Sarvaion Miigaion Through Muli-Channel Coordinaion in CSMA Muli-hop Wireless Neworks Jingpu Shi, Theodoros Salonidis, and Edward W. Knighly Deparmen of Elecrical and Compuer Engineering Rice Universiy, Houson, TX 775 {jingpu,hsalon,knighly}@ece.rice.edu ABSTRACT Exising muli-channel proocols have been demonsraed o significanly increase aggregae hroughpu compared o single-channel proocols. However, we show ha despie such improvemens in aggregae hroughpu, exising proocols can lead o flow sarvaion in a muli-hop nework, a phenomenon ha also occurs wih singlechannel proocols. In his paper, we devise Asynchronous Mulichannel Coordinaion Proocol (AMCP), a disribued medium access proocol ha no only increases aggregae hroughpu, bu more imporanly, addresses he fundamenal coordinaion problems ha lead o sarvaion. Based on AMCP s couner-sarvaion mechanisms, we analyically derive and experimenally validae an approximae lower bound on he hroughpu of any flow in an arbirary opology. We also demonsrae ha AMCP can deliver significanly higher per-flow hroughpu han boh IEEE 82.11 and exising muli-channel soluions. In addiion o is performance properies, AMCP is boh simple in ha i operaes using he primiives of IEEE 82.11 DCF, and cos-effecive in ha i requires only a single half-duplex ransceiver and no infrasrucure suppor. Caegories and Subjec Descripors C.2.1 [Compuer-Communicaion Neworks]: Nework Archiecure and Design Wireless communicaion General Terms Design, Performance Keywords CSMA, CSMA/CA, Muliple Channels, Sarvaion 1. INTRODUCTION Disribued CSMA-based random access proocols such as IEEE 82.11 DCF are well known o produce unfairness or even flow sarvaion when applied o muli-hop wireless neworks. The main This research is suppored by NSF ITR Grans ANI-33162 and ANI-325971, and by he Cisco ARTI program. Permission o make digial or hard copies of all or par of his work for personal or classroom use is graned wihou fee provided ha copies are no made or disribued for profi or commercial advanage and ha copies bear his noice and he full ciaion on he firs page. To copy oherwise, o republish, o pos on servers or o redisribue o liss, requires prior specific permission and/or a fee. MobiHoc 6, May 22 25, 26, Florence, Ialy. Copyrigh 26 ACM 1-59593-368-9/6/5...$5.. reason is ha in muli-hop seings no all ransmiers are wihin range of each oher and hence may have a differen view of he channel sae. Sarvaion can be addressed by appropriaely scheduling inerfering ransmissions over muliple orhogonal channels. Indeed, if every node were equipped wih a large number of channels and ransceivers, he access problem would be eliminaed and sarvaion would no occur. However, mos commercially available wireless cards have a single ransceiver and suppor a limied number of orhogonal channels. Scheduled access mehods [7, 16, 19, 25, 27, 29, 31] can operae under ransceiver and channel consrains and can address sarvaion by coordinaing ransmissions across muliple channels (be i ime slos, frequency bands, or spread specrum codes) in an opimal manner. However, such echniques ypically require global knowledge of opology and raffic requiremens as well as some form of infrasrucure suppor: eiher global ime slo synchronizaion (TDMA) or pre-disribuion of codes (CDMA). Such operaional requiremens canno be easily suppored in he disribued ad hoc nework seing. In his paper we uilize muliple channels o miigae sarvaion under he consrain of a single half-duplex radio a each node and he absence of cenralized knowledge or infrasrucure suppor. Our soluion, called Asynchronous Muli-channel Coordinaion Proocol (AMCP), uses he simple disribued access primiives of IEEE 82.11 DCF and provides analyical minimum rae guaranee for each flow in he nework. Despie is simpliciy, AMCP is designed o address he roo cause of sarvaion of CSMA proocols in singlechannel muli-hop wireless neworks as well as he dual coordinaion problems ha arise by he inroducion of muliple channels. While previous muli-channel MAC proocols [1, 2, 18, 24, 3, 33] have been shown o increase aggregae nework hroughpu, hey do no provide mechanisms ha preven sarvaion in muli-hop wireless neworks. We show ha wihou proper coordinaion of ransmissions, he aggregae hroughpu may increase wih he number of channels, bu cerain flows may sill receive zero hroughpu. We firs presen ha sarvaion in single-channel CSMA sysems arises due o misaligned ransmissions, which eiher cause a ransmier of a flow o defer hrough carrier sense for exended ime periods, or cause collisions a he flow s receiver. Muli-channel wireless echnologies have he poenial o address sarvaion by moving he misaligned inerfering ransmissions o differen channels. Sill, achieving his goal is challenging, especially when each node can ransmi or receive on only a single channel and link a a ime. Alhough packes can be ransmied on differen channels, ransmissions in a muli-channel sysem are sill no aligned. This resuls in he following generic muli-channel coordinaion problems: 1) Conrol packes sen on a cerain channel fail o inform neighboring nodes currenly communicaing on a differen channel and 2) Conrol packes inended for a cerain receiver may fail because he receiver is currenly on a differen channel. These prob-
lems may also lead o sarvaion if no addressed appropriaely. In ligh of he above coordinaion problems we revisi he basic design principles of muli-channel MAC proocols. A fundamenal design choice is wheher o use a dedicaed conrol channel or ransmi boh conrol and daa informaion on all channels. AMCP uilizes a dedicaed conrol channel o address boh single-channel and muli-channel coordinaion problems and effecively alleviae sarvaion in a muli-hop wireless nework. To comba he boleneck caused by he conrol channel, we compue he maximum number of daa channels ha can be suppored by he conrol channel as a funcion of he proocol parameers. This allows one o quaniaively perform appropriae sizing on he conrol channel capaciy. Nex, we derive an approximae lower bound on he hroughpu of any AMCP flow in an arbirary opology. The basic echnique is o consruc a hypoheical, low-hroughpu scenario on he conrol channel and o model he impac of he aggregae channel hopping paern of he inerfering flows. The lower bound depends on sysem parameers and he number of inerfering nodes wihin he neighborhood of each flow. Therefore, i can be compued using only local informaion. Through exensive simulaions we demonsrae he properies of AMCP in boh single-hop and muli-hop neworks. We show ha he hroughpu achieved by AMCP can approach he approximae lower bound in highly congesed conenion regions while being much higher in muli-hop scenarios. We design experimens o isolae and expose each fundamenal single- and muli-channel coordinaion problem and show how AMCP addresses he issue and describe why exising muli-channel soluions do no. As AMCP swiches channels a packe level, we evaluae via simulaions he performance degradaion of AMCP due o channel swiching delay. The remainder of he paper is organized as follows. In Secion 2 we presen he coordinaion problems ha resul in sarvaion in single channel sysems, hen poin ou he issues involved when muliple channels are used o address sarvaion. In Secion 3 and 4 we presen and analyze AMCP. In Secion 5 we evaluae he performance of AMCP hrough simulaions. Relaed work is discussed in Secion 6 and Secion 7 concludes. 2. MOTIVATION AND PROTOCOL DESIGN ISSUES In his secion, we firs presen Informaion Asymmery (IA) and Flow-in-he-Middle (FIM), wo coordinaion problems ha have been shown o cause sarvaion in single-channel CSMA muli-hop wireless neworks [12]. We hen show ha muliple channels can be used o address sarvaion and compare wo broad classes of soluions. Then we sudy wo generic coordinaion problems inheren in a muli-channel sysem, namely he Muli-Channel Hidden Terminal problem idenified in [3] and he Missing Receiver problem which we idenify in his paper. These muli-channel coordinaion problems manifes in boh classes of soluions and may cause performance degradaion if no addressed properly. 2.1 Sarvaion in CSMA single-channel mulihop wireless neworks When all ransmiers are wihin range of each oher i can be shown ha CSMA proocols provide fair access opporuniies o all flows. Unforunaely, in a muli-hop opology where no all nodes are wihin range of each oher, such proocols do no perform well, even if coordinaion enhancemens such as RTS/CTS conrol packe exchanges [3] are used. More specifically, hroughpu disribuions arise in which a few flows capure all bandwidh while many oher flows ge very low or even zero hroughpu. Such sarvaion phenomena are no merely due o having a differen number of conenders for each flow, which is naural in a muli-hop opology; raher, hey are due o coordinaion problems when CSMAbased access is used in a muli-hop environmen. Here we illusrae hese coordinaion problems ha cause sarvaion hrough wo characerisic examples. Informaion Asymmery (IA). The IA problem arises when he senders of wo conending flows are no wihin radio range and have an asymmeric view of he channel sae. Fig. 1(a) is an example opology of he IA problem ([3]), where he ransmier B of flow Bb is wihin radio range of he receiver a of flow Aa no in range of ransmier A. If boh flows are backlogged, flow Bb will receive significanly higher hroughpu han flow Aa. This is because he ransmier B of flow Bb knows exacly when o conend for he channel (hrough he conrol packes sen by he receiver of flow Aa). On he oher hand, sender A canno sense he aciviy of flow Bb and has o discover an available ime slo only hrough random back-off. Since for efficiency purposes he raio of daa ransmission inerval o he idle slo size is usually large, mos of hese random aemps occur during he ransmission of flow Bb and resul in collisions a receiver a. Repeaed collisions rigger imeous a sender A, which repeas doubling is conenion window. As a resul, he collision probabiliy of flow Aa is close o 1, while he collision probabiliy of flow Bb is close o. Figure 1(b) shows he channel sae experienced by flow Aa. Bb Aa A a B (a) Example Topology RTS(A) RTS(A) RTS(A) RTS(A) BO BO BO BO... (b) Flow aciviies: Flow Bb does no experience collisions. The random aemps of node A o find an idle inerval wihin he ransmissions of flow Bb resul in RTS failures and exponenial back-off. Figure 1: Informaion Asymmery (IA) example Flow-in-he-Middle (FIM). The FIM problem arises when he sender of a flow senses he aciviy of neighboring nodes ha are no wihin range wih respec o each oher. This behavior is illusraed in he hree-link scenario of Fig. 2. If all flows are backlogged, he middle flow Bb will receive very low hroughpu, while he ouer flows (Aa and Cc) will receive maximum hroughpu. This is no due o high loss probabiliy, bu raher o he lack of ransmission opporuniies for he middle flow. More specifically, when one of he ouer flows (say flow Aa) capures he medium, he ransmier of he middle flow Bb will sense and defer bu he ransmier of he oher ouer flow Cc will coninue conending and iniiae ransmission. When flow Aa ends ransmission, i will conend and iniiae ransmission, while flow Bb now defers due o flow Cc. Fig. 2(b) shows ha he misaligned concurren ransmissions of he ouer flows may be sensed by he ransmier B of he middle flow for exended periods of ime. The middle flow has a chance o access he medium only when boh ouer flows are in he back-off phase (he verical lines inerval in Fig. 2(b)). Unforu- b
naely, such occurrences become increasingly rare especially as he raio of daa ransmission inerval o he back-off inerval increases. Channel 2 RTS(B) CTS(b) RTS(B)CTS(b) b Channel 1 RTS(A) CTS(a) RTS(A)CTS(a) a A B C c (a) Soluion1 (a) Example Topology Daa Channel 2 Aa Daa Channel 1 Bb RTS(B)CTS(b) Conrol Channel RTS(A) CTS(a) RTS(B)CTS(b) Cc (b) Soluion2 (b) Channel aciviy sensed by he middle flow. Figure 3: The use of muliple channels o address sarvaion. Figure 2: Flow In The Middle (FIM) example Boh IA and FIM problems are no specific o he 82.11 DCF access mechanism. They are generic coordinaion problems ha arise due o he asymmery of he muli-hop opology and due o he use of carrier sense. In a general opology he sarving flows experience he combined effec of boh IA and FIM problems and heir hroughpu may even reach zero. For an analyical model of sarvaion phenomena in single-channel CSMA muli-hop neworks, see [12]. For convenience, in he res of he paper we use he erm advanaged flows o refer o flows wih geomery advanage (such as flow Bb in Fig. 1(a) and he ouer flows Aa and Cc in Fig. 2(a)) and he erm disadvanaged flows o refer o flows wih geomery disadvanage (such as flow Aa in Fig. 1(a) and flow Bb in Fig. 2(a)). We also mainain he convenion of using capial leer for he ransmier and lowercase leer for he receiver of each flow. 2.2 Sarvaion avoidance hrough muliple channels In boh he IA and FIM sarvaion scenarios, he disadvanaged flow is unable o idenify an idle inerval because ransmissions are generally misaligned and heir duraions are much larger han he back-off inerval. Clearly, sarvaion would be eliminaed if all ransmissions occurred on orhogonal channels. Poenial soluions can be classified ino wo approaches, as exemplified in Fig. 3 for he case of wo flows. In he firs approach, he enire ransmission (including conrol and daa ransmissions) of each flow is scheduled on a differen channel (Fig. 3(a)). The reason his approach can avoid sarvaion is sraighforward: an advanaged flow will no sarve a disadvanaged flow because hey boh ransmi on differen channels. In he second approach, conrol packes are ransmied on a separae conrol channel and daa packes of differen flows are disribued o differen daa channels (Fig. 3(b)). This approach also alleviaes sarvaion: as he daa packes have moved o differen channels, conenion occurs only on he conrol channel beween shor conrol packes, whose lengh is comparable o he back-off inerval. Each approach has is advanages and disadvanages. The advanage of he firs approach is ha i does no require he overhead of a dedicaed channel for conrol messages and can poenially reduce he conenion beween he advanaged flows and he disadvanaged flows o zero. However, i can lead o logical pariion where wo nodes wihin range are unable o communicae. This is a significan challenge, especially when a node has only one ransceiver and can only ransmi or lisen o one channel a a ime. In he second approach nodes immediaely reurn o he conrol channel afer finishing heir daa ransmissions. The advanage of his approach is ha nodes have a common channel (bu no ime) reference o coordinae heir ransmissions. The downside is ha a dedicaed conrol channel inroduces overhead, which can be significan if is capaciy is no appropriaely designed. 2.3 Muli-channel coordinaion problems Regardless of he soluion approach, i is challenging o coordinae ransmissions over differen channels in an asynchronous seing where each node has a single radio ransceiver. Transmissions occurring on differen channels can sill be misaligned. When a node communicaes on a channel, i is no aware of he sae on oher channels. Hence, when i finishes communicaion i may aemp o exchange informaion wih is neighbors while hey are currenly on oher channels. To design an efficien proocol we mus be able o accuraely characerize his lack of coordinaion. We invesigae wo generic coordinaion problems, classified wih respec o heir effec on he inended acions of conrol packes. In he Muli-channel Hidden Terminal Problem, conrol packes sen on a cerain channel fail o inform neighboring nodes currenly communicaing on a differen channel. An insance of his generic problem was firs idenified in [3]. To illusrae his problem we use a naive proocol ha is a sraighforward exension o IEEE 82.11 DCF for a muli-channel seing: The RTS/CTS conrol packes are exchanged on a dedicaed conrol channel and reserve daa channels for daa packes. Nodes reurn o conrol channel immediaely afer hey finish heir daa ransmissions. Now we consider again he wo-flow opology of he IA scenario in Fig. 1. In his example, we assume ha he proocol operaes
wih wo daa channels. As shown in Fig. 4, a conrol packe exchange of disadvanaged flow Aa may occur when he advanaged flow Bb ransmis on daa channel 2. Suppose Aa selecs daa channel 1 and iniiaes a ransmission. When flow Aa ransmis, flow Bb will reurn o he conrol channel. Since i has no heard he reservaion of Flow Aa, i may selec daa channel 1. In his case, flow Aa will experience a collision, while he ransmission of Bb succeeds. Flow Aa can be sarved if here are many advanaged flows wihin is radio range. Daa Channel 2 Daa Channel 1 Conrol Channel RTS(A)CTS(a) RTS(B)CTS(b) Figure 4: The Muli-Channel Hidden Terminal Problem. Alhough in his example we used a naive proocol where all conrol messages are exchanged in a dedicaed conrol channel, i is eviden ha he problem is also presen when conrol messages are ransmied on differen channels. The muli-channel hidden erminal problem limis he abiliy of conrol packes o block inerfering flows. If no proper measures are aken i may resul in very poor performance. The Missing Receiver Problem arises when conrol packes sen on a cerain channel o access an inended receiver fail because his node is currenly on a differen channel (acing eiher as ransmier or receiver). To illusrae he problem, we consider he simple hree-node scenario of Fig. 5, where node A ransmis o node B and node B ransmis o node C. We firs consider he naive proocol version where all conrol messages are ransmied on differen channels. In Fig. 5, an access aemp of A for B on channel 1 will fail if B is on channel 2. Then node A will perform random back-off and rery on channel 1. Unless proper measures are aken, his problem will cause large packe delay for flow AB and decrease is hroughpu. A B Channel 1 Channel 2 Figure 5: Missing Receiver Problem. The problem also persiss wih a proocol ha separaes he conrol channel from daa channels. Suppose A sars conending for B and B sars conending for C on he conrol channel. As long as one of hem wins he conenion, he oher node will be able o synchronize and resume conenion a he end of he daa ransmission. Unforunaely, synchronizaion is los when he nodes coun-down simulaneously. In his case, boh nodes will no be able o hear each oher s RTS while hey ransmi. Therefore he RTS from B o C succeeds, while he RTS from A o B fails. Afer his poin, node A will ry o discover node B using random back-off. This is difficul o occur since A will need o find a shor inerval where B reurns for is own back-off on he conrol channel. I is more likely for B o conac A when i conends in he conrol channel for he nex packe for C. In his case, A synchronizes wih he end of ransmission of B bu i will already have a large back-off inerval and will no be able o compee fairly for B. Hence flow AB will sarve if no proper measures are aken. C I is eviden ha similar inefficiencies arise in he oher version of he Missing Receiver Problem, where node B acs as receiver on link BC. Noe ha he Missing Receiver Problem does no exis in a singlechannel sysem because A can carrier sense he daa ransmissions of B and immediaely defer unil he end of BC ransmission. 3. ASYNCHRONOUS MULTI-CHANNEL CO- ORDINATION PROTOCOL (AMCP) We firs illusrae he basic principles of AMCP and hen presen is implemenaion. Finally, we show how i addresses he mulichannel coordinaion problems. 3.1 Overview Following he second approach of Secion 2.2, AMCP uses a dedicaed conrol channel on which nodes conend o reserve daa channels by exchanging RTS/CTS packes according o 82.11 DCF. Upon successful conrol packe exchange, boh he sender and he receiver swich o he reserved daa channel, denoed by x, and ransmi a daa packe. Afer a daa packe is successfully ransmied on channel x, he sender and receiver reurn o he conrol channel and se all channels as unavailable excep x. They may conend for daa channel x immediaely or conend for oher daa channels afer he imers of hese channels expire. The RTS/CTS conrol packes serve a dual purpose: firs, hey aid wo link endpoins o negoiae on commonly available daa channels; second, hey inform neighboring nodes o se he overheard daa channels piggy-backed in RTS/CTS as unavailable for an enire daa ransmission inerval. However, a node overhearing an RTS/CTS will no always defer for he enire daa ransmission; under cerain condiions, i may iniiae conenion afer he overheard RTS/CTS. The exac deferring rules (described in Secion 3.2) implemen an efficien coordinaion scheme where nodes say on he conrol channel long enough o learn abou which channels o compee, while a he same ime no always waiing for he enire daa packe ransmission, hus increasing hroughpu. We proceed o describe he exac proocol operaions. 3.2 Proocol Descripion 3.2.1 Srucures and variables We assume one conrol channel and N daa channels, indexed from 1 o N. All channels are orhogonal wih respec o each oher. Each node has a single ransceiver, hence i can eiher ransmi or lisen, bu no boh. Also i can lisen o or ransmi on one channel a a ime. To execue AMCP, each node mainains he following srucures and variables: A local N-enry Channel Table. Each able enry corresponds o a daa channel and consiss of a bi called avail bi indicaing channel availabiliy, and a imer called avail imer indicaing he remaining ime a channel is no available. Each ime he channel becomes unavailable (avail bi = ), is imer is se o expire afer a daa ransmission duraion. When he imer expires, he corresponding channel becomes available (avail bi = 1). By defaul, when a node joins he nework all is avail bis are se o zero. An ineger prefer variable akes values from o N. If nonzero, his variable indicaes ha a node prefers o compee for he daa channel indexed by prefer. If zero i indicaes no preference.
3.2.2 Reservaion/ransmission cycle Iniially all nodes reside on he conrol channel. We now describe he proocol acions ha occur when node A has a packe inended o node a. We denoe a neighboring node of A or a as node C. Sep 1: Channel selecion. Node A selecs a daa channel by inspecing is channel able. Among he available daa channels, he channel indexed by prefer is seleced if prefer is non-zero and available. Oherwise one of he available daa channels is randomly chosen. If no daa channel is available, he node wais unil any of he avail imers expires. Sep2: Channel conenion. Suppose ha daa channel x is seleced. Node A insers he index x o is RTS packe and conends on he conrol channel using he 82.11 DCF CSMA/CA mechanism. In AMCP, a conrol channel s NAV inerval expires a he end of a RTS/CTS ransmission, raher han he end of a DATA/ACK ransmission as in IEEE 82.11. Sep 3: Channel negoiaion. When node a receives he RTS packe, i inspecs he saus of channel x in is channel able. If x is available, node a replies o A wih a Confirming CTS packe conaining index x. Then, i swiches o daa channel x and wais for a DATA packe. If channel x is no available, node a replies o A wih a Rejecing CTS packe conaining index and a lis of is available daa channels, and remains on he conrol channel. If node A receives a Confirming CTS, i swiches o channel x and ransmis he DATA packe o a. If A receives a Rejecing CTS, i randomly selecs a channel available in boh is channel able and he channel lis included in he CTS packe, hen i insers he index of his channel in a RTS packe and begins a new conenion cycle on he conrol channel. Sep 4: Daa ransmission. Upon recepion of he DATA packe, node a responds wih an ACK on daa channel x, hen swiches back o he conrol channel. Upon recepion of he ACK packe, A also swiches back o he conrol channel. The packe ransmission has compleed successfully. Sep 5: Seing channel availabiliy. Afer A reurns o he conrol channel i ses is prefer variable o x; A also ses he avail bi unavailable and sars avail imer for all oher daa channels excep x. Node a ses is prefer variable and Channel Table in he same way. Node A resars sep 1 if here is a packe in is ransmission queue. We noe ha errors in he ransmied conrol and daa packes are handled wih imeou mechanisms similar o 82.11. If a imeou occurs while a node resides on a daa channel, he node reurns o he conrol channel, ses is prefer variable o, ses he avail bi unavailable and sars avail imer for all daa channels. 3.2.3 Overhearing nodes deferral rules Le C be a neighbor of eiher A or a. When C overhears an RTS packe, i firs updaes is channel able by seing is avail bi(x) = and ses avail imer(x) o expire a he end he full daa packe ransmission (for a duraion equal o CTS + DATA + ACK). When node C hears a Confirming CTS, i ses is avail bi(x) = and sars avail imer(x) in he same way. When i hears a Rejecing CTS, no acion is needed. Noe ha his deferring rule is only wih respec o channel x. Node C can compee for oher available channels afer deferring for he duraion of an RTS/CTS exchange. There is only one excepion o he above deferring rules. When C wans o ransmi o A and hears an RTS from A, inended o a, i will defer unil he end of he enire ransmission of flow Aa, and se is conenion window size o he minimum value. Similarly, when C wans o ransmi o a and hears a CTS from a inended o A, i will defer unil he end of he enire ransmission of flow Aa and se is conenion window size o he minimum value. This scheme provides an opporuniy for C o address he Missing Receiver Problem. 3.3 Addressing muli-channel coordinaion problems To presen how AMCP solves he coordinaion problems described in 2.3, we consider he opology in Figure 1(a) and suppose here are 2 daa channels and 1 conrol channel. Muli-channel Hidden Terminal Problem. Consider again he Muli-channel Hidden Terminal Problem example of Fig. 4. Recall ha when flow Bb arrives on he conrol channel during Aa ransmission on daa channel 1, i does no have sufficien informaion abou he sae of channel 1 because i has no heard he RTS/CTS packe of flow Aa while ransmiing is own daa packe on daa channel 2. If i selecs channel 1 i will cause a collision o he disadvanaged flow Aa. Under AMCP, node B ses channel 1 as unavailable and ses a imer o expire afer he duraion of a RTS/CTS/DATA/ACK ransmission (as specified in sep 5 of proocol operaions). Noe ha channel 1 may or may no acually be available, bu node B ses i o unavailable, precisely because i does no have his informaion. Node B will compee for channel 1 only afer he imer expires by ha ime any ransmission on channel 1 will have compleed. If any RTS/CTS for channel 1 is heard during his period, node B will defer furher bu will have synchronized for conenion on channel 1. However, node B does no necessarily remain idle afer he channel 1 imer is se. Since is ransmission on channel 2 was successful, his channel is available and B will sar conending immediaely for his channel (hrough is prefer variable). Fig. 6(a) shows he scenario where B succeeds in capuring channel 2. In case B fails due o anoher flow Cc ha conended for channel 2, i will also se a imer for channel 2 and defer conenion unil eiher of he wo channel imers expires. The wors-case scenario for flow Bb upon is arrival on he conrol channel is depiced in Fig. 6(b). Here flow Cc wins channel 2 and hen flow Aa wins channel 1 before he channel 1 imer expiraion. Alhough B has los boh conenions, i has synchronized on boh daa channels and will conend when eiher of hese ransmissions ends. Flow Bb has an advanage in capuring eiher channel in fuure access aemps: i can compee for boh channels, couning down a single back-off couner. On he oher hand, each of flows Aa and Cc will only compee for is preferred channel, according o a fresh back-off couner. Summarizing, he simple waiing scheme of AMCP on he conrol channel effecively addresses he Muli-channel Hidden Terminal Problem by providing fair channel access opporuniies o conending flows. Missing Receiver Problem. Consider he scenario shown in Fig. 5, where A wans o ransmi o B when B is ransmiing o C on a differen channel. AMCP handles he Missing Receiver Problem as follows. If A receives from B an RTS inended o C, A will defer unil he end of he ongoing ransmission of B and examine is back-off sage. If i is already in high back-off sage, A ses is conenion window size o he minimum value. In his way, A will fairly conend for he aenion of B when B is in idle sae. In conras, in he naive proocol, B will ransmi many packes before A decremens is back-off couner o zero. Wih AMCP, he key reason of A quickly synchronizing wih B is ha all conrol messages are ransmied on a dedicaed conrol channel, where A can hear anoher RTS from B when A s firs RTS o B collides wih he RTS of B o C. In case node B is he receiver on link BC, node A performs he same acions as above when i hears he CTS of B o C. Therefore, AMCP effecively addresses boh manifesaions of he Missing Receiver Problem. 4. PROTOCOL ANALYSIS We now derive he analyical properies of AMCP. In Secion
Daa Channel 2 can drive more daa channels if we reserve a daa channel for muliple daa packes. Also noe ha M derived above is for a single conenion region. In a muli-hop nework, M can be much larger because he conrol channel is spaially reused. Daa Channel 1 Conrol Channel RTS(A)CTS(a) RTS(B)CTS(b) Tsuccess (a) Flow Bb conends and capures channel 2, upon reurn o he conrol channel. Daa Channel 2 Daa Channel 1 Conrol Channel RTS(A) CTS(a) RTS(C)CTS(c) DATA(C)+ACK(c) Tsynch RTS(A) CTS(a) (b) Flow Bb synchronizes o conend on boh channel 1 and channel 2, upon reurn o he conrol channel. Figure 6: AMCP addressing he Muli-channel Hidden Terminal problem. 4.1 we derive he maximum number of daa channels ha can be suppored by he conrol channel. In Secion 4.2 we derive an approximae lower bound on he hroughpu achieved by any flow in an arbirary opology. 4.1 Boleneck Analysis For any muli-channel proocol having a dedicaed conrol channel, wo poenial bolenecks exis: he number of daa channels and he bandwidh of he conrol channel. Le M be he he maximum number of daa channels occupied by daa ransmissions when he conrol channel is sauraed by conrol message exchanges. Le T d be he duraion of a successful daa ransmission (including DATA and ACK), T r be he duraion of an RTS packe, and T c be he duraion of an CTS packe, all expressed as ime slos. We observe ha in seady sae when he conrol channel is sauraed, here are always M on-going ransmissions on M daa channels. Furhermore, wihin he ime period of T d + T r + T c, exacly M flows reurn o he conrol channel. Hence in seady sae, M flows should successfully exchange conrol packes and swich o M respecive daa channels. Since he conrol channel is sauraed, here is no idle inerval beween wo successive RTS/CTS exchanges, if we neglec small overhead inervals, such as SIFS and DIFS. Thus M is given by, M = T d + T r + T c T r + T c. (1) From Equaion (1), we observe ha M increases when he daa ransmission ime T d increases. For example, he conrol channel 4.2 Lower Bound Analysis In his secion, we compue an approximae lower bound of perflow hroughpu achieved by our proocol in an arbirary muli-hop wireless nework. We firs consruc a hypoheical, low-hroughpu scenario for a agged flow, hen compue is collision probabiliy p by modeling he process by which conrol packes of oher nodes arrive on he conrol channel as a Poisson process. We hen apply he analyical model proposed in [13] o compue he hroughpu of he agged flow, which serves as an approximae lower bound on he hroughpu achieved by any flows in an arbirary opology. Consrucion of he hypoheical scenario. We consider a agged flow Aa ha has N neighboring nodes in a nework employing AMCP. We consruc he scenario where flow Aa achieves very low hroughou given N neighbors as he case ha all of is N neighbors are backlogged and always ransmi o receivers ha are no in range of Aa using he minimum back-off window. We also assume ha hese N nodes are ransmiing independenly, in he sense ha hey do no sense and hence coordinae wih each oher s ransmission. Furhermore, hey are placed such ha hey are advanaged wih respec o flow Aa. More specifically, we consider an IA scenario where hese N nodes are wihin range of receiver a and ou of range of ransmier A. In his scenario, mos conrol packes of flow Aa will collide, hus forcing flow Aa o double is conenion window. Noice ha ransmier A is no able o sense he aciviy of he inerfering flows. This scenario is hypoheical and only used o derive an approximae lower bound of he hroughpu of flow Aa given N neighboring nodes in is conenion region. Since in his scenario he inerfering nodes ransmi independenly, heir conrol packes arrive on he conrol channel independenly. Consequenly, we assume ha he aggregae process formed by he conrol packe arrivals of he N inerfering nodes is Poisson. While his process is no sricly Poisson, we validae he bound via simulaions below. Compuaion of he condiional packe loss probabiliy. To compue he hroughpu of flow Aa in he hypoheical, low-hroughpu scenario, we firs need o compue he collision probabiliy p when node A aemps o ransmi an RTS packe o a. Similar o [4], we refer o p as he condiional collision probabiliy. Le X() be he Poisson process ha represens he number of successful conrol packe arrivals of he N inerfering nodes, given a saring poin in ime. Le α be he arrival rae of conrol packes and T be he arrival inerval. Noe ha α is a deerminisic value and T is a random variable. We assume nodes can always find a daa channel o ransmi a daa packe upon successful RTS/CTS exchange. The arrival rae α of X() is given by: α = N T d + T r + T c. (2) Since X() is a Poisson process, any inerval T beween wo successive conrol packe exchanges of he inerfering flows is exponenially disribued wih he following CDF, F T() = P(T ) = 1 e α. (3) The RTS/CTS exchange beween A and a will fail if i canno fi wihin an idle gap T (T r + Tc) beween wo successive conrol packe exchanges. This corresponds o he even T (T r + Tc) < T r (or T < 2T r + Tc), which occurs wih probabiliy p = F T(2T r +Tc). Combining wih Equaions (2) and (3), we derive he final expression for he condiional packe loss probabiliy
p: p = 1 e (2Tr+Tc) T d +Tr+Tc. (4) Throughpu compuaion. We compue he hroughpu of he agged flow Aa using a general model for backlogged flows sharing an 82.11 muli-hop nework inroduced in [13]. In ha model, he channel view of each node comprises of a sequence of ime inervals ha correspond o 4 differen saes: (i) idle channel; (ii) channel occupied by successful ransmission of he agged saion; (iii) channel occupied by a collision of he saion; (iv) busy channel due o aciviy of oher saions, deeced by means of eiher physical or virual carrier sensing (he NAV). The ime inervals during which he saion remains in each of he four saes above are denoed by σ, T s, T c, and T b, respecively. According o he model in [13], he hroughpu of he agged flow Aa is given by: τ(1 p) T P = τ(1 p) T s + τp T c + (1 τ)(1 b)σ + (1 τ) b T, b (5) where τ is he probabiliy ha he node aemps o send a packe afer an idle slo, b is he probabiliy ha he channel becomes busy afer an idle slo due o aciviy of oher nodes and p is he condiional packe loss probabiliy. The probabiliy τ is a deerminisic funcion of p and is given by [2]: 2q(1 p m+1 ) τ = q(1 p m+1 ) + W ˆ1 p p(2p) m `1 + p q, m m (6) where q = 1 2p, W is he minimum window size, m is he maximum rery limi, and m is he backoff sage a which he window size reaches is maximum value. The average duraions T s and T c are fixed and can be found in [4]. In his hypoheical scenario, he ransmier node A does no defer is ransmission due o he aciviy of oher nodes. Seing b = in Equaion ((5)) yields: T P = τ(1 p) τ(1 p) T s + τp T c + (1 τ)σ + (1 τ), (7) Using Equaions (4) and (6), in Equaion (7), we can now compue he hroughpu of he agged flow Aa in he hypoheical scenario which serves as a lower bound approximaion on he hroughpu achieved by any flow in an arbirary opology as a funcion of number of inerfering flows and sysem parameers. Lower bound validaion. We now validae he approximae lower bound wih simulaions obained wih ns. Boh RTS/CTS packes and daa packes are ransmied a 2 Mbps. We vary he number of flows N and place hem in a 7m 7m area such ha hey belong o he same conenion region. This means ha only one flow can ransmi successfully a a ime, however i is no necessary ha all ransmiers or receivers are wihin range. For each N, we generae 1 daa poins each corresponding o he minimum rae achieved by a differen conenion region. Fig. 7 shows he minimum raes as daa poins and he lower bound as he analyical curve, as compued by our model. We observe ha in general he minimum raes are greaer han he lower bound while in several cases he bound is igh. 5. PERFORMANCE EVALUATION We evaluae AMCP in boh single-hop and muli-hop opologies using he ns-2 simulaor wih CMU wireless exensions. Unless oherwise specified we use he MAC parameers of Table 1. According o hese parameers, he maximum rae achieved by a backlogged flow in isolaion is 184 pk/s. The simulaor physical N Throughpu (pk/s) 2 18 16 14 12 1 8 6 4 2 Prediced Bound Simulaion 2 4 6 8 1 12 14 16 18 2 Number of flows Figure 7: Comparison of lower bound o minimum hroughpu achieved in an arbirary conenion region as a funcion of he number of inerfering flows. layer parameers have been se so ha he ransmission range of each node is approximaely 25m. SIFS 1 µs DIFS 5 µs EIFS 364 µs σ 2 µs BasicRae 2 Mbps DaaRae 2 Mbps PLCP lengh 192 bis @ 1 Mbps MAC header (RTS,CTS,ACK,DATA) (2,14,14,28) byes @ BasicRae Packe size 1 byes (CW min, CWmax) (31,123) Rery Limi (Shor,Long) (7,4) Channel swiching delay 224µs MMAC ATIM window 2ms MMAC Beacon inerval 1ms Table 1: MAC layer parameers We begin wih experimens on single-hop opologies o sudy he main proocol properies and illusrae he inerplay beween various parameers number of channels, raffic load, conrol channel capaciy, number of nodes, channel swiching delay ha affec performance. Performance is measured in erms of aggregae hroughpu gain wih respec o IEEE 82.11 DCF using a single channel. We hen move o muli-hop opologies, where we demonsrae he properies of AMCP: sarvaion miigaion, increase of aggregae uilizaion and addressing he fundamenal coordinaion problems of boh single channel and muli-channel sysems, as elaboraed in secion 2. We also compare AMCP wih MMAC, a singleradio, muli-channel proocol proposed in [3]. MMAC uses a globally synchronized conrol/daa periodic frame (ermed beacon inerval). During he conrol subframe (ermed ATIM window) flows conend on a defaul channel o reserve channels (including he defaul channel) for he daa subframe. The flows ha succeed in reserving a channel during he ATIM window conend during he daa subframe using RTS/CTS 82.11 access mechanism. Our experimens use he same MMAC parameers as [3] (Table 1). 5.1 Single-hop opologies In his series of experimens all nodes are wihin range of each oher and are equally divided in a ransmier and receiver se. This yields a se of single-hop disjoin flows wih disinc ransmierreceiver pairs. The case where a node is boh sender and receiver is considered in he muli-hop experimens. Effec of number of channels. Fig. 8 depics he aggregae hroughpu achieved by AMCP as a funcion of he oal number of channels for 15 backlogged flows (3 nodes). The capaciy of
he conrol channel and each daa channel is 2 Mpbs. The case of AMCP wih 2 channels is equivalen o single-channel 82.11, which provides he reference line in Fig. 8. The aggregae hroughpu increases linearly unil 7 channels. Afer ha poin, i increases wih a slower rae wih addiional channels; a 8 channels i reaches he limi of 11 pk/s where he conrol channel is sauraed. This behavior agrees wih our boleneck analysis: for he parameers in his experimen, Equaion (1) predics ha he conrol channel can drive up o 8 daa channels. µs is small compared o he duraion of a daa ransmission. Afer 3ms, hroughpu goes below he single-channel maximum hroughpu of 184 pk/s. For hardware wih such high channel swiching delays, he overhead can be addressed by reserving a channel for muliple daa packes. Such funcionaliy is easy o incorporae in he AMCP channel reservaion mechanism. 5 Throughpu (pk/s) 14 12 1 8 6 4 2 82.11 AMCP 2 4 6 8 1 12 Number of Channels Aggregae hroughpu (pk/s) 4 3 2 1 1 2 5 1 2 5 Channel swiching ime (us) Figure 1: Effec of channel swiching delay. Figure 8: Throughpu as a funcion of number of channels. Effec of raffic load. We evaluae he performance of AMCP under non-backlogged condiions. Fig. 9 depics he aggregae hroughpu of AMCP and IEEE 82.11 in a 15-flow opology as he inpu rae of each flow increases, when a oal of 4 channels are used. Unil 1 packes/s, he load is oo low o exploi he addiional daa channels and AMCP yields similar performance o 82.11. Afer ha poin, channelizaion becomes effecive and AMCP reaches an aggregae hroughpu gain equal o he number of daa channels. We noe ha exising muli-channel MAC proocols can achieve similar or slighly higher aggregae hroughpu han AMCP. For example, for 4 channels and under heavy load, DCA [33] also achieves hree imes he aggregae hroughpu of 82.11, similar o AMCP. This is because boh AMCP and DCA dedicae a separae channel for conrol raffic. On he oher hand, MMAC ransmis conrol and daa packes over 4 channels and achieves an addiional gain of 2%-3%. However, DCA requires wo radio ransceivers per node and MMAC requires global synchronizaion. AMCP uses a single ransceiver and no global synchronizaion. Throughpu (pk/s) 8 7 6 5 4 3 2 1 1 1 82.11 AMCP 1 Packe arrival rae per flow (pck/sec) 1 Figure 9: Aggregae hroughpu when arrival rae varies. Effec of channel swiching delay. Since AMCP swiches channels a he packe level, channel swiching delay due o hardware limiaions can be a source of overhead. According o he IEEE 82.11 specificaion [14] his parameer can reach 224µs. Fig. 1 shows a graceful decrease of aggregae hroughpu as channel swiching delay increases from o 5ms. A 224µs, he hroughpu decrease is very small. This can be explained by he fac ha 224 5.2 Muli-hop opologies In his series of experimens we compare he performance of AMCP, MMAC, and single-channel 82.11 in saic and mobile muli-hop opologies using boh single-hop and muli-hop flows. We also consider specific scenarios ha isolae inefficiencies ha arise in he design of muli-channel proocols, namely he random channel selecion problem due o collisions of conrol packes and he head-of-line (HOL) problem due o lack of packes o fill a channel reservaion window. 5.2.1 Single-hop flows Single-channel sarvaion scenarios. We firs invesigae he abiliy of AMCP and MMAC o address he IA and FIM coordinaion problems (Fig. 1 and Fig. 2, respecively). These scenarios can easily be addressed by random channel selecion if a large number of channels are available. Here we consider he case when a oal of hree channels are available. We observe from Fig. 11(a) ha AMCP provides equal and maximum hroughpu o each flow, despie ha, opologically, flow Bb has more informaion abou he channel. Furhermore, he simulaion shows ha he wo flows persis ransmiing on differen channels. This is a desired propery and shows ha AMCP successfully separaes he wo flows and reduces heir ineracion. Under MMAC, flow Bb achieves 8% he maximum hroughpu of 82.11 and AMCP. This is he maximum hroughpu allowed by MMAC since he ATIM window is 2% of he beacon period. However, he key observaion is ha he disadvanaged flow Aa receives only 2/3 of he maximum MMAC hroughpu. This is because he IA problem sill exiss in boh he conrol subframe and he daa subframe: he ATIM packe size is comparable o he backoff window size; 1 since is conrol packes collide, he ransmier of flow Aa is no informed abou channel reservaions in is neighborhood and is forced o perform random channel selecion. Aa may choose he same daa channel as Bb and, consequenly, is daa packes may be desroyed due o he IA problem. Similarly, in Fig. 11(b) AMCP provides equal and maximum hroughpu o all flows. As in he IA scenario, he flows quickly coordinae and keep ransmiing on he righ channels: flow Aa and flow Cc on one daa channel and he middle flow Bb on he oher. In conras, MMAC does no equalize he hroughpus bu is again 1 To allow more ATIM conrol packe exchanges in he 2 ms ATIM window, he back-off window size can no be se oo small.
Throughpu (pk/s) 25 2 15 1 5 AMCP 1 2 MMAC 1 2 Flow ID 82.11 (a) IA scenario (Fig. 1(a)). 1 and 2 denoe flow Aa and Bb, respecively. 1 2 82.11 as well as he AMCP lower bound, all wih respec o he AMCP hroughpu sored in decreasing order. Throughpu (pk/s) 25 2 15 1 5 AMCP MMAC 82.11 AMCP bound 1 2 3 4 5 Flow ID Throughpu (pk/s) 25 2 15 1 5 AMCP 1 2 3 MMAC 1 2 3 Flow ID 82.11 1 2 3 (b) FIM scenario (Fig. 2(a)). 1, 2 and 3 denoe flow Aa, Bb and Cc, respecively. Figure 11: AMCP performance in he basic single-channel sarvaion scenarios. subjec o random channel selecion: in his case, he ransmier of he middle flow is no able o decode he colliding ATIM conrol packes of he ouer flows. We now explain how AMCP addresses he IA and FIM problems. In he IA scenario, flow Bb does no experience collisions and will persis ransmiing on one of he wo daa channels (e.g. channel 1). The receiver of flow Aa will be informed abou his decision hrough he conrol packes of flow Bb. The ransmier A of flow Aa sars wihou any knowledge of which channel o use. Since conenion in he conrol channel has been reduced by he removal of daa packes, i is easier for A o access he receiver. In case i picked channel 1, A will be informed by he receiver and will compee and acquire channel 2 in is nex access aemp. Afer ha poin boh flows will coninue ransmiing on differen channels. In he FIM scenario flows do no experience collisions and herefore prefer o ransmi on he same channel. Since he ouer flows are no wihin range here may be an undesirable siuaion where hey have preference for channels 1 and 2, respecively. In his case he middle flow Bb is blocked bu only emporarily, unil is channel imers expire. I will hen conend on he conrol channel for any of he wo daa channels. When i acquires any of he wo channels (e.g. channel 1) he ouer flows are informed and will compee for channel 2. From his poin on, since he flows do no experience collisions hey will coninue ransmiing on orhogonal channels. Arbirary opology / single-hop flows. We now consider an arbirary opology of 1 nodes placed in a 2m 2m area. The nodes are arbirarily divided in 5 disjoin single-hop flows. Fig. 12 depics he per-flow hroughpu under AMCP, MMAC, Figure 12: Per-flow hroughpu in an arbirary opology of single-hop backlogged flows using 12 channels. As expeced, boh AMCP and MMAC achieve higher aggregae hroughpu han 82.11. Furhermore, 82.11 clearly sarves since 16 ou of 5 flows receive close o zero hroughpu. AMCP achieves higher per-flow hroughpu han MMAC and 82.11. Under AMCP, all flows receive above 15 pk/s and each flow receives higher hroughpu han is prediced lower bound. Under MMAC all flows receive hroughpu above 75 pk/s, ye always lower han AMCP; furhermore, 27 ou of 5 flows receive hroughpu below he corresponding AMCP lower bound. Par of his inefficiency is due o he 2% ATIM window overhead; however, he flows wih much lower hroughpu indicae ha he random channel selecion problem can be a source of inefficiency even if several channels (12 in his case) are available in he sysem. 5.2.2 Muli-hop flows Nex, we move o more sophisicaed scenarios involving mulihop flows. Muli-hop flows induce non-disjoin single-hop flows which include he missing receiver problem and he head-of-line (HOL) problem in addiion o he problems we have experimened so far. We firs consider a scenario ha isolaes and illusraes hese wo addiional problems. Finally, we consider an arbirary scenario where all he problems are presen and also evaluae he effec of mobiliy. Download scenario. In he saic 2-node opology of Fig. 13(a), a designaed gaeway node sends raffic o all oher nodes hrough a ree srucure. In his download scenario muliple channels are of lile help because he boleneck is he radio consrain a he roo node. The maximum per-flow fair rae is 184 / 19 = 9.68 pk/s. The per-flow hroughpus under backlogged condiions are shown in Fig. 13(b). Two key observaions are in place. Firs, AMCP delivers close o maximum per-flow hroughpu in a scenario where he missing receiver problem is srongly presen. Second, MMAC delivers subsanially lower hroughpu han boh AMCP and 82.11. This is no due o he missing receiver problem because MMAC uses synchronized conenion. I is also no due o he random channel selecion problem because he number of channels is no he boleneck in his scenario. The problem arises because each node inends packes o muliple ougoing neighbors. During he 2ms conrol subframe, each node conends for he link corresponding o he HOL packe in is queue. Upon success, for he nex 8ms-daa subframe i will conend and ransmi in he reserved channel only for his link. Hence, he daa subframe can be fully uilized only if a sufficienly high number of packes of his link immediaely follow he HOL packe. Unforunaely his is no likely o happen if his node inends packes o muliple neighbors and is he source of inefficiency in his scenario.
Y (meers) 5 45 4 35 3 25 2 15 1 5 5 1 15 2 25 3 35 4 45 5 X (meers) he roues fixed during each run. We hen consider only he experimens where no roue breakages occurred. In his way, we can es how he MAC proocols reac o mobiliy viewed as changes of he nework conenion regions. Throughpu (pk/s) 4 35 3 25 2 15 AMCP(12 channels) MMAC(12 channels) 82.11 (a) Topology: Gaeway downloads o all oher nodes hrough a ree srucure. Arrows denoe direcion of ransmissions and doed edges denoe inerference. 1 5 1 2 3 4 5 6 7 8 9 Flow ID Throughpu (pk/s) 1 8 6 4 2 AMCP(12 channels) MMAC(12 channels) 82.11 2 4 6 8 1 12 14 16 18 Flow ID (b) Per-flow hroughpus. MMAC use 12 channels. AMCP and Figure 13: Download scenario. There appears o be no easy soluion o he HOL problem. On one hand a node could be allowed o reserve a channel for muliple links during he daa subframe. This would require boh significan changes o he MAC proocol as well as sophisicaed queue managemen ha would increase proocol complexiy. On he oher hand, he daa subframe can be reduced o fi packe ransmissions of a single link. However his increases he overhead due o he conrol subframe. Opimal sizing of he global conrol/daa subframe is hard o perform wihou a-priori knowledge of raffic requiremens. In addiion, no sizing would sui all nodes in he nework. The HOL problem is no specific o MMAC. I exiss in any muli-channel proocol ha aemps o reserve a channel for several packe ransmissions (e.g. SSCH [2]). If no addressed properly, i can produce subsanial overhead ha couner-balances he gain due o muliple channels. On he oher hand, he HOL problem is no presen in AMCP because conenion occurs on a per-packe basis. Muli-hop flows and mobiliy. To sudy mobiliy and he join effecs of he above facors, we consider a mobile scenario of 5 nodes in a 1m 1m area and form 1 muli-hop flows wih arbirary source-desinaion pairs. We use he random waypoin mobiliy model where nodes move a 1 m/s. To es MAC proocol performance we need o operae a relaively high loads. Under such condiions, a dynamic MANET rouing proocol can cause frequen roue changes due o los rouing packes, which in urn can have a dominaing degrading effec in overall performance. To decouple he effec of rouing, we precompue shores pah roues based on he iniial opology and keep Figure 14: Per-flow hroughpu in an arbirary opology of muli-hop flows using 12 channels. Fig. 14 shows he achieved hroughpus under a per-flow UDP load of 3 pk/s. Each daa poin is he average of 1 mobiliy scenarios. AMCP appears robus in erms of delivery raio since each flow achieves hroughpu close o 3 pk/s. In conras, several flows receive much lower hroughpu under MMAC and 82.11. Overall, MMAC ouperforms 82.11. However, in 7 ou of he 1 flows i receives subsanially lower hroughpu han AMCP and flows 6 and 7 receive very low hroughpu similar o 82.11. This inefficiency is due o he superposiion of ATIM window overhead, he random channel selecion problem and he HOL problem. 6. RELATED WORK There is an exensive body of work on disribued MAC proocols for muli-hop wireless neworks each operaing under differen environmens and assumpions. In his secion we review exising MAC proocols in view of he sarvaion problem by broadly classifying hem based on 1) number of channels used (singlechannel/muli-channel) and 2) he access mehod hey use (conenionbased access or scheduled access). 6.1 Single channel proocols Disribued CSMA-based MAC proocols ha provide fair access by enhancing coordinaion of ransmissions over a single channel have been proposed in [1, 11, 15, 17, 22, 23, 26]. Alhough a fair access proocol may address sarvaion i may no always be able o mee a desired minimum rae posed by he sysem. Topologyransparen proocols are an alernaive o CSMA-based access for provision of minimum rae guaranees [5]. Each node uses a differen ransmi-lisen periodic schedule ha has been compued based on projeced number of neighbors (a sysem parameer). By consrucion, he schedule of each node possesses an ineresing mahemaical propery: a any ime, a leas one ransmi slo will be received conflic-free by all neighbors of his node. An operaional problem is ha a mechanism is needed o disribue hese schedules o he nodes. Also he schedules are saic in he sense ha nodes canno modify heir ransmission decisions wihou compromising he minimum rae guaranee. The above proocols operae over a single channel. When muliple channels are suppored by he sysem and need o be exploied for boh fairness and aggregae hroughpu increase, he problem of efficienly coordinaing ransmissions, especially under ransceiver limiaions, becomes much harder.