Centralized Channel Assignment and Routing Algorithms for MultiChannel reless Mesh Networks - PDF document

CentralizedChannelAssignmentandRoutingAlgorithmsforMulti-ChannelWirelessMeshNetworksAshishRaniwalaKartikGopalan
Tzi-ckerChiuehraniwala@cs.sunysb.edukartik@cs.fsu.educhiueh@cs.sunysb.eduDepartmentofComputerScience,StonyBrookUniversity,StonyBrook,NY11794
ComputerScienceDepartment,FloridaStateUniversity,Tallahassee,FL32306TheIEEE802.11WirelessLANstandardsallowmultiplenon-overlappingfrequencychan-nelstobeusedsimultaneouslytoincreasetheaggregatebandwidthavailabletoend-users.Suchbandwidthaggregationcapabilityisroutinelyusedininfrastructuremodeoperation,wherethetrafctoandfromwirelessnodesisdistributedamongmultipleinterfacesofanaccesspointoramongmultipleaccesspointstobalancethetrafcload.However,bandwidthaggregationisrarelyusedinthecontextofmulti-hop802.11-basedLANsthatoperateintheadhocmode.MostpastresearcheffortsthatattempttoexploitmultipleradiochannelsrequiremodicationstotheMACprotocolandthereforedonotworkwithcommodity802.11interfacehardware.Inthispaper,weproposeandevaluateoneoftherstmulti-channelmulti-hopwirelessad-hocnetworkarchitecturesthatcanbebuiltusingstandard802.11hardwarebyequippingeachnodewithmultiplenetworkinterfacecards(NICs)operatingondifferentchannels.Wefocusourattentiononwirelessmeshnetworksthatserveasthebackboneforrelayingend-usertrafcfromwirelessaccesspointstothewirednetwork.Theideaofexploitingmultiplechannelsisparticularlyappealinginwire-lessmeshnetworksbecauseoftheirhighcapacityrequirementstosupportbackbonetrafc.Toreapthefullperformancepotentialofthisarchitecture,wedevelopasetofcentralizedchannelassignment,bandwidthallocation,androutingalgorithmsformulti-channelwire-lessmeshnetworks.Adetailedperformanceevaluationshowsthatwithintelligentchannelandbandwidthassignment,equippingeverywirelessmeshnetworknodewithjust2NICsoperatingondifferentchannelscanincreasethetotalnetworkgoodputbyafactorofupto8comparedwiththeconventionalsingle-channeladhocnetworkarchitecture.I.IntroductionDespitesignicantadvancesinphysicallayertech-nologies,today'sWirelessLANstillcannotofferthesamelevelofsustainedbandwidthastheirwiredbrethren.Theadvertised54MbpsbandwidthforIEEE802.11a/gbasedhardwareisthepeaklink-leveldatarate.Whenalltheoverheads–MACcontention,802.11headers,802.11ACK,packeterrors–areac-countedfor,theactualgoodputavailabletoapplica-tionsisalmosthalved.Inaddition,themaximumlink-layerdataratefallsquicklywithincreasingdis-tancebetweenthetransmitterandreceiver.Theband-widthproblemisfurtheraggravatedformulti-hopadhocnetworksbecauseofinterferencefromadjacenthopsinthesamepathaswellasfromneighboringpaths[1].Figure1showsanexampleofsuchinter-ference.Fortunately,theIEEE802.11b/802.11gstan-dards[2]andIEEE802.11astandard[3]provide3and12non-overlappingfrequencychannels,respectively,thatcouldbeusedsimultaneouslywithinaneigh-borhood.Abilitytoutilizemultiplechannelswithinthesamenetworksubstantiallyincreasestheeffec-tivebandwidthavailabletowirelessnetworknodes.Suchbandwidthaggregationisroutinelyusedwhenan802.11-basedwirelessLANoperatesininfrastruc-turemode,wheretrafctoandfromwirelessnodesisdistributedamongmultipleinterfacesofanaccesspointoramongmultipleaccesspointstoavoidcon-gestion.However,bandwidthaggregationisrarelyappliedto802.11-basedLANsthatoperateintheadhocmode.Inthispaper,weproposeoneoftherstarchitecturesthatusesmultiplefrequencychan-nelsinanadhocnetworkbyequippingnodeswithmultipleNICs,developtheassociatedchannelassign-mentandroutingalgorithms,andpresenttheresultsofacomprehensiveperformancestudyoftheseal-gorithmsusingbothns-2simulationsaswellasrealtestbed.Althoughtherehavebeenseveralresearcheffortsthataimtoexploitmultipleradiochannelsinanadhocnetwork,mostofthemwerebasedonpro-prietaryMACprotocols[4][5][6][7][8][9],andthere-forecannotbedirectlyappliedtowirelessnetworksusingcommodity802.11interfaces.Incontrast,thearchitecturethisworkproposesfocusesspecicallyon802.11-basednetworks,andrequiresonlysystemsoftwaremodication.Asingle-NICarchitectureinherentlylimitsthewholenetworktooperateinonesinglechannel.Thisisbecauseanattempttousemultiplechannelsin50MobileComputingandCommunicationsReview,Volume8,Number2 1253469Path-1Path-28710110Communication-range neighbors of '3'Figure1:Intra-pathandInter-pathinterferenceinasingle-channelmulti-hopadhocnetwork.Nodes1,2,4,5areintheinterferencerangeofnode3,andhencecannottransmit/receivewhennode3isactive.Nodes8,9,and10belongingtoanothernode-disjointpathalsofallintheinterferencerangeofnode3.Thusnoneofthewirelesslinksshowninthegurecansi-multaneouslyoperatewhennode3istransmittingtonode4.
 (b) Dual-NIC ad hoc network321431operating on 2 channels(a) Single-NIC ad hoc networkoperating on 4 channelsNIC tuned to channel 3Wireless link operatingDual-NIC nodeoperating on channel 3tuned to channel 3Single-NIC nodeNetwork Partitionoperating on channel 3Wireless link operating Figure2:(a)Single-NICnetworkgetsdisconnectedwhenoperatinginmultiplechannels,(b)Evenplac-ing2NICsoneachnetworknodeenablesformingaconnected4-channelnetwork.single-NICnetworkdisconnectsthesubsetofnodesusingonechannelfromothernodesthatarenotusingthesamechannel(Fig2(a)).Cross-channelcommunicationrequireseitherchannel-switchingca-pabilitywithineachnode,ormultipleNICspernodeeachtunedtooperateinadifferentchannel.Channel-switchingrequiresne-grainedsynchroniza-tionamongnodesastowhenanynodewilltrans-mit/receiveoveraparticularchannel.Suchne-grainedsynchronizationisdifculttoachievewith-outmodifying802.11MAC.Therefore,inourarchi-tecture,wechoosetoenablecross-channelcommuni-cationbyequippingeachnodewithmultiple802.11commodityNICseachoperatinginadifferentchan-nel(Fig2(b)).Amulti-NIC-per-nodewirelessmeshnetworkarchitectureraisestworesearchquestions:1.Whichofthe3or12radiochannelsshouldbeassignedtoagiven802.11interface?Fortwonodestocommunicatewitheachother,theirin-terfacesneedtobeassignedtoacommonchan-nel.However,asmoreinterfaceswithinanin-terferencerangeareassignedtothesamera-diochannel,theeffectivebandwidthavailabletoeachinterfacedecreases.Therefore,achan-nelassignmentalgorithmneedstobalancebe-tweenthegoalsofmaintainingconnectivityandincreasingaggregatebandwidth.2.Howpacketsshouldberoutedthroughthismulti-interfacewirelessadhocnetwork?Therout-ingstrategyinthenetworkdeterminestheloadoneach802.11interface,andinturnaffectsthebandwidthrequirementandthusthechannelas-signmentofeachinterface.Afullmulti-channelwirelessmeshnetworkarchi-tecturerequirestopologydiscovery,trafcproling,channelassignment,androuting.However,thefocusofthispaperisontheunexploredproblemofchannelassignmentanditsintegrationwithrouting.Topologydiscoveryalgorithmshavebeenexploredin[10]and[11].Similarly,trafcprolingtechniqueshavebeendiscussedin[12]and[13].Avastarrayofroutingalgorithmshavebeenproposedandreviewedinsev-eralarticles[14][15].Weevaluateourchannelassign-mentalgorithmusingtwosuchroutingprotocols-(1)Shortestpathrouting,and(2)Randomizedmulti-pathrouting.Inthispaper,wemakethefollowingresearchcontributions-\nWeproposeamulti-channelwirelessmeshnet-workarchitectureinwhicheachnodeisequippedwithmultipleIEEE802.11interfaces,presenttheresearchissuesinvolvedinthisarchitecture,anddemonstratethroughanextensivesimulationstudythepotentialgaininaggregatebandwidthachievablebythisarchitecture.\nWedevelopandevaluate2novelchannelassign-mentandbandwidthallocationalgorithmsfortheproposedmulti-channelwirelessmeshnet-works.TherstalgorithmNeighborPartition-ingSchemeperformschannelassignmentbasedonlyonnetworktopology.ThesecondalgorithmLoad-AwareChannelAssignmentreapsthefullpotentialofproposedarchitecturebyfurtherex-ploitingtrafcloadinformation.Evenwiththeuseofjust2NICspernode,thetwoalgorithmsimprovethenetworkcross-sectiongoodputbyfactorsofupto3and8respectively.Althougheach2-NICnodecanonlyoperateon2channels,theoverallnetworkcanutilizemanymorechannels(Figure2(b)).Thisbreakseachcollisiondo-mainintomultiplecollisiondomainseachoperatingMobileComputingandCommunicationsReview,Volume8,Number251 onadifferentfrequencychannel.Thisisthefunda-mentalreasonfornon-linearimprovement(8times)inthroughputwithrespecttoincreaseinnumberofNICspernode(from1to2).Inthispaperwefocusourattentiononwirelessmeshnetworkswherethebandwidthissueismostlimiting.Inthesenetworks,staticnodesformamulti-hopbackboneofalargewirelessaccessnetworkthatprovidesconnectivitytoend-users'mobileterminals.Thenetworknodescooperatewitheachothertorelaydatatrafctoitsdestinations.Wirelessmeshnetworksaregainingsignicantmomentumasaninexpensivesolutiontoprovidelast-mileconnectivitytotheInter-net[16][17][18][19][20].Here,someofthenodesareprovidedwithwiredconnectivitytotheInternet,whiletherestofthenodesaccesstheInternetthroughthesewire-connectednodesbyformingamulti-hopwire-lessmeshnetworkwiththem.Asdeploymentandmaintenanceofwiredinfrastructureisamajorcostcomponentinprovidingubiquitoushigh-speedwire-lessInternetaccess[19],useofmeshnetworkonthelast-hopbringsdowntheoverallISPcosts.Forsimi-larreasons,wirelessmeshnetworkcanbeanattractivealternativeeventowiredbroadbandtechnologiessuchasDSL/cablemodem.Wirelessmeshnetworkscanalsoserveasenterprise-scalewirelessbackboneswhereaccesspointsinter-connectwirelesslytoformaconnectiv-itymesh[21][22].Mostoftoday'senterprisewirelessLANdeploymentisonlylimitedtotheaccessnetworkrole,whereacomprehensivewiredbackbonenetworkisstillneededtorelaytheaggregatedtrafcgeneratedfromordestinedtowardsthesewirelessLANaccesspoints.Useofwirelessmeshbackbonenetworkef-fectivelyeliminatesthiswiringoverheadandenablestrulywirelessenterprises.Thehighbandwidthrequirementofwirelessmeshnetworksineachoftheseapplicationdomainssug-geststhatbandwidthaggregationtechniqueshouldbeappliedwheneverpossible.Forbandwidthallocationpurpose,weutilizethefollowingpropertiesofawire-lessmeshnetwork–1.Thenodesinthenetworkarenotmobile.Thenetworktopologycanstillchangebecauseofoc-casionalnodefailures/maintenance,andjoiningofnewnodes.Ourmulti-channelarchitecturecanaccommodatebothofthesepossibilities.2.Thetrafccharacteristics,beingaggregatedfromalargenumberofend-usertrafcows,donotchangeveryfrequently.Thispermitsnetworkoptimizationbasedonmeasuredtrafcprolesoveratimescaleofhoursordays,ratherthansecondsorminutes.Inthispaper,weassumewecanobtainsuchtrafcproleinformationthroughmeasurementsand/orprovisioning,anduseittomodifychannelassignmentandroutingdecisionsonaperiodicbasis.Therestofthepaperisorganizedasfollows.Sec-tionIIreviewspastworkinwirelessadhocnet-worksaswellasotherbranchesofwirelessnetworksthatarerelatedtothisresearch.SectionIIIgivesanoverviewoftheproposedmulti-channelwirelessmeshnetworkarchitecture.SectionIVdescribesthechannelassignment,bandwidthallocation,androut-ingalgorithms.SectionVpresentstheresultsandanalysisofadetailedstudyoftheproposedarchitec-tureandalgorithmsbasedonns-2simulationsaswellasrealtestbed.SectionVIconcludesthepaperwithasummaryofresearchcontributionsandfutureresearchdirection.RelatedWorkSeveralproposals[4][5][6][7][8][9]havebeenmadetomodifytheMAClayertosupportmulti-channeladhocnetworks.Theapproachtakenbymostofthisbodyofresearchistondanoptimalchannelforasinglepackettransmission,essentiallyavoidinginterferenceandenablingmultipleparalleltransmis-sionsinaneighborhood.Unlikeinallthesepreviousproposals,ourarchitecturedoesnotperformchannelswitchingonapacket-by-packetbasis;ourchannelas-signmentlastsforalongerduration,suchashoursordays,andhencedoesnotrequirere-synchronizationofcommunicatingnetworkcardsonadifferentchan-nelforeverypacket.Thispropertymakesitfeasi-bletoimplementourarchitectureusingcommodity802.11hardware.Additionally,oursystemtakesamoreglobalapproachbyadjustingchannelassign-mentsandroutesbasedontheoverallnetworktrafcpatterns.Avastamountofresearchhasbeenconductedinsingle-channelmulti-hoproutinginadhocnetworks.Acomprehensivesurveyoftheseroutingprotocolscanbefoundin[14]and[15].Ourarchitecturedoesnottietoanyspecicroutingmechanism,anditshouldbepossibletouseanydesiredroutingal-gorithmforthegivenscenario.Thechannelalloca-tionalgorithmworkswithgivenroutingalgorithmtoassignnetworkbandwidthtowirelesslinksbasedontheloadimposedbyrouting.Forevaluationpurposes,weuseshortestpathandrandomizedload-balancedmulti-pathrouting.Theideaofusingmulti-pathrout-ingforloadbalancingadhocnetworkshaspreviouslybeendiscussedin[24]and[25].Useofrandomizationtoachieveload-balancedroutinghasbeenproposedinthecontextofwirednetworks[26].Toperformchan-nelandbandwidthassignment,weborrowthecon-ceptofexpected-loadofnetworklinksasameasureoftheircriticalitiestooverallnetworkcommunicationfromLCBR[27].52MobileComputingandCommunicationsReview,Volume8,Number2 SeveralcommercialaswellasresearchprojectsaimtoutilizeWirelessMeshNetworkstoprovidelast-milewirelessconnectivity.MeshNetworksInc'sMeshEn-abledArchitecture[16]andRadiantNetworks'Mesh-Works[18]aretwooftherecentcommercialwirelessmeshnetworks.Bothofthesearchitecturesusepro-prietaryhardwarethatisnotcompliantwith802.11standard.Nokia'sRoofTopWirelessNetwork[17]isanothercommercialmeshnetworkbuiltusingpro-prietary2.4GHzwirelessrouters.Nokia'sRoofTopNetworkusesacommoncontrolchannelandmulti-pledatachannelstoreduceinterferenceamongdif-ferenttransmissions.TransitAccessPointNetwork[19]isaproposedmeshnetworkarchitectureusingnodesequippedwithbeamformingantennas.Theauthorsplantopropose802.11modicationstoim-provebandwidthefciency.Inessence,mostofthesewirelessmeshnetworkprojectsareeitherbasedonsingle-channelorbasedonproprietarymodicationsto802.11protocoltoutilizemultiplechannels.Themulti-NICapproachhasalsobeenmentionedinsomepastwork[29][30];thetrueperformancepo-tentialofthemulti-NICapproachhashowevernotbeendiscoveredearlier.In[29],authorsusemultiple802.11NICspernodeinanadhocnetworksetting.ThisworkassumesanaprioriandidenticalchannelassignmenttotheNICs.Thechannelassignmentofeachnodeisthesame-NIC-1isassignedchannel-1,NIC-2isassignedchannel-2,andsoon.Thisap-proachtousemultipleNICscanonlyyieldafactor2ofimprovementusing2NICs,ascomparedtoafac-tor8improvementpossiblewithourchannelassign-mentscheme.In[30]also,authorsmentionuseofmultipleNICsoneachmeshnode.TheirapproachtoutilizemultipleNICsrequireseachnodetohaveasmanyNICsasithasneighbors.Theyalsorequireasufcientlylargenumberofavailablechannels.Es-sentially,noneoftheseapproachesrealizethetruepotentialofmulti-NICarchitecture.ThekeytothisperformancepotentialliesinthechannelassignmenttechniquethatdecideswhichchanneltouseforwhichNICinthenetworkandinturnhowmuchbandwidthismadeavailabletoeachNICinthenetwork.Noneofthepastresearchhasbroughtoutthisfactorpresentedanysophisticatedchannelassignmenttechniques.Achannelallocationproblemalsooccursincellu-larnetworkswherebecauseoflimitednumber,theavailablechannelsneedtobere-usedfromcell-to-cell,whilemaintainingtheminimumre-usedistance.Thisleadstotheproblemofchannelallocationwhereeachcellneedstobeassignedcertainchannels,basedonitstrafcandchannelsusedinnear-bycells.Vari-ousstaticanddynamictechniqueshavebeenproposedandusedtosolvethisproblem[31].Theasymme-tryofcomponentrolesandcommunicationbehaviorsincellularnetworkmakesitdifferentfromadhocnetworks.Inacellularnetwork,allmobiledevicesAggregation Node11122233335454Coverage Area for Aggregation NodeWired InternetMulti-channel Wireless Mesh Network2Virtual Link4Figure3:SystemarchitectureofMulti-channelWire-lessMeshNetwork.Endusers'mobiledevicescon-necttothenetworkthroughaccesspoint-liketrafcaggregationnodes,whichformamulti-channelwire-lessmeshnetworkamongthemselvestorelaythedatatrafcto/fromenduserdevices.Thelinksbetweennodesdenotedirectcommunicationoverthechannelindicatedbythenumberonthelink.Inthisnetwork,eachnodeisequippedwith2wirelessNICs.There-forethenumberofchannelsanynodeusessimultane-ouslycannotbemorethan2;thenetworkasawholeuses5distinctchannels.communicatewiththeircorrespondingbasestations,whilethebase-stationtobase-stationcommunicationiscarriedoveraseparatenetworkandthecellularchannelallocationdoesnotaddressthatissue.III.ProblemFormulationInthissection,wedescribetheproposedmulti-channelwirelessmeshnetworkarchitecture,andfor-mulatethekeyresearchissuesinvolvedinthearchi-tecture–channelassignmentandrouting.Inpartic-ular,weillustratewhysimplesolutionstothechan-nelassignmentproblemdonotworksatisfactorilyandderivedesirablepropertiesfortheoptimalchannelas-signmentalgorithm.III.A.SystemArchitectureTheproposedmulti-channelwirelessmeshnetworkarchitecture,showninFigure3,consistsofstatictraf-caggregationnodessimilartowirelessLANaccesspoints.Eachtrafcaggregationnodeprovidesnet-workconnectivitytoend-usermobilewirelessdeviceswithinitscoveragearea.Inturn,thesestaticnodesformamulti-hopadhocnetworkamongthemselvesMobileComputingandCommunicationsReview,Volume8,Number253 torelaytrafctoandfromend-userdevices.Notallnodeshavetheaggregationcapability.Somenodesinthemeshnetworkworkaspurerouters[16],whileothernodesserveasgatewaystothewiredInternet.Eachnodeinamulti-channelwirelessmeshnet-workisequippedwithmultiple802.11-compliantNICs,eachofwhichistunedtoaparticularradiochannelforarelativelylongperiodoftime,suchashoursordays.Fordirectcommunication,twonodesneedtobewithincommunication/hearingrangeofeachother,andneedtohaveacommonchannelas-signedtothem.Additionally,apairofnodesusingasamechannelthatiswithinsense/interferencerangeinterfereswitheachother'scommunication,eveniftheycannotdirectlycommunicate.Nodepairsusingdifferentchannelscancommunicatesimultaneouslywithoutinterference.Forexample,inFigure3,eachnodeisequippedwith2NICs.The“virtuallinks”shownbetweenthenodesdepictdirectcommunica-tionbetweenthem;therearenophysicallinksbe-tweenthem.Theradiochannelusedbyavirtuallinkbetweenapairofnodesisshownasthenumberla-beledontheedge.Thisexamplenetworktotallyuses5distinctfrequencychannels.NotethateachmobilenodehasonlyoneNIC,andthecommunicationbe-tweenmobilenodesandaggregationnodesisbasedonthestandardIEEE802.11infrastructuremodeop-eration.Giventheplacementofwirelessmeshnetworknodesandatrafcprolethatdescribesthetrafcloadbetweeneachpairofnodes,themaindesignprob-lemsare(1)howtoassignaradiochanneltoeach802.11interface,and(2)howtoroutetrafcbetweenallpairsofnodes,insuchawaythatthetotalgoodputofthewirelessmeshnetworkismaximized.Wedis-cusseachofthesetwoproblemsinmoredetailinthenexttwosubsections.III.B.TheChannelAssignmentProblemThegoalofchannelassignmentinamulti-channelwirelessmeshnetworkistobindeachnetworkinter-facetoaradiochannelinsuchawaythattheavail-ablebandwidthoneachvirtuallinkisproportionaltoitsexpectedload.Asimpleapproachtothechannelassignmentproblemistoassignthesamesetofchan-nelstotheinterfacesofeachnode,e.g.,channel1totherstNIC,channel2tothesecondNIC,andsoonforeachnodeasdescribedin[29].Thisidenticalchannelassignmentindeedprovidesthroughputgainsbyutilizingmultiplechannels.ThegainsarehoweverlimitedbecauseusingthisapproachanetworkwithNICspernodecanonlyspanatotalofchannels,eventhoughthenumberofavailablenon-overlappingchannelscouldbemuchgreaterthan.ItisalsonotsufcienttosimplyassigneachNICtoadifferentchannel,saythe“least-usedchannel”intheneighbor-hoodasisthecasewithcellularnetworks[31].Thisapproachdoesnotevenguaranteebasicnetworkcon-nectivity.Anodeneedstoshareacommonchannelwitheachofitscommunication-rangeneighborswithwhichitwantstocommunicate.Ontheotherhand,toreduceinterferenceanodeshouldnothavetoomanycommonchannelswithanysingleneighbor.Moregenerally,oneshouldbreakeachcollisiondomainintoasmanychannelsaspossiblewhilemaintainingtherequiredconnectivityamongneighboringnodes.Thechannelassignmentproblemintheproposedmulti-channelwirelessmeshnetworkarchitecturecanbedividedintotwosubproblems-(1)neighbor-to-interfacebinding,and(2)interface-to-channelbinding.Neighbor-to-interfacebindingdeterminesthroughwhichinterfaceanodecommunicateswitheachofitsneighbors.Becausethenumberofinter-facespernodeislimited,eachnodetypicallyusesoneinterfacetocommunicatewithmultipleofitsneigh-bors.Interface-to-channelbindingdetermineswhichradiochannelanetworkinterfaceuses.Themaincon-straintsthatachannelassignmentalgorithmneedstosatisfyare1.Thenumberofdistinctchannelsthatcanbeas-signedtoawirelessmeshnetworknodeislim-itedbythenumberofNICsonit,2.Twonodesinvolvedinavirtuallinkthatisex-pectedtocarrysometrafcshouldbeboundtoacommonchannel,3.Thesumoftheexpectedloadsonthelinksthatinterferewithoneanotherandthatareassignedtothesamechannelcannotexceedthechannel'srawcapacity,and4.Thetotalnumberofradiochannelsisxed.Atarstglance,thisproblemappearstobeagraph-coloringproblem.However,standardgraph-coloringalgorithmscannotreallycapturethespeci-cationandconstraintsofthechannelassignmentprob-lem.Anode-multi-coloringformulation[32]failstocapturethesecondconstraintwherecommunicat-ingnodesneedacommoncolor.Ontheotherhand,anedge-coloringformulationfailstocapturetherstconstraintwherenomorethan(numberofNICspernode)colorscanbeincidenttoanode.Whileacon-strainededge-coloringmightbeabletoroughlymodeltheremainingconstraints,itisincapableofsatisfyingthethirdconstraintoflimitedchannelcapacity.Oneapproachtothechannelassignmentproblemistostartwithonenode,partitionitsneighborsintogroupsandassignonegrouptoeachofitsinterfaces.Eachofthisnode'sneighbors,inturn,partitionsitsneighborsintogroups,whilemaintainingthegroup-ingdonebytherstnodeasaconstraint.Thispro-54MobileComputingandCommunicationsReview,Volume8,Number2 111111222222333344444Node with 2 NICsVirtualChannel-id for LinkFigure4:Resultofneighborpartitioningschemeforagrid-likewirelessmeshnetwork.Thechannelassign-mentisbasedonthenetworktopologyinformation.cessisiterativelyrepeateduntilallnodeshaveparti-tionedtheirneighbors.Eachgroupcanthenbeboundtotheleast-usedchannelintheneighborhood.Ingen-eral,thisschemerequiresawaytopartitionneighborsthatresultsinauniformchannelassignmentacrossthenetwork.Foragridnetwork,thisneighborpartition-ingcanbebasedonpatternssuchasshowninFigure4.Inthisexample,eachnodehas2NICs,buttheresultingnetworkuses4channels.Forageneralnet-work,partitioningofneighborscouldbedoneusingrandomizationtechniques.Whiletheaboveneighborpartitioningschemein-deedallowsanetworktousemorechannelsthanthenumberofinterfacespernode,itdoesnottakeintoaccountthetrafcloadonthevirtuallinksbetweenneighboringnodes.Theschemethuswouldworkwellonlyifeachvirtuallinkinthenetworkhasthesametrafcload.However,thisdoesnotholdtrueinmostcasesassomelinkstypicallycarrymoretrafcthanothers.Conceptually,linksthatneedtosupporthighertrafcloadshouldbegivenmorebandwidththanothers.Thismeansthattheselinksshouldusearadiochannelthatissharedamongafewernumberofnodes.Suchload-awarechannelassignmentwoulddistributeradioresourceamongnodesinawaythatmatchesthespatialdistributionofthetrafcload.III.C.TheRoutingProblemChannelassignmentdependsontheexpectedloadoneachvirtuallink,whichdependsonrouting.Givenasetofcommunicatingnodepairs,theexpectedtrafcbetweenthem,andthevirtuallinkcapacities,therout-ingalgorithmdeterminestheroutethroughthenet-workforeachcommunicatingnodepair.Theresult-ingroutespopulatetheroutingtablesofallthenodesandthusgovernthepathtakenbyfuturetrafc.Apartfromdeterminingthetrafcrouteforeachcommuni-catingnodepair,routingalsoplaysanimportantroleintheload-balancingofthenetwork[24][25].Load-balancinghelpsavoidbottleneckcreationinthenet-work,andinturnincreasesthenetworkresourceuti-lizationefciency.Thenotionofnetwork-wideloadbalancingisconceptuallysimple,butissurprisinglydifculttocapturequantitatively.Finally,routingcanalsoincreasethetoleranceofnetworkagainstnodefailuresbycomingupwithmultiplenode-independentroutesforeachpairofend-hosts[23].Atrun-time,ifanodefailsleadingtoapathfailure,theaffectednodescanhavealternatepathstoroutetheirpackets.III.D.EvaluationMetricTheultimategoaloftrafcengineeringabackbonenetworkistomaximizeitsoverallgoodput,orthenumberofbytesitcantransportbetweennodeswithinaunittime.Thisenablesthenetworktosupportmoreend-userows,andinturnmorenumberofusers.Toformalizethisgoal,weusetheideaofcross-sectiongoodputofthenetwork.Thecross-sectiongoodputofanetworkisdenedas
 \n\r(1)Here,\n\ristheusefulnetworkbandwidthas-signedbetweenapairofingress-egressnodes\n\r.Ifthetrafcprolehasanexpectedtrafcloadof\n\rbetweenthenodepair\n\r,thenonlyupto\n\roftheassignedbandwidthbetweenthenodepair\n\risconsidereduseful.Thiscriteriaensuresthatweonlycounttheusablebandwidthofthenet-worktowardsitscross-sectionthroughput,hencethetermcross-sectiongoodput.Thegoalofthechannelassignmentandroutingalgorithmsistomaximizethiscross-sectiongoodput.IV.Load-AwareChannelAssignment/RoutingIV.A.OverviewWeassumeavirtuallinkexistsbetweenanytwonodesthatarewithincommunicationrangeofeachother.Tomaximizeanetwork'soverallgoodput,theroutingal-gorithmneedstoroutetrafctobalancetheloadonthenetwork'svirtuallinksorsimplylinkstoavoidbottlenecks.However,theproposedwirelessmeshnetworkarchitectureoffersonemoredegreeoffree-dom–modifyingavirtuallink'scapacitybyassigningaradiochanneltothelink.Thisispossiblebecausethecapacityofavirtuallinkdependsonthenumberofotherlinksthatarewithinitsinterferencerangeandthatareusingthesameradiochannel.Becauseroutingdependsonthevirtuallinks'ca-pacity,whichisdeterminedbychannelassignment,andchannelassignmentdependsonthevirtuallinks'expectedload,whichisaffectedbyrouting,thereisthusacirculardependencybetweenradiochannelas-signmentandpacketrouting.Tobreakthiscircular-ity,westartwithaninitialestimationoftheexpectedMobileComputingandCommunicationsReview,Volume8,Number255 loadoneachvirtuallinkwithoutregardtothelinkca-pacity,andtheniterateoverchannelassignmentandroutingstepsuntilthebandwidthallocatedtoeachvir-tuallinkmatchesitsexpectedloadascloselyasitcan.Moreconcretely,givenasetofnodepairsandtheex-pectedtrafcloadbetweeneachnodepair,theroutingalgorithmdevisestheinitialroutesforthenodepairs.Giventheseinitialroutesforthenodepairsandthusthetrafcloadoneachvirtuallink,theradiochannelassignmentalgorithmassignsaradiochanneltoeachinterface,suchthattheamountofbandwidthmadeavailabletoeachvirtuallinkisnolessthanitsex-pectedload.Thenewchannelassignmentisfedbacktotheroutingsteptoarriveatmoreinformedroutingdecisions,i.e.usingactuallinkcapacitiesbasedoncurrentchannelassignment.Attheendofeachitera-tion,ifsomeofthelinkloadsaremorethantheirca-pacities,thealgorithmgoesbacktondabetterchan-nelassignmentusingthelink-loadsfrompreviousit-eration,redotherouting,andcomparesthenewlinkloadswithnewlinkcapacities.Thisiterativeprocesscontinuesonuntilnofurtherimprovementispossible.Figure5depictsthisprocess.Itshouldbenotedthatsomeprobleminputsmightnothavecorrespondingfeasiblesolutions;ourgoalthereforeistoreducethedifferencebetweenlinkcapacitiesandtheirexpectedloadsasmuchaspossible.Insummary,theinputstothecombinedchannelas-signmentandroutingalgorithmare(1)anestimatedtrafcloadforallcommunicatingnodepairs,(2)awirelessmeshnetworktopology,and(3)thenum-berof802.11networkinterfacesavailableoneachnodeandthenumberofnon-overlappingradiochan-nels.Theoutputsofthisalgorithmare(1)thechannelboundtoeach802.11interfaceand(2)thesetofpathsforeverycommunicatingnodepairsinthewirelessmeshnetwork.IV.B.InitialLinkLoadEstimationThecombinedchannelassignmentandroutingalgo-rithm,rstderivesaroughestimateoftheexpectedlinkload.Onepossibilityistoassumethatallinter-feringlinkswithinaneighborhoodequallysplitthecombinedbandwidthofallradiochannels.Speci-cally,weassumethecapacityoflink,
,tobe


(2)whereisthenumberofavailablechannels,isthecapacityperchannel,and
arethenumberofvirtuallinkswithintheinterferencerangeof.Theequationessentiallydividestheaggregatedchannelcapacitiesamongallinterferinglinks,withoutregardtonumberofNICspernode.Basedonthesevirtuallinkcapacities,theroutingalgorithmdeterminestheinitialroutesandthustheinitiallinkloads.Initial Link Load EstimationChannel AssignmentLink Capacity EstimationRoutingFor all links�Capacity = Expected Load ?NoTraffic ProfileExpected Link LoadsYesChannelsSeed Link LoadsChannelsLink CapacitiesExpected Figure5:Basicowchartdiscussingvariousaspectsoftrafcengineeringinmulti-channelmeshnetworkarchitecture.Atthebeginning,aroughestimationoflinkloadsisusedastheseed.Thechannelassignmentalgorithmgovernsthecapacitiesoflinks.Therout-ingalgorithmusesthesecapacitiestocomeupwithroutes,andinturnfeedsmoreaccurateexpectedloadsonthelinkstothenextiteration.Amoreaccurateestimateofexpectedlinkloadisbasedonthenotionoflinkcriticality[27].Tocom-puteinitialexpectedlinkloads,weassumeperfectloadbalancingacrossallacceptablepathsbetweeneachcommunicatingnodepair.Let'scallthenum-berofacceptablepathsbetweenapairofnodes\n\r,\n\r,andthenumberofacceptablepathsbetween\n\rthatpassalink,
\n\r.Thentheexpected-loadonlink,
,iscalculatedusingtheequation


\n\r\n\r\n\r(3)where\n\ristheestimatedloadbetweenthenodepair\n\rinthetrafcprole.Thisequationsaysthattheinitialexpectedloadonalinkisthesumofloadsfromallacceptablepaths,acrossallpossiblenodepairs,thatpassthroughthelink.Becauseoftheas-sumptionofuniformmulti-pathrouting,theloadthatanacceptablepathbetweenanodepairisexpectedtocarryisthenodepair'sexpectedloaddividedbythetotalnumberofacceptablepathsbetweenthem.Whiletheresultingestimatesofthisapproacharenot100%accurate,itprovidesagoodstartingpointtokickofftheiterativerenementprocess.56MobileComputingandCommunicationsReview,Volume8,Number2 1134568DACE1134568BDACE22763B2Edge A-B is assigned channel 6Figure6:Illustrativeexampletoshowthe3rdcaseofchannelassignment.NodeA'schannel-listis[1,6],andthatofnodeBis[2,7].SinceAandBhavenon-intersectingsetsofchannelsinuseandeachnodehas2NICs,linkA-Bneedstobeassignedoneofthechannelsfrom[1,2,6,7].Basedonresultingchannelexpected-loads,linkA-Bisassignedchannel6,andchannel7isrenamedtochannel6.IV.C.ChannelAssignmentGiventheexpectedloadoneachvirtuallink,thegoalofchannelassignmentalgorithmistoassignchannelstonetworkinterfacessuchthattheresultingavailablebandwidthontheseinterfacesisatleastequaltotheirexpectedtrafcload.Thechannelassignmentprob-lemisNP-hard;ahardnessproofcanbefoundintheAppendixA.Inthissubsection,wepresentagreedyload-awarechannelassignmentalgorithm.Inthisal-gorithm,thevirtuallinksinthewirelessmeshnetworkarevisitedinthedecreasingorderoflinkcriticality,ortheexpectedloadonalink.Whenavirtuallinkistraversed,itisassignedachannelbasedonthecur-rentchannelassignmentoftheincidentnodes,callednode1andnode2respectivelyinthefollowing.Thechannellistofanodereferstothesetofchannelsas-signedtoitsvirtuallinks.AssumingthereareNICspernode,thereare3possiblecases-1.Bothnode1andnode2havefewerthanmem-bersintheirchannellist.Inthiscase,weassignanychannelthathastheleastdegreeofinterfer-encetothevirtuallinkinquestion.2.Oneofthenodes,saynode1,hasmembersinitschannellist,andtheothernode'schannellisthasfewerthanmembers.Inthiscase,wechooseoneofthechannelsinnode1'schannellist,assignittothevirtuallink,andaddittonode2'schannellistifitisnotalreadythere.Thechannelchosenfromnode1'schannellististheonethatminimizesthedegreeofinterferenceforthevirtuallink.3.Bothnode1andnode2havemembersintheirchannellists.Iftherearecommonchannelssharedbynode1andnode2,wepickthecommonchannelthatminimizesthedegreeofinterferenceandassignittothevirtuallink.Otherwise,wepickachannelfromnode1andachannelfromnode2,mergethemintoonechannel,andassignthismergedchanneltothevirtuallink.Inthiscase,alltheotherinstancesofthetwochannelsbeingmergedneedtoberenamedintothenewchannelaswell,asshowninFigure6.Again,thechoiceoftwochannelstobemergedissuchthatthecombineddegreeofinterferenceofthetwochannelsisminimized.Bythedegreeofinterference,wemeanthesumofexpectedloadfromthevirtuallinksintheinterferenceregionthatareassignedtothesameradiochannel.Asincreasingthenumberofvirtuallinkswithinanin-terferencerangetendtodecreasethebandwidthshareavailabletoeachoneofthem,decreasingthedegreeofinterferenceofalinkincreasesitsavailableband-width.Byvisitingthevirtuallinkinthedecreasingorderoflinkcriticality,moreloadedlinksarelikelytobeassignedtoachannelwithlessinterference,andthusgivenahighercapacity.LinkCapacityEstimation:Toevaluatetheeffec-tivenessofachannelassignmentalgorithm,weneedtocalculatethecapacityofeachvirtuallink,andcom-pareitagainstthethelink'sexpectedload.Thepor-tionofchannelbandwidthavailabletoavirtuallink,orthelinkcapacity,isdeterminedbythenumberofallvirtuallinksinitsinterferencerangethatarealsoassignedtothesamechannel.Ofcourse,theexactshort-terminstantaneousbandwidthavailabletoeachlinkdependsonsuchcomplexsystemdynamicsascaptureeffect,coherenceperiod,physicalobstacles,strayRFinterferences,anddistance.Ourattempthereistocomeupwithanapproximationofthelong-termbandwidthshareavailabletoavirtuallink.Weap-proximateavirtuallink'scapacity
by

 \n\r\r\r(4)whereistheexpectedloadonlink,\nisthesetofallvirtuallinksintheinterferencezoneoflink,andisthesustainedradiochannelcapacity.Therationaleofthisformulaisthatwhenachannelisnotoverloaded,thechannelshareavailabletoavirtuallinkisproportionaltoitsexpectedload.Thehighertheexpectedloadonalink,themorechannelshareitwouldget.Theaccuracyofthisformuladecreasesas\napproaches.MobileComputingandCommunicationsReview,Volume8,Number257 IV.D.RoutingAlgorithmTheload-awarechannelassignmentalgorithmisnottiedtoanyspecicroutingalgorithm.Itcanworkwithdifferentroutingalgorithms.Forevaluationpur-poses,weexploretwodifferentroutingalgorithms–(1)shortestpathrouting,and(2)randomizedmulti-pathrouting.TheshortestpathroutingisbasedonstandardBellman-Fordalgorithmwithminimumhop-countmetric.Theshortestpathherereferstotheshortest“feasible”path,i.e.,apathwithsuf-cientavailablebandwidthandleasthop-count.Themulti-pathroutingalgorithmattemptstoachieveload-balancingbydistributingthetrafcbetweenapairofnodesamongmultipleavailablepathsatruntime.Theexactsetofpathsbetweenacommunicatingnodepairischosenrandomlyoutofthesetofavailablepathswithsufcientbandwidth.Althoughinthiscase,thetrafcbetweenanodepairissplitacrossmultiplepaths,packetsassociatedwithanetworkconnectionstillfollowasinglepathtoavoidTCPre-ordering.IV.E.PuttingItAllTogetherFigure7depictstheiterativeprocessofthecombinedchannelassignmentandroutingalgorithm.Attheverybeginning,weestimatetheinitiallinkloadsus-ingtheschemedescribedinsectionIV.B.Next,weiteratemultipletimesthroughthechannelassignmentandroutingsteps.Wecalltheseiterationstheexplo-rationphase.Eachtimeweseeachannel/routecon-gurationthatprovidesabetternetworkcross-sectiongoodput,weentertheconvergencephase.Thecon-vergencephaseissimilartotheexplorationphaseex-ceptthattheroutingalgorithmnowonlyre-routesthenon-conformingows,i.e.onesthathavenotfoundapathwithsufcientbandwidthtomeettheirtrafcdemands.Theconvergencephaseisthenrepeatedun-tilthecross-sectiongoodputoftheresultingnetworkconverges.Forboth,theexplorationphaseandtheconver-gencephase,theroutingorderfordifferentowsisxedatthebeginningofalgorithmbasedonthehop-countdistancebetweenthetwoend-points–oneswithshorterhopcountdistanceareroutedrst.Thepartic-ularorderischosensoastorouteowsthatconsumelessernetworkresources,rst.Axedroutingorderalmostalwaysensuresconvergence.Finally,weiter-ateoverboththeexplorationandconvergencephasesuntileitherallnodepairsaresuccessfullyrouted,ornobetternetworkconguration(channelassignmentsandroutes)isseeninseveraliterations.V.PerformanceEvaluationTostudytheoverallperformanceoftheproposedmulti-channelwirelessmeshnetworkarchitectureandLink CapacitiesLink CapacitiesTraffic ProfileChannel AssignmentFull RoutingChannel AssignmentNoNoConvergence PhaseExploration PhaseYesExpected Link LoadsLink LoadsExpected Cross-section b/w?ImprovedSave Best ConfigurationFor all linksOR converged ?YesExpected Link LoadsExpected Link LoadsConverged ? Re-Route non-conforming flowsNo�capacity load ?Figure7:Overalliterations.Intheexplorationphase,fullroutingisperformedineverysteptoal-lowalgorithmtoexplorenewcongurations.Intheconvergencephase,onlynon-conformingowsarereroutedtone-tunespecicchannel/routecongu-rationscomingoutofexplorationphase.theeffectivenessoftheassociatedchannelassignmentandroutingalgorithms,weperformedanextensivesimulationstudyusingNS-2.WemodiedNS-2tosupportmultiplewirelesscardsonmobilenodesandrandomizedmulti-pathrouting.Inaddition,togaugetheinter-channelinterferencethatisnotmodeledbyNS-2,webuiltasmallmulti-channeladhocnetworkusing802.11bhardwareandevaluatedthefeasibilityofbuildingamulti-interfacePC-basedwirelessmeshnetworknode.V.A.SimulationResultsInthissubsection,wepresentthesimulationresultsdemonstratingtheperformanceimprovementsofde-ployingmultipleinterfacesoneachwirelessmeshnet-worknode,andthecontributionofchannelassign-mentschemes.Wealsodiscusstheimpactofvarioustunablesystemparameters,andtheeffectofnetworktopology/trafcpatterns.V.A.1.ImprovementsduetoMultipleNICsandLoad-awareChannelAs-signmentFigure8presentsthecross-sectiongoodputofa100-nodesquare-gridnetworkforvarioustrafcproles58MobileComputingandCommunicationsReview,Volume8,Number2 12345678910Traffic Profile Number048121620242832Network Cross-section Goodput (Mbps)Single-channel NetworkMulti-channel, Identical CAMulti-channel, Neighbor Partitioning CAMulti-channel, Load-aware CAFigure8:Thenetworkcross-sectionalgoodputfor20randomlychosenpairsofingress-egressnodeswithshortest-pathrouting.Theguresshowthatevenwiththesimpleneighborpartitioningapproach,thereissubstantialimprovementinnetworkcross-sectionalgoodputbyuseofjust2NICspernode.UsingtheLoad-awarechannelassignmentscheme,how-ever,yieldsthefullpotentialofmulti-channelwirelessmeshnetworks.Thechannelassignmentfromneigh-borpartitioningalgorithmistheonecorrespondingtoFigure4.eachcontaining20pairsofrandomlychoseningress-egressnodes.Recallthatthecross-sectiongoodputisdenedasthesumofusefulbandwidthassignedbetweenallcommunicatingingress-egressnodepairs.Foreachprole,theamountoftrafcbetweeneachingress-egressnodepairwaschosenatrandombe-tween0and3Mbps.Theratiobetweeninterferenceandcommunicationrangewasxedat2.Dependingonitsposition,eachnodecouldcommunicatewithupto4neighbors.AllexperimentswereconductedwithRTS/CTSmechanismenabled.Unlessspecied,theroutingalgorithmusedistheshortest-pathrouting,andinitialloadwascomputedusingequation3.Toderivethenetworktosaturation,thebandwidthofalltheowsisproportionallyvarieduntilthenet-workcanonlyroute75%oftheaggregateinputtraf-c.Therelativeperformanceofdifferentalgorithmsdoesnotchangeforothervaluesofsaturationthresh-old,e.g.100%atwhichweensurethateachowhastobeassigneditsfullrequiredbandwidth.Thesatura-tionthresholdcanalsobeper-owtoensurefairnessacrossows,e.g.onecanensurethateachowhastobeassignedatleastacertainpercentageofitstraf-crequirement.Weveriedthecross-sectiongoodputassignedbythevariousalgorithmsusingns-2simula-tions,whereweemulatedthetrafcprolebyrunningCBRUDP-owsbetweeningress-egressnodepairs.Thereceivedtrafcwasmeasuredontheeachoftheegressnodesandaddedtogethertoyieldthecross-sectiongoodput.Forbrevity,weonlyshowtheoverallcross-sectiongoodputforallthegraphs.12345678910Traffic Profile Number0481216202428323640Network Cross-section Goodput (Mbps)Single-channel NetworkMulti-channel, Identical CAMulti-channel, Neighbor Paritioning CAMulti-channel, Load-aware CAFigure9:Networkcross-sectionalgoodputwithran-domizedload-balancedrouting.Thisguredemon-stratestheadaptabilityofchannelassignmenttolinkloadsimposedbydifferentroutingschemes.NotethedifferentY-scalefrompreviousFigure.ThegraphsinFigure8showthecross-sectiongoodputmadeavailableforsingle-channelnetworkandfor12-channel/2-nic-per-nodenetworkwithdif-ferentchannelassignmentschemes.Comparedwithconventionalsingle-channelwirelessmeshnetworkarchitecture,theidenticalchannelassignmentschemeIII.Bachievesapproximatelytimesimprovementincross-sectiongoodput.Incontrast,theneighborparti-tioningscheme(asshowninFigure4)achievesbe-tween
and
timesimprovementoversingle-channelarchitecture.Theload-awarechannelassign-mentschemebringsoutthefullpotentialofthepro-posedmulti-channelwirelessmeshnetworkarchitec-ture,byachievingovertimesimprovementincross-sectiongoodputwithjust2NICspernode.Intuitively,equippingeachwirelessmeshnetworknodewithmul-tipleinterfacesallowsthenetworktouseseveralra-diochannelssimultaneously.Thisbreakseachcolli-siondomainintoseveralcollisiondomainsoperatinginadifferentfrequencyrange.Acollisiondomainisfurthersub-dividedspatiallywhentheingress-egressnodepairsoriginallypassingthroughthecollisiondo-main,takedifferentpathstoroutethetrafc.Thisdivisionofeachcollisiondomainacrossmultiplefre-quencyandspatialdomainsisthekeyreasonforthenonlineargoodputimprovement(8times)withrespecttotheincreaseinthenumberofNICs(from1to2).Moreover,theinterferenceamongadjacenthopsofanindividualpathoramongneighboringpathsismuchreduced.Figure9showsthesameperformancecompari-sonwhentheroutingalgorithmischangedtoran-domizedmulti-pathrouting.Becausewedonotper-formanyexplicitloadbalancinginmulti-pathrout-ingscheme,theperformanceimprovementwhengo-ingfromsingle-pathroutingtomultiple-pathroutingisnotveryconsistent.Thisistrueforboththesingle-MobileComputingandCommunicationsReview,Volume8,Number259 012345678910Link Load / Assigned Bandwidth01020304050607080Number of LinksLoad-aware Channel AssignmentNeighbor Partitioning Channel AssignmentFigure10:Ratioofloadimposedbyroutingalgorithmandbandwidthassignedbythechannelassignmental-gorithmforalllinksinthenetwork.Aratiocloseto1forallthelinksinload-awarechannelassignmentcase,impliesassignedbandwidthcloselymatchestheimposedload.channelcaseandmultiple-channelcase.However,thegoodputgainofthemulti-channelnetworkarchitec-turewithproperchannelassignmentalgorithmsovertheconventionalsingle-channelarchitecturedoesnotseemtodependonaparticularroutingalgorithm.Thisadaptabilityofthechannelassignmentalgorithmen-ablesonetochoosearoutingschemeappropriatetothedeploymentscenario.Theimprovementachievedwithuseofrandomization-basedmulti-pathroutingisbecauseofbetterload-balancingofthenetwork.Withtheuseofamoreexplicitload-balancedrouting,thenetworkperformanceshouldimproveevenfurther.Figure10demonstratestheeffectivenessofthechannelassignmentdonebyload-awarechannelas-signmentscheme.Foreachlinkinthenetwork,theratioofloadimposedbytheroutingalgorithmandthebandwidthassignedbythechannelassignmental-gorithmwasmeasured.Aratiocloseto1indicatesthatmorebandwidthisallocatedtolinksthatrequiremorebandwidth.Weobservethatalthoughthelinkloadimposedbyroutingvariedanywherefrom0to3.9Mbpsacrossnetworklinks,theratioiscloseto(orlessthan)1fortheload-awarechannelassign-mentscheme.Achievingthisdistributionofchannelresourceamongthenodestomatchthespatialdistri-butionoftrafcloadisthekeytogoodperformanceofthescheme.Fortheneighborpartitioningscheme,mostofthelinksareoverloadedresultinginthevari-ationofratiofrom0.5to8.9,thereasonisthatthelatterperformsaload-insensitiveassignmentofchan-nels.ThehistogramforIdenticalchannelassignmentscheme(notshown)issimilarinnaturetotheNeigh-borpartitioningapproach.12345Number of Network Cards / Node081624324048566472Network Cross-section Goodput (Mbps)3 Channels6 Channels9 Channels12 Channels24 ChannelsFigure11:Impactofincreasingthenumberofradiochannelsand/orcardspernode.Asmorechannelsaremadeavailable,thechannelassignmentalgorithmusesthemtoincreasetheoverallnetworkthroughput.Experimentswithdifferenttrafcprolesshowsimi-largraphs.V.A.2.EffectsofAvailableResources(In-terfaces/Node&Channels)Figure11showstheimpactofincreasingthenum-berofNICsoneachwirelessnodeand/orthenum-berofnon-overlappingradiochannelsavailablefromthephysicalwirelessnetworktechnology.Theex-perimentalsetupforthesesimulationsisthesame(10x10grid-networkwith20pairsofrandomlycho-seningress-egressnodes).Thenumberofchan-nels,3and12,correspondtothenumberofnon-overlappingchannelsavailableinIEEE802.11band802.11arespectively.The6and9channelscorre-spondtothecaseswhensomeofthewirelesschan-nelsmightbealreadyinusebytheaccessnetworkorsomeothernetworks.Theexperimentsdemonstratethattheload-awarechannelassignmentalgorithmcaneffectivelyadaptitselfwiththenumberofavailablechannels/NICs.Asnewchannelsbecomeavailable,thealgorithmcanincreasethereusedistanceandthusincreasethecross-sectiongoodput.Thegraphssug-gestthatincreasingtheNICsoneachnodedonothelpasmuchasincreasingthechannelsinthenetwork.Thereasonisthatevenwith2NICsthenetworkisabletospanaround9channels,thusthechannellim-itationcomesrst.AsFCCmakesmorechannels(12)availableforuseby802.11a,increasingthenum-berofNICspernodebeyond3willindeedimprovetheperformancefurtherasshownbythehypotheticalgraphdrawnfor24availablechannels.V.A.3.EffectofInputNetwork&TrafcIngure12,wevariedthenumberofingress-egresspairsinthe10x10network(eachnodeequippedwith2NICs)whilekeepingtheaggregatedofferedloadtobethesame.Asmoreingress-egresspairsarein-60MobileComputingandCommunicationsReview,Volume8,Number2 01020304050Number of Ingress-Egress Node Pairs048121620242832Network Cross-section Goodput (Mbps)Single-channel NetworkMulti-channel, Identical CAMulti-channel, Neighbor Partitioning CAMulti-channel, Load-aware CAFigure12:Impactofvaryingthenumberofingress-egresspairsongoodputimprovements.12345678910Traffic Profile Number0481216202428Network Cross-section Goodput (Mbps)Single-channel NetworkMulti-channel, Identical CAMulti-channel, Load-aware CAFigure13:Comparisonofmulti-channelnetworkagainstsingle-channelnetworkforMITRoofnettopology[33].troduced,thetrafcrequirementismoredistributedacrossthenetworkleadingtoanoverallincreaseinnetworkutilization.Theload-awareschemeadaptsthechannelassignmenttothesedifferentsetsoftrafcrequirementsmaintainingtheperformanceimprove-mentsoversingle-channelnetwork.Experimentswithdifferenttrafcprolesproducedsimilarresults.Wealsoexperimentedwithdifferentnetworktopologies.Figure13showstheperformancecompar-isonofthe29-nodeMITRoofnetnetwork[33]simu-latedinns-2.Thedataforgraphconnectivityisbasedonsignal-strengthnumbersfromthetestbed.Eachpointinthegraphcorrespondstoarandomlygener-atedtrafcprolesof10ingress-egressnodepairs.The8+timesimprovementinnetworkperformancedemonstratestheusefulnessofmulti-channelarchi-tectureforrealnetworks.Weobservedsimilarim-provementsforothertopologies-hexagonalgrid,andincompletemesh.Theperformanceimprovementus-ingneighborpartitioningscheme,however,dependsonthetopology.Amoregenericwaytopartitionthe012345678910Number of Worst-case Node Failures048121620Network Cross-section Goodput (Mbps)Figure14:Impactofworst-casenodefailuresonfu-turechannelassignments.Theload-awarechannelas-signmentshiftsthechannels/bandwidthtorestofthenetworktotackletheserepeatednodefailures,grace-fullyreducingthenetworkcross-sectiongoodput.neighborsisneededinthelatterschemetohandlegen-eralmeshnetworks.Figure14demonstratestheadaptabilityofloadawarechannelassignmenttoworst-casenodefailures.Eachtimeafterperformingthechannelassignmentandshortest-pathrouting,thenodeinthenetworkwiththemaximumload(andwhichwasnotaningressoregressnode)wassimulatedtofail.Thechannelassignmentprocesswasrepeated,andthenewcross-sectiongoodputmeasured.Again,thenodewiththemaximumloadwassimulatedtofail,channelassign-ment/routingredone,andcross-sectiongoodputmea-sured.Theprocesswasrepeatedforupto10%nodefailures.Thegracefuldegradationinnetworkband-widthindicatestheadaptationofchannelassignmenttonodefailures.Ingeneral,nodefailuresareproba-blymorerandom,andthereforebandwidthdegrada-tionshouldbeevenmoregraceful.Inapracticalset-ting,onecanusearoutingschemethatassignsbackuppathsforcommunicatingnodepairsupfront,thusthechannelassignmentandroutingdonotneedtobedoneimmediatelyafteranodefailure.Thetremen-dousimprovementinnetworkbandwidthwithmulti-channelarchitecturemakesitpossibletoallocatesuchbackuppathswhilemaintaininghighthroughputoverprimarypaths.V.B.ImplementationExperiencesInthissubsection,wediscussourimplementationexperienceswithreal802.11bhardware.Speci-cally,wepresentempiricalmeasurementsofinter-channelinterferencefortwocardsresidingonasin-glenode,techniquestoovercomesuchinterference,andnallythethroughputimprovementsfora4-nodemulti-channelmeshnetworkbuiltusing802.11bin-terfacehardware.MobileComputingandCommunicationsReview,Volume8,Number261 Table1:Interferencebetweentwointernal-antennaequipped802.11bcardsplacedonthesamemachineandoperatingonchannels1and11.Thelastcolumnindicatesthetotalgoodputachievedasa%ofsumofindividualgoodputswithoutinterference.Thelink-layerdatarateforalltheseexperimentswasclampedto11Mbps.NIC-1NIC-2NIC-1NIC-2%ofMaxActionActionGoodputGoodputGoodputsendsilent5.52--recvsilent5.23--silentsend-5.46-silentrecv-5.37-sendsend2.442.7747.6%recvsend2.214.0258.3%sendrecv4.222.4261.0%recvrecv4.021.8955.8%V.B.1.Inter-channelInterferenceNS-2simulatormakestheassumptionthatthereisnointerferencebetweennon-overlappingchannels.Thisassumption,however,isnotentirelytrueinpractice.Inourexperimentswithreal802.11bhardware,weobservedsubstantialinterferencebetweentwocardsplacedonthesamemachinedespiteoperatingonnon-overlappingchannels.Theextentofinterferencede-pendsontherelativepositionsofthecards.Placingcardsrightontopofeachotherleadtomaximumin-terference,andachievesonlyamaximum20%gaininaggregategoodputoverthesinglechannelcase(shownintable1).Ifthecardsareplacedhorizon-tallynexttoeachother,asinOrinocoAP-1000ac-cesspoints,theinterferenceisminimumleadingtoalmost100%gaininaggregategoodput.Inaddition,thedegradationduetointer-channelinterferencewasfoundindependentoftheguardband,i.e.thedegrada-tionwasalmostthesamewhenchannel1and6wereusedascomparedtothecasewhenchannel1and11wereused.Wesuspectthisinterferencearisesbecauseoftheimperfectfrequency-lterpresentinthecom-moditycards.Thisresulthasanimplicationovertheplacementofmultiplecardsonthesamemachine.Theelectro-magneticleakagefromthecardsneedstobetakenintoaccount,andonecardshouldnotbeplacedinthezonewherethestrengthoftheleakageradiationsbytheothercardishigh.OnepossiblewaytoachievethisistouseUSBcardsinsteadofPCI/PCMCIAcardsandplacethemside-by-sideinsimilarcongurationasinOrinocoAP-1000accesspoints.Anotherpossibilityistoequipcardswithexternalantennasandplacetheexternalantennasslightlyawayfromeachother.UsingexternalantennasalonemayTable2:Reducedinterferencewiththeuseofexter-nalantennas.Here,thecardswereoperatedoncloserchannels–1and6.NIC-1NIC-2NIC-1NIC-2%ofMaxActionActionGoodputGoodputGoodputsendsilent5.93--recvsilent5.75--silentsend-5.96-silentrecv-5.78-sendsend5.525.9696.6%recvsend5.375.8996.2%sendrecv5.425.4192.5%recvrecv5.665.1793.9%Node-1Channel 1Channel 11Channel 6Channel 1Flow-1Flow-4Flow-2Node-2Node-4Flow-3Figure15:Multi-channel802.11bTestbed.Eachnodeisequippedwith2cardswhosechannelsweredeter-minedbasedontheload-awarechannelassignmental-gorithm.sufce:itisalsonecessarythattheinternalan-tennaofthecardisdisabled.WeusedOrinocoGoldPCIadaptersthatcomewithexternalantennasthaten-abledustobuildmulti-channelwirelessmeshnetworkusingstandardPCs.Table2showstheresults.Theexactinterferencedependsontheplacementandcardactions(send/receive).Theuseofexternalantennasisabletohandlemostoftheinterferenceeffectsasshownbytable2;theremaininginterferenceisbe-causeofRFleakagefromcablesandfromcard'sin-ternalcomponents.YetanotheroptionistousetheupcomingEngimchipsets[34]whichsolvetheinterferenceproblematRF-levelitself.Engimchipsetsreceivethecom-pletespectrum,digitizeitandprocessittocompensateforinter-channelinterference.ThiswidebandspectralprocessingcapabilitycanhelpbuildsingleNICwithmulti-channelcommunicationcapabilitywhileintro-ducingminimalinter-channelinterference.62MobileComputingandCommunicationsReview,Volume8,Number2 Table3:Performanceofmulti-channel802.11btestbed.Theperformanceimprovementinthiscaseislimitedbythenumberofnon-overlappingavailablechannelsfor802.11bstandard.FlowIdSingle-channelMulti-channel802.11b802.11b10.922.4020.701.6130.872.4040.852.39Total3.348.80V.B.2.3-channel802.11bNetworkFigure15showsthe4-nodemulti-channeltestbedbuiltwithinourlab.Node-1andnode-2werebasedondesktopseachequippedwithoneOrinoco802.11bPCIcardandanotherCiscoAironet350PCMCIAcardaddedusingaPCI-PCMCIAconvertor.Node-3andnode-4werebasedonlaptopseachequippedwithoneCiscoAironet350PCMCIAcardandan-otherSyntax802.11bUSBcard.Thenodeswerear-rangedinagridtopologyasshowninFigure15,and4differentows(eachgoingover2-hops)weregen-erated.Theassignmentofchannelsandtheroutesfortheowsweredeterminedusingtheload-awarechan-nelassignmentalgorithm,andareshowninthegure.Theexperimentswerethenrepeatedwithusingonlyonecardoneachnodetunedtothesamechannel.Ta-ble3showsthebandwidthachievedbyeachowinthetwocases.Themulti-channelnetworkachieves2.63timesthethroughputascomparedtothesingle-channelnetwork.Thenumberofnon-overlappingchannelsin802.11bstandard,i.e.3,isthelimitingfactorforthisperformance.Theperformancehow-everdoesnotreach3-timesthesingle-channelnet-workperformancebecauseoftheinter-channelinter-ferencethatcouldnotbecompletelyeliminated.VI.ConclusionDespitemanyadvancesinwirelessphysical-layertechnologies,limitedbandwidthremainsapressingissueforwirelessLANs.Thebandwidthissueismostsevereformulti-hopwirelessmeshnetworksduetointerferenceamongsuccessivehopsofanindividualpathaswellasamongneighboringpaths.Asare-sult,conventionalsingle-channelwirelessmeshnet-workscannotadequatelyfullltheroleofanextendedlast-mileaccessnetwork,letaloneawirelesscam-pusbackbonethatcompletelyreplaceswiredEther-net.Inthispaper,weproposeamulti-channelwirelessmeshnetworkarchitecturebasedon802.11hardwarethateffectivelyaddressesthisbandwidthproblem,andshowhowwithproperchannelassignmentandroutingalgorithmssuchanetworkarchitecturecanbecomeaseriouscontenderforacampus-scalebackbonenet-work.Althoughthemulti-NIC-per-nodeapproachhasbeeninvestigatedinthepast,ithasnotbeenfullyexplored.Weshowthatchannelassignmentplaysacrucialroleinrealizingthefullpotentialofthepro-posedmulti-channelwirelessmeshnetworkarchitec-ture,anddiscussvariousissuesinvolvedinchannelassignment.Wepresenttwonovelchannelallocationalgorithms,andevaluatetheirperformancefortwodifferentroutingalgorithms:shortestpathroutingandrandomizedmulti-pathrouting.Oursimulationstudyshowsthatbydeployingjust2NICspernode,itispossibletoachieveafactorofupto8improvementintheoverallnetworkgoodputwhencomparedwiththeconventionalsingle-NIC-per-nodewirelessadhocnetwork,whichisinherentlylimitedtoonesinglera-diochannel.Finally,weempiricallyshowedthatitispossibletobuildaPC-basedmulti-NICwirelessmeshnetworknodewiththeuseofexternalantennas.Theperformanceevaluationpresentedinthispa-perdemonstratesthatthemulti-channelwirelessmeshnetworkarchitectureisquitepromising,anddeservesfurtherinvestigation.Thereareseveralinterestingquestionsthatweareexploringcurrently,e.g.(1)Whatisthedistributedversionoftheproposedload-awarechannelassignmentalgorithmthatonlyutilizeslocalneighborhoodtrafcinformationtoperformon-the-ychannelassignment?(2)Howshouldonear-chitectamobilemulti-channeladhocnetwork?APPENDIXChannelAssignmentisNP-hard.Giventheexpectedloadoneachvirtuallink,thegoalofchannelassignmentalgorithmistoassignachanneltoeachnetworkinterface,suchthattheresult-ingavailablebandwidth
oneachvirtuallinkisatleastequaltoitsexpectedload\r.Therearephysicalchannelseachwithacapacityofeachinanygiveninterferencezone.Finally,eachnodeisequippedwith
wirelessnetworkinterfaces.WeprovetheNP-hardnessofthechannelassign-mentproblembyreducingtheMultipleSubsetSumProblem[28]tothechannelassignmentproblem.Themultiplesubsetsumproblemcanbestatedasfol-lows.Wearegivenasetofitemswithweights,,..,..,andidenticalbinsofcapacityeach.Theobjectiveistopacktheseitemsinthebinssuchthatthetotalweightofitemsinthebinsismaximized.Aninstanceofmultiplesubsetsumproblemiscon-vertedintoanequivalentinstanceofchannelassign-mentproblemasfollows.Weconstructasinglecolli-siondomainnetworkofnodeswhereeachnodeisequippedwith2networkcards.WenowaddavirtualMobileComputingandCommunicationsReview,Volume8,Number263 1352n-32n-12n-22nWWWW123n-1nWCCCCCCFigure16:Constructednetworkgraphforaninstanceofmultiplesubsetsumproblem.linkbetweennodes1and2withbandwidthrequire-mentofwhichistheweightoftherstiteminthegivenmultiplesubsetproblem.Wenextaddanothervirtuallinkbetweennodes3and4withbandwidthre-quirementof,andsoon.Next,weintroducevir-tuallinksbetweennodes2and3,nodes4and5,andsooneachwithbandwidthrequirementof.Wealsoaddalinkbetweennodes\rbackto1.Theconstruc-tionisshowninFigure16.Thecapacityofthechan-nelisthesameasthebincapacity,andthenumberofchannelsisequalto
.Letusnowseewhatasolutiontothisconstructedproblemlookslike.Firstofall,eachofthebluelinkswithbandwidthrequirementhastobeassignedoveradedicatedchannel.Thus,thesolutionmustusetheremainingchannelstosatisfyalltheblacklinks.Eachbluelinkalsousestwonetworkinterfaces–oneoneachofthetwonodesitisincidentupon.Thus,thesolutionmustusetheremainingsingleinterfaceoneachnodetosatisfytheblacklinksbandwidthre-quirements.Now,alltheblacklinks,say,,.., \nthatthesolutionputsoveranyoneofthechannelsmusthaveasumlessthan.Thismeans,thatfortheorig-inalmultiplesubsetproblem,allof,,...,\rcangointoonebin.Similarly,alltheitemscorrespondingtovirtuallinksscheduledoveranyotherchannelcangotothecorrespondingbin.Thus,ifthechannelas-signmentproblemweresolvableinpolynomialtime,sowouldbethemultiplesubsetsumproblem.SincethemultiplesubsetsumproblemisNP-hard,thechan-nelassignmentproblemisalsoNP-hard.ReferencesK.Jain,J.Padhye,V.N.Padmanabhan,L.Qiu;“Impactofinterferenceonmulti-hopwirelessnetworkperformance”;MobiCom,2003.[2]“IEEE802.11bStandard”;standards.ieee.org/getieee802/download/802.11b-1999.pdf[3]“IEEE802.11aStandard”;standards.ieee.org/getieee802/download/802.11a-1999.pdf[4]A.MuirandJ.J.Garcia-Luna-Aceves;“AChan-nelAccessProtocolforMultihopWirelessNet-workswithMultipleChannels”;IEEEICC'98.[5]J.SoandN.Vaidya;“Multi-ChannelMACforAdHocNetworks:HandlingMulti-ChannelHiddenTerminalsUsingaSingleTransceiver”;ACMMobiHoc,May2004.[6]Wing-ChungHung,K.L.EddieLaw,A.Leon-Garcia;“ADynamicMulti-ChannelMACforAd-HocLAN”;21stSymp.onComm.'02[7]AsisNasipuriandSamirR.Das;“AMultichan-nelCSMAMACProtocolforMobileMultihopNetworks”;Proc.ofIEEEWCNC1999.[8]R.Garces,J.G.L.Aceves;“CollisionAvoid-anceandResolutionMultipleAccessforMulti-channelWirelessNetworks,”;Infocom2000.[9]J.So,N.H.Vaidya.“AMulti-ChannelMACProtocolforAdHocWirelessNetworks”;UIUCTechReport,Jan2003.[10]Ying-YiHuang,andMartL.Molle;“Anim-provedtopologydiscoveryalgorithmfornet-workswithwormholeroutinganddirectedlink”;IEEEComputerNetworks,1999.[11]R.Chandra,C.Fetzer,K.Hogstedt;“AdaptiveTopologyDiscoveryinHybridWirelessNet-works”;Informatics'02[12]K.C.Claffy,H.W.Braun,G.C.Polyzos;“AparameterizablemethodologyforInternettrafcowproling”;IEEEJSAC,1995.[13]Y.J.Lin,M.C.Chan;“AScalablemonitoringapproachbasedonaggregationandrenement”;IEEEJSAC,2002.[14]A.Iwata,C.-C.Chiang,G.Pei,M.Gerla,andT.-W.Chen.“ScalableRoutingStrategiesforAd-hocWirelessNetworks.”;IEEEJSAC,1999[15]E.M.RoyerandC-KToh.;“AReviewofCur-rentRoutingProtocolsforAdHocMobileWire-lessNetworks”;IEEEPersonalCommunica-tions,April1999.[16]MeshNetworksInc;www.meshnetworks.com[17]NokiaInc.;“NokiaRoofTopWirelessRouting”;www.americasnetwork.com/americasnetwork/data/articlebrief/americasnetwork/412002/34898/article.pdf[18]RadiantNetworks;www.radiantnetworks.com[19]R.Karrer,A.Sabharwal,E.Knightly;“EnablingLarge-scaleWirelessBroadband:ACaseforTAPs”;HotNets'03[20]P.Bhagwaty,B.Ramanz,D.Sanghi;“Turning802.11Inside-Out”;HotNets'03.64MobileComputingandCommunicationsReview,Volume8,Number2 [21]FireTideInc.;www.retide.com[22]StrixNetworksInc.;www.strixsystems.com[23]S.LeeandM.Gerla;“SplitMultipathRoutingwithMaximallyDisjointPathsinAdHocNet-works”;IEEEICC'01.[24]KuiWu,JanelleHarms;“PerformanceStudyofaMultipathRoutingMethodforWirelessMo-bileAdHocNetworks”;MASCOTS,2001.[25]LianfangZhang,ZenghuaZhao,YantaiShu,LeiWang,andOliverW.W.Yang;“LoadBalanc-ingofMultipathSourceRoutinginAdHocNet-works”;Proc.ofIEEEICC2002.[26]S.Bak,J.A.Cobb,E.L.Leiss;“Load-BalancedRoutingviaRandomization”;PDCS,1999.[27]K.Gopalan;“EfcientNetworkResourceAllo-cationwithQoSGuarantees”;TR#133,ECSL,SUNY-SB.[28]A.Caprara,H.Kellerer,U.Pferschy;“TheMul-tipleSubsetSumProblem”;SIAMJournalonOptimization,V11,'00.[29]V.Bahl,A.Adya,J.Padhye,A.Wolman.“Re-consideringtheWirelessLANPlatformwithMultipleRadios”;WorkshoponFutureDirec-tionsinNetworkArchitecture'03.[30]P.H.Hsiao,A.Hwang,H.T.Kung,andD.Vlah;“Load-BalancingRoutingforWirelessAccessNetworks”;Proc.ofIEEEINFOCOM2001.[31]I.KatzelaandM.Naghshineh;“Channelassign-mentschemesforcellularmobiletelecommuni-cationsystems:acomprehensivesurvey”;IEEEPersonalComm.June'96[32]T.R.Jensen,B.Toft;“GraphColoringProb-lems”;WileyInterscience,NewYork,1995.[33]D.Couto,et.al.;“AHigh-ThroughputPathMet-ricforMulti-HopWirelessRouting”;MobiCom,2003.[34]EngimInc;www.engim.comMobileComputingandCommunicationsReview,Volume8,Number265