but it highlights the

1292: importance for discrimination between fading and contention losses.

1293: For example, possible discriminators such as channel utilization as a

1294: measure of contention or the use of physical level detection of

1295: collisions at the receiver with feedback to the sender are likely to

1296: provide effective discriminators for controlling the MAC backoff

1297: algorithm.  This is a topic of future work.

1298: 

1299: %

1300: %

1301: %

1302: %

1303: %

1304: %

1305: %

1306: 

1307: \section{Effects on Upper Layers}~\label{sec:routing}

1308: 

1309: A number of problems due to fading have been encountered in real

1310: testbeds and several solutions have been developed.  The most heavily

1311: studied problem is the issue of link quality and exposing that to the

1312: routing protocol to enable it to evaluate path

1313: quality~\cite{decouto-03a,aguayo-04}.  Moreover, the problem of

1314: discovering low quality hops that may become visible due to fading

1315: is well known~\cite{chin-02,maltz-00}.  The problem occurs if the protocol

1316: attempts to discover only high quality links -- lower quality ones are

1317: still occasionally visible to route discovery.  One of the proposed

1318: solutions is to use MAC filtering to assess the quality of the

1319: link before accepting it~\cite{chin-02,maltz-00,dube-97,goff-01}; this approach

1320: improves performance but does not eliminate the problem.   We were able to

1321: reproduce these problems in our simulations.   

1322: In this section, we outline additional problems and interactions that arise 

1323: at the upper protocol layers due to fading.  We are pursuing solutions to several of these problems.

1324: 

1325: 

1326: \subsection{Link and Route Qualities}

1327: Exposing link quality to the routing protocol is an important step

1328: towards effective evaluation of route qualities.  However, we believe

1329: that current link quality metrics and their combination into a route

1330: costs are not representative of actual behavior.  Packet delivery

1331: ratio based metrics such as ETX appear to perform better than delay

1332: metrics~\cite{draves-04,woo-03}.  The route quality is then obtained by

1333: adding up the costs of the individual links.  However, both link delay

1334: and throughput are not linear functions of packet delivery.  Simply

1335: adding the individual links does not provide an accurate estimate of

1336: the delay or expected throughput of the connection.

1337: 

1338: A more subtle effect occurs at the connection level as well.  Due to

1339: the unfairness problem, the problem of self-contention among hops of a

1340: single connection exhibits markedly different performance when

1341: fading is considered.  Specifically, under idealized assumptions,

1342: there is a bias towards hops closer to the source since these hops

1343: supply packets.  Thus, later hops cannot be unfair against earlier

1344: hops -- the best they can hope for is to match

1345: them~\cite{xu-02}.  When fading is considered, the

1346: effect of this problem is much more pronounced.  More specifically,

1347: because a better quality link can dominate the available bandwidth

1348: when contending with a lower quality one, the following effect is

1349: observed.  If a strong connection precedes a weaker one in a multi-hop

1350: chain, the strong connection will dominate the weaker one.  Thus,

1351: packets would be sent to the source of the weaker connection, and get

1352: dropped there because it is unable to get a share of the bandwidth to

1353: send the packets it received.

1354: 

1355: If the weaker link comes first, it has a regulating effect on the

1356: stronger one -- the stronger link is limited in terms of supply

1357: packets to what the weaker link can provide it.  Thus, the quality of

1358: the route is not simply the sum of the qualities of the links -- the

1359: order of the links has to be considered as well.  Alternative,

1360: sophisticated packet scheduling techniques can be used to bypass this

1361: problem.  We believe that addressing the root problem of MAC level

1362: unfairness will provide a better solution.  

1363: 

1364: \begin{figure}

1365: \centerline{\epsfig{file=order.eps,width=0.7\linewidth,silent=}}

1366: \caption{Effect of Hop Order}

1367: \label{order}

1368: \end{figure}

1369: 

1370: The effect of the link

1371: order on chain performance can be seen in Figure~\ref{order}.  In this experiment, the throughput of a two hop connection is tracked.  The figure shows 4 situations depending on the quality of the two links.  Most interesting is the performance of the middle scenarios.  A

1372: connection with a weak link followed by a strong one performance much

1373: better than a connection with the strong link first.  In fact, when the strong link comes first, the throughput is {\em worse} than a connection with 

1374: two weak links (the fourth scenario).  

1375: Our future research regarding this problem targets coming up with

1376: generalized route quality estimates based on link qualities and order.  

1377: 

1378: \subsection{Effect of Spurious Route Errors}

1379: 

1380: The dual of discovering unwanted low quality hops is mistakenly

1381: thinking that good quality ones are no longer there.  Some routing

1382: protocols such as DSR~\cite{johnson-01} assume that a failure to send

1383: a packet is due to mobility.  Again, this is due to idealized

1384: propagation assumptions causing packet losses to be almost always due

1385: to mobility.  However, packet failures due to repeated collisions have

1386: been shown to cause route errors under heavy loads as

1387: well~\cite{xu-02}.  Route errors can have a major effect on

1388: performance -- leading to expensive route searches and connection

1389: interruption while route discovery is accomplished.

1390: 

1391: Under fading, even in good quality links occasional packet failures

1392: do arise.  For example, even when the transmission delivery ratio is

1393: 90\%, there is approximately a 2\% chance of a packet failure with 4

1394: retries (as would happen with no RTS/CTS).  This would lead to route

1395: error getting generated every 50 packets; the effect is even worse

1396: when one considers a multi-hop connection -- each data packet has a

1397: chance of causing a route error at every hop.

1398: 

1399: The Link Quality Source Routing (LQSR) protocol provides a framework

1400: for managing this and similar problems~\cite{draves-04}.  LQSR

1401: constantly monitors path quality by providing feedback to the sender

1402: potentially with every packet delivered.  While LQSR still generates the

1403: Route Error Packet when delivery failure occurs, it does not interpret

1404: such a packet as a loss of route at the source -- simply, the source 

1405: penalizes the link (increasing its cost estimate).  

1406: 

1407: Such approaches are needed to filter out transient losses due to

1408: fading from real loss of route.  However, a large amount of

1409: overhead can be generated due to these route error packets.  We have

1410: developed an approach for locally discriminating between fading

1411: losses and loss of connection due to mobility.  Essentially, an

1412: exponential average of the delivery ratio along an active link is

1413: tracked.  When packet drops occur, they generate route errors only if

1414: the stable estimate of the link quality is low (indicating possibility

1415: of disconnection).

1416: 

1417: \subsection{Effect on MAC broadcasts and Route Discovery}

1418: 

1419: 

1420: 

1421: \begin{figure} 

1422: \centerline{\epsfig{file=macbcast.eps,width=0.7\linewidth,angle=270,silent=}}

1423: \caption{Effect on Network Wide Broadcast}

1424: \label{paul-fig}

1425: \end{figure}

1426:                                                                                 

1427: 

1428: Fading has especially high impact on broadcast operations.  Unlike

1429: unicast packets, MAC level broadcast is not acknowledged -- if a

1430: broadcast is loss, the loss is undetected and no retransmissions

1431: occur.  As a result, broadcast transmissions are only delivered at the

1432: link level transmission ratio (see Figure~\ref{fig-cbrdeliveryratiobycalc}), making

1433: them especially vulnerable to fading.  This has important

1434: implications on all protocols that use MAC broadcast such as flooding

1435: and group communication operations.  

1436: 

1437: Figure~\ref{paul-fig} shows the effect of this problem on flooding

1438: coverage as the network density increases.  In this scenario, a number of

1439: nodes are deployed randomly in a 1000x1000 meter area; the number of nodes is varied to increase the density.  We

1440: simulated the effect of fading using probabilistic packet drop to

1441: have control on the loss rate.  Clearly, the coverage of flooding

1442: suffers as the quality of the links drops.  This is especially true

1443: for sparse networks (and for sparse areas of networks).  Moreover, since many optimized Network-Wide

1444: Broadcast operations reduce the redundancy in flooding, they end up

1445: becoming more vulnerable to fading losses.  We characterize this

1446: problem and investigate solutions to it in another paper submitted to

1447: MobiHoc. 

1448: 

1449: Because of fading packets may fail multiple consecutive

1450: retransmissions and be dropped because the retransmission limit is

1451: reached.  Many routing protocols use MAC level transmission failure as

1452: an indicator that the link is no longer available (due to mobility).

1453: It is important to have more effective discrimination between packet

1454: losses due to fading and those due to mobility.  

1455: 

1456: 

1457: %

1458: %

1459: %

1460: %

1461: %

1462: 

1463: 

1464: 

1465: 

1466: 

1467: 

1468: 

1469: 

1470: 

1471: 

1472: 

1473: 

1474: %

1475: %

1476: 

1477: \section{Conclusions}~\label{sec:conclude}

1478: 

1479: In this paper, we presented the result of a bottom-up investigation of

1480: the effect of fading on MANET protocol performance.  We first did

1481: an analysis of the effect of fading on the MAC layer and showed

1482: that packet delivery ratio, packet delay and effective throughput all

1483: suffer as a result of fading.  

1484: 

1485: Based on this analysis, we made the case the Collision Avoidance (the

1486: RTS/CTS mechanism) is harmful for performance.  This case has two

1487: sides: under fading, it become significantly harder to deliver

1488: packets if 4 successive transmissions must be delivered correctly

1489: (RTS/CTS/DATA/ACK) vs. only 2 transmissions if RTS/CTS is not used.

1490: In addition, we showed that RTS/CTS has minor benefit in preventing

1491: collisions.  More specifically, with aggressive carrier sense (as is

1492: commonly used in commercial radios), CSMA can prevent all collisions

1493: that are also preventable by CA (and in fact, considerably more).  The

1494: benefit of RTS/CTS is then isolated to making the cost of contention

1495: for situations where an interferer is in range with either the sender

1496: or the receiver less (due to contention using the smaller RTS

1497: packets).  Overall, we believe that the large benefit in performance

1498: due to eliminating RTS/CTS is not recuperated by the small benefit to

1499: the cost of collisions for a subset of the collisions.

1500: 

1501: Another primary contribution of this paper is to identify an

1502: unfairness problem that arises due to fading.  More specifically,

1503: the MAC layer backoff algorithm presumes that all losses are due to

1504: contention.  Thus, losses that occur due to fading increase the

1505: backoff even when no contention exists.  Thus, weaker links are at a

1506: disadvantage to stronger links because they end up backing-off more

1507: frequently.  We showed that this problem can be severe even when the

1508: difference in the quality between the links is minor.  In the long

1509: term, we believe that the root of this problem must be attacked: the

1510: backoff algorithm should be able to discriminate between contention

1511: and transmission losses.  We are investigating techniques to estimate

1512: the contention to avoid the inefficiency and unfairness problems that

1513: arise due to it.  In addition, we outlined the effect of fading on

1514: broadcast and multicast operations.

1515: 

1516: We also presented an overview of additional problems that arise at the

1517: upper layers.  Some of these problems are known, but the solutions to

1518: them are ad hoc in nature and not systemic (for example, the problem

1519: of discovering low quality links; also the related problem of dropping

1520: good quality ones when a packet fails).  One of the problems we

1521: outlined is an artifact of the unfairness problem at the MAC layer.

1522: As a result of this problem, path quality is sensitive to the order of

1523: hops in a multi-hop connection.  More specifically, if a strong link

1524: is closer to the source than a weaker one, it ends up dominating it

1525: due to the unfairness problem: packets would constantly be delivered

1526: to the source of the weaker hop only to be dropped there.  However, if

1527: the weaker link comes first, the stronger link cannot dominate because

1528: its packet supply comes from the weaker link.  There is a need for a

1529: generalized approach for estimating link quality that takes such

1530: behavior into consideration.

1531: 

1532: The fading model we use has several limitations.  It treats links

1533: symmetrically and assumes that losses are independent; both these

1534: assumptions are not valid in real testbeds.  Nevertheless, we believe

1535: that the problems that are isolated are real but the magnitude of

1536: their effect may be different when a more accurate model is used.

1537: Moreover, we did not consider the effect of multi-rate MAC

1538: protocols~\cite{sadeghi-02,seok-multirate}.  Upgrading the propagation

1539: model to better reflect wireless propagation and generalizing our

1540: analysis is also a topic of future work.

1541: 

1542: \bibliography{local}

1543: \bibliographystyle{IEEE}

1544: \end{document}

1545: 

1546: 

1547: