InterferenceAware Channel Assignment in MultiRadio Wireless Mesh Networks Krishna N - PDF document

Interference-AwareChannelAssignmentinMulti-RadioWirelessMeshNetworksKrishnaN.Ramachandran,ElizabethM.Belding,KevinC.Almeroth,MilindM.BuddhikotyUniversityofCaliforniaatSantaBarbarayLucentBellLabs,Holmdelfkrishna,ebelding,almerothg@cs.ucsb.edumbuddhikot@lucent.comAbstract—Thecapacityprobleminwirelessmeshnetworkscanbealleviatedbyequippingthemeshrouterswithmultiplera-diostunedtonon-overlappingchannels.However,channelassign-mentpresentsachallengebecauseco-locatedwirelessnetworksarelikelytobetunedtothesamechannels.Theresultingincreaseininterferencecanadverselyaffectperformance.Thispaperpresentsaninterference-awarechannelassignmentalgorithmandprotocolformulti-radiowirelessmeshnetworksthataddressthisinterferenceproblem.Theproposedsolutionintelligentlyassignschannelstoradiostominimizeinterferencewithinthemeshnetworkandbetweenthemeshnetworkandco-locatedwirelessnetworks.Itutilizesanovelinterferenceestimationtechniqueimplementedateachmeshrouter.Anextensiontotheconictgraphmodel,themulti-radioconictgraph,isusedtomodeltheinterferencebetweentherouters.Wedemonstrateoursolution'spracticalitythroughtheevaluationofaprototypeimplementationinaIEEE802.11testbed.Wealsoreportonanextensiveevaluationviasimulations.Inasamplemulti-radioscenario,oursolutionyieldsperformancegainsinexcessof40%comparedtoastaticassignmentofchannels.I.INTRODUCTIONTypicaldeploymentsofstaticmulti-hopwirelessnetworks,calledwirelessmeshnetworks,utilizeroutersequippedwithonlyoneIEEE802.11radio.IEEE802.11radiosaretypicallysingle-channelradios.Asaresult,single-radiomeshnetworkscansufferfromseriouscapacitydegradationduetothehalf-duplexnatureofthewirelessmedium[10].Fortunately,theIEEE802.11PHYspecicationpermitsthesimultaneousoperationofmultiplenon-overlappingchannels.Forexample,threenon-overlappingchannelsinthe2.4GHzbandcanbesimultaneouslyused.TheIEEE802.11aspeci-cationallowsuptotwelvenon-overlappingchannelsinthe5.0GHzband.Bydeployingmulti-radioroutersinwirelessmeshnetworksandassigningtheradiostonon-overlappingchannels,therouterscancommunicatesimultaneouslywithminimalinterferenceinspiteofbeingindirectinterferencerangeofeachother.Therefore,thecapacityofwirelessmeshnetworkscanbeincreased.Inequippingrouterswithmultipleradios,ana¨vestrategywouldbetoequipeachrouterwiththenumberofradiosequaltothenumberoforthogonalchannels.However,thisstrategyiseconomicallyprohibitiveduetothesignicantnumberofnon-overlappingchannels.Furthermore,smallform-factorembeddedsystemsusedformanufacturingrouterssupportonlyalimitednumberofradios.Consequently,usingallnon-overlappingchannelsonameshrouterisstillnotaviableoption.Theassignmentofchannelstoameshrouterthenbecomesaproblemofchoosingwhichchannelstoassigntowhichofitsradios.Asimpletechniqueistousestaticchannelassignment.However,withtheexplosivegrowthin“WiFi”deploymentsthatoperateinthesame(unlicensed)spectrumaswirelessmeshnetworks,anystaticassignmentwilllikelyresultintheoperationofthemeshonchannelsthatarealsousedbyco-locatedWiFideployments.Theresultingincreaseininterferencecandegradetheperformanceofthemeshnetwork.Thispaperaddressesthechannelassignmentproblemandspecicallyinvestigatesthedynamicassignmentofchan-nelsinawirelessmeshnetwork.Wepresentacentralized,interference-awarechannelassignmentalgorithmandacorre-spondingchannelassignmentprotocolaimedatimprovingthecapacityofwirelessmeshnetworksbymakinguseofallavail-ablenon-overlappingchannels.Thealgorithmintelligentlyselectschannelsforthemeshradiosinordertominimizeinterferencewithinthemeshnetworkandbetweenthemeshnetworkandco-locatedwirelessnetworks.Eachmeshrouterutilizesanovelinterferenceestimationtechniquetomeasurethelevelofinterferenceinitsneighborhoodbecauseofco-locatedwirelessnetworks.Thealgorithmutilizesanextensiontotheconictgraphmodel[14],theMulti-radioConictGraph(MCG),tomodelinterferencebetweenthemulti-radioroutersinthemesh.TheMCGisusedinconjunctionwiththeinterferenceestimatestoassignchannelstotheradios.Onepotentialpitfallofdynamicchannelassignmentisthatitcanresultinachangeinthenetworktopology.Topologychangescanleadtosub-optimalroutingandevennetworkpartitioningincaseofnodefailures.Theproposedsolution,therefore,ensuresthatchannelassignmentdoesnotalterthenetworktopologybymandatingthatoneradiooneachmeshrouteroperateonadefaultchannel.Asecondpotentialpitfallisthatchannelassignmentcanresultindisruptionofowswhenthemeshradiosarereconguredtodifferentfrequencies.Topreventowdisruption,linkredirectionisimplementedateachmeshrouter.Thistechniqueredirectsowsoverthedefaultchanneluntilthechannelassignmentsucceeds.WeevaluateourproposedsolutionthroughsimulationsinQualnet.WeutilizetheOptimizedLinkStateRouting(OLSR)protocol[8]andtheWeightedCumulativeExpectedTrans-missionTime(WCETT)metric[9]forrouteselection.Wedemonstratethepracticalityofourproposedsolutionviatheevaluationofaprototypeimplementationinamulti-radioIEEE802.11btestbed. A.ResearchContributionsTothebestofourknowledge,oursistherstsolutiontoaddresstheproblemofdynamicchannelassignmentinwirelessmeshnetworksinthepresenceofinterferencefromco-locatedwirelessnetworks.Akeygoalinthedesignoftheproposedsolutionhasbeentomakethesolutionamenabletoeasyimplementationusingcurrentlyavailableradios.Thisdifferentiatesourworkfromseveralproposedsolutions(sur-veyedinSectionIX)whichrequireeitherspecializedasyetunavailableradiosorknowledgeaboutthenetworksuchasanticipatedtrafcpatternsandthespecicpathstobetraversedbynetworkows.Specically,thecontributionsofthispaperareasfollows:Adynamic,interference-awarechannelassignmental-gorithmthatminimizesinterferencebetweenthemeshnetworkandco-locatedwirelessnetworks.Amulti-radioconictgraph,anextensiontothewell-knownconictgraphmodel,tomodeltheinterferencerelationshipbetweenmulti-radioroutersinawirelessmeshnetwork.Anovelinterferenceestimationschemethatroutersusetoestimatetheinterferencelevelintheirneighborhoods.Alinkredirectionprotocolthatpreventsthedisruptionofowsduringchannelassignment.Acomprehensiveperformancestudythatshowssigni-cantthroughputimprovementsinthepresenceofvaryinginterferencelevels,whicharevalidatedthroughempiricalmeasurementsonaprototypeimplementation.B.PaperOutlineTheremainderofthepaperisorganizedasfollows:Sec-tionIIdiscussestheeffectofchannelassignmentonnet-worktopology.InSectionIII,weformulatethechannelassignmentproblem.SectionIVdescribesourinterferenceestimationtechniqueandthemulti-radioconictgraphmodel.InSectionV,wepresentourcentralizedchannelassignmentalgorithm.WediscussthechallengesweaddressedduringthedevelopmentofourprototypeimplementationinSectionVI.SectionVIIpresentsresultsfromoursimulation-basedevalua-tion,whileresultsfromourprototypeevaluationarepresentedinSectionVIII.InsectionIX,wesummarizerelatedwork,andSectionXconcludesthepaper.II.CHANNELASSIGNMENTANDNETWORKTOPOLOGYInamulti-radiomeshnetwork,channelassignmenttoradioscanalterthenetworktopology.ConsidertheexamplefournodetopologyinFigure1(a).Here,nodeCisequippedwiththreeradiosandtheothernodes(A,B,andD)haveoneradioeach.Eachlinkinthegureislabeledwithitschannelnumber.Figure1(a)illustratesthetopologywhenallradiosaretunedtochannelone.Figure1(b)illustratesthechangeinnetworktopologyafterchannelassignment.Alterationsinthenetworktopologyhavethreemaindraw-backs.First,subsequentnodefailureshaveahigherprobabilityofcausingnetworkpartitions.Consequently,portionsofthemeshmaybecomeunreachable,resultinginthedisruption DCBA11111 (a) DCBA123 (b) DCBA123 (c)Fig.1.Networktopologywithvaryingchannelassignments.ofows.ThisisclearlyseeninFigure1(c).WhennodeCfails,thefournodenetworkispartitionedintothreedifferentclusters.Reconnectionofthenetworkwouldrequirecomplexsynchronizationschemestobeimplementedatthemeshrouters[25].Second,topologyalterationscanresultinsub-optimalroutesbetweennodepairswithrespecttosomemetric,suchasthroughput,delay,orreliability.Toillustratehowthiscanoccur,consideragainFigure1(a).NodeAcancommunicatewithnodeBonaonehoppath.Afterchannelassignment,AcancommunicatewithBonlyoveratwohoppathviaC,asshowninFigure1(b).Selectionofapathwithahigherhop-countisnotpreferredforthreereasons:(1)longer,frequencydiversiedpathsoftenyieldworseperformancethanshorterpaths;(2)theinterference“foot-print”ofaowonahigherhop-countpathisnaturallygreater;and(3)alongerpathismorepronetofailure.Note,however,thatwedonotclaimthatalllongerpathsarelikelytoperformpoorlycomparedtoshorterpathsbecausetheperformanceofeachpathalternativeislikelytovarywithfactorssuchastrafcpattern,nodeplacement,radiocharacteristics,andterrain.Nevertheless,westressthatitischallengingtoaccuratelypredict,inpractice,whichchannelassignmentalternativeandresultingnetworktopologycongurationwillyieldoptimumperformance.Thethirddrawbackofalteringanetwork'stopologyisthatitaffectsexistingows.Forexample,letusassumethatlinkCDinFigure1(b)isassignedanewchannel.Theprocessofchannelassignmentmustbeaccuratelycoordinated;otherwise,casesmayarisewhereoneradioonthelinkswitchestothenewfrequencybutthesecondradiodoesnotbecauseacontrolmessageiseitherlostordelayed.Consequently,anyowsfromDtotherestofthenetworkthatexistedatthetimeofthechannelassignmentaredisruptedduringtheswitch.Overcom-ingsuchcasesischallenginginpracticebecausecongurationoftheradiosrequirestime-synchronizedcoordinationbetweenthemeshroutersduringchannelassignment.Becauseofthesedrawbacksassociatedwithnetworktopol-ogychanges,weadvocatethattopologyalterationsshouldbeavoided.Wemandatethisbyrequiringthatallroutersinthemeshnetworkdesignateoneoftheirradiostobeadefaultradiointerface.Thisdefaultradioisofthesamephysicallayertechnology,either802.11a,802.11b,or802.11g,andistunedtoacommonchannelthroughoutthemesh.Thedefaultchannelcarriesbothcontrolanddatatrafc.Thisstrategyhasseveraladvantages.First,itpreventschangesinthetopologyofthenetworkbecauserouterswilldiscoverotherwisedisconnectedneighborsbycommunicatingoverthedefaultradiointerface.Second,overcomingnode failureissimpliedbecausearouterwillbeabletochoosealternatepathstoroutearoundafailednode.Third,theroutingprotocolwillnowhavetheoptionofselectingapaththatisnotfrequencydiversiedifithasbetterperformancechar-acteristicsthanafrequencydiversiedalternative.Asanaladvantage,anydisruptionofowsduringchannelassignmentcanbeavoidedbyredirectingowsoverthedefaultradiountiltheassignmentcompletes.TheredirectiontechniqueisfurtherelaboratedinSectionVI.WeconsiderthereassignmentofthedefaultchannelinSectionVI-D.III.PROBLEMFORMULATIONThechannelassignmentalgorithmweproposeinthispaperisdesignedforwirelessmeshnetworks.Routersinsuchnetworksarestationary.However,userdevices,suchaslaptopsandPDAs,canbemobile.Suchdevicesassociatewithroutersthatalsofunctionasaccesspoints.Figure2illustratesourmodelofamulti-radiomeshnet-work.Inourmodel,themeshroutersareassumedtobeequippedwithmultipleIEEE802.11radios,suchas802.11a,802.11b,or802.11g.Theroutersneednotallbeequippedwiththesamenumberofradiosnordotheyneedallthreetypesofradios.Dependingonthenumberofradiosateachmeshrouter,weclassifytheroutersintotwocategories:(1)Multi-Radiomeshrouters(MRs);and(2)Single-Radiomeshrouters(SRs).WemandatethateachMRandSRinthenetworkbeequippedwithoneradio,calledthedefaultradio,whichisofthesamephysicallayertype,e.g.802.11b,andtunedtothesamechannelasmotivatedinSectionII.Atleastonerouterinthemeshisdesignatedasagateway.Thegatewaysprovidesconnectivitytoanexternalnetwork.Inordertosimplifytheexplanationofthechannelassignmentsolution,weassumethepresenceofonlyonegateway.AccessPoints(APs)provideconnectivitytouserdevicesandareco-locatedwithmeshrouters.Amajorityofthetrafcwithinthemeshiseitherfromtheuserdevicestothegatewayorvice-versa.Thistrafcpatternistypicalinwirelessmeshdeployments.Becausethetrafcpatternisskewedto-and-fromthegateway,thepathstakenbytheresultingowsarelikelytoformatreestructureinwhichthegatewayisthe“root”andtheuserdevicesarethe“leaves”.Trafcowswilllikelyaggregateatroutersclosetothegateway.Therefore,inordertoimproveoverallnetworkcapacity,itispreferabletoplaceMRsclosetothegatewayandinregionsofthemeshthatarelikelytoexperienceheavyutilization.Itisimportantthattheplacementoccuraftercarefulnetworkplanninginordertooptimizenetworkperformance,reduceequipmentcosts,andaddresslogisticalconstraints.ThedottedlinesinthegureillustratelinksbetweenMRsthataretunedtonon-overlappingchannels.Inourexample,vesuchchannelsareused.Asixthchannel,indicatedbysolidlines,isthedefaultchannel.TheChannelAssignmentServer(CAS),whichisco-locatedwiththegatewayinthegure,performschannelassignmenttoradios.Inassigningchannels,theCASshouldsatisfythefollowinggoals: Fig.2.Multi-Radiowirelessmesharchitecture.Minimizeinterferencebetweenroutersinthemesh:Insatisfyingthisgoal,threesub-goalsneedtobeachieved.First,theCASshouldsatisfytheconstraintthatforalinktoexistbetweentworouters,thetwoend-pointradiosonthemmustbeassignedacommonchannel.Second,linksindirectcommunicationrangeofeachothershouldbetunedtonon-overlappingchannels.Third,becauseofthetreeshapedtrafcpatternexpectedinwirelessmeshnetworks,channelassignmentpriorityshouldbegiventolinksstartingfromthegatewayandthentolinksfanningoutwardstowardstheedgeofthenetwork.Minimizeinterferencebetweenthemeshnetworkandwirelessnetworksco-locatedwiththemesh:Insatisfyingthisgoal,theCASshouldperiodicallydeterminetheamountofinterferenceinthemeshduetoco-locatedwirelessnetworks.Theinterferencelevelisestimatedbyindividualmeshrouters.TheCASshouldthenre-assignchannelssuchthattheradiosoperateonchannelsthatexperiencetheleastinterferencefromtheexternalradios.Giventhesegoalsforthechannelassignmentalgorithm,nextwepresentdetailsoninterferenceestimationanddescribetheinterferencemodelingtechnique.IV.INTERFERENCEESTIMATIONANDMODELINGThissectionpresentsanoverviewoftheinterferenceestima-tionprocedure.ImplementationdetailsarelefttoSectionVI.ThissectionalsointroducestheMulti-radioConictGraph(MCG)model.A.InterferenceEstimationThegoalofinterferenceestimationistoperiodicallymea-suretheinterferencelevelineachmeshrouter'senvironment.Accuratemeasurement,however,ischallengingandrequiresthatexpensivehardwarebeused[1].Instead,asanapproximation,werelyonthenumberofinterferingradiosoneachchannelsupportedbyeachrouterasanestimationofinterference.Aninterferingradioisdenedasasimultaneouslyoperatingradiothatisvisibletoarouter butexternaltothemesh.Avisibleradioisonewhosepacket(s)passFrameCheckSequence(FCS)checksandarethereforecorrectlyreceived.WeassumethattheCASinformstherouterofradiosinternaltothemesh.TheinformationcouldconsistofanIPaddressrangeoranexhaustivelistofallradioMACaddressesinthemesh.Onecaveattotheaboveestimationprocedureisthatcarrier-sensingradios,i.e.,thoseradiosthatarewithinanestimatingrouter'scarriersensingrangebutoutsideitsreceptionrange,willnotbeaccountedforintheestimation.ThisisbecausepacketstransmittedbysuchradioswillfailFCSchecksperformedbytherouter.However,carrier-sensingradiosmaystillinterferewiththerouter.Ourinterferenceestimationtechniquedoesnotconsidersuchradiosfortworeasons.First,recentstudies[12],[24]suggestthatcurrentIEEE802.11MACimplementationsareoverlyconservativeintheircarriersensemechanismandoftenoverestimatetheadverseimpactofinterferingradios.Therefore,eveninthepresenceofmultiplecarrier-sensingradios,theperformancedegradationduetocarrier-sensingneighborsmaynotbeassevereaspreviouslyunderstood.Second,evenifweweretoincorporatecarrier-sensingradiosinourinterferenceestimationsolution,itisimpossibletodeterminethepresenceofsuchradiosusingcommodityhardwarebecauseoftheinabilityofcurrentrmwareimplementationstoidentifythem1.Sanzgirietal.proposetousespecializedhardwaretoovercomethermwarelimitations[22].Suchhardwarearelikelytobeavailableinthefutureandcanbeleveragedwhenavailable.Measurementofonlythenumberofinterferingradios,how-ever,isnotsufcientbecauseitdoesnotindicatetheamountoftrafcgeneratedbytheinterferingradios.Forinstance,twochannelscouldhavethesamenumberofinterferingradiosbutonechannelmaybeheavilyutilizedbyitsinterferingradioscomparedtotheother.Therefore,inaddition,eachmeshrouteralsoestimatesthechannelbandwidthutilizedbytheinterferingradios.Theinterferenceestimationprocedureisasfollows:ameshrouterconguresoneradioofeachsupportedphysicallayertypetocapturepackets2oneachsupportedchannelforasmallduration.Therouterusesthecapturedpacketstomeasurethenumberofinterferingradiosandpersecondchannelutilization.ThenumberofinterferingradiosissimplythenumberofuniqueMACsexternaltothemesh.Theutilizationoneachchannelduetotheinterferingradiosiscomputedfromthecaptureddataframesbytakingintoaccountthepacketsizesandtheratesatwhichthepacketsweresent[13].TheoverheadoftheMAClayerisaccountedforinourutilization1Wirelessdevices,suchasonesusingthePrism2/2.5chipset,sometimesallowthecaptureofpacketstransmittedbycarrier-sensingradiosthatfailtheFCScheck.Thismechanismatrstmightsuggestatechniquetoidentifythecarrier-sensingradios.However,theutilityofthiscapturemechanismislimitedbecausetheinformationcontainedinthegarbledpacketsisbynaturefaulty.2PacketcapturemodeasimplementedoncurrentlyavailableIEEE802.11radioscannotcapturepacketsfromradios,suchascordlessphonesorBlue-toothdevices,thatuseotherphysicallayertechnologies.Wenote,however,thattheinterferencefoot-printofsuchdevicesislikelytobesmall.Software-denedradiosarelikelytoaddressthislimitationinthefuture. C-2B-1A-1 (a) ACBCCDAB (b) D-1 : C-1D-1 : C-2 (c)Fig.3.(a)Asimplenetworktopology,G.(b)Correspondingconictgraph,F.(c)Correspondingmulti-radioconictgraph,F0.calculation.Wesetthedurationofthepacketcapturetothreesecondsinourimplementation.Thethreeseconddurationislargeenoughtoallowfortheaveragingofthevariationsinpersecondmeasurementsandissmallenoughtoenabletheinterferenceestimationtocompletequickly.Eachmeshrouterthenderivestwoseparatechannelrank-ings.Therstrankingisaccordingtoincreasingnumberofinterferingradios.Thesecondrankingisaccordingtoincreasingchannelutilization.Themeshrouterthenmergestherankingsbytakingtheaverageoftheindividualranks.TheresultingrankingissenttotheCAS.B.InterferenceModelingConictgraphsareusedextensivelytomodelinterferenceincellularradionetworks[14].Aconictgraphforameshnetworkisdenedasfollows:consideragraph,G,withnodescorrespondingtoroutersinthemeshandedgesbetweenthenodescorrespondingtothewirelesslinks.Aconictgraph,F,hasverticescorrespondingtothelinksinGandhasanedgebetweentwoverticesinFifandonlyifthelinksinGdenotedbythetwoverticesinFinterferewitheachother.Asanexampleofaconictgraph,Figure3(a)showsthetopologyofanetworkwithfournodes.Eachnodeinthegureislabeledwithitsnodenameanditsnumberofradios.Figure3(b)showstheconictgraph.Atarstglance,theproblemofassigningchannelstolinksinameshnetworkappearstobeaproblemofvertexcoloringtheconictgraph.However,vertexcoloringfailstoassignchannelscorrectlybecauseitdoesnotaccountfortheconstraintthatthenumberofchannels(colors)assignabletoaroutermustbeequaltoitsnumberofradios.Asanexampleofwhythisisthecase,letusassumethatthefourverticesintheconictgraphshowninFigure3(b)areeachassignedoneofthreedifferentchannelsusingavertexcoloringalgorithm.Thismeansthatthetworadiosrepresentedbyeachvertexintheconictgraphoperateonthefrequencyassignedtothatvertex.ThisimpliesthatnodeCintheillustratednetworkoperatesonthreedifferentchannels,whichisimpossiblebecauseitisequippedwithonlytworadios.Theconictgraphdoesnotcorrectlymodelroutersequippedwithmultipleradios.Therefore,weextendtheconictgraphtomodelmulti-radiorouters.Intheextendedmodel,calledtheMulti-radioConictGraph(MCG),werepresentedgesbetweenthemeshradiosasverticesinsteadofrepresentingedgesbetweenthemeshroutersasverticesas intheoriginalconictgraph.TocreatetheMCG,F0,werstrepresenteachradiointhemeshnetworkasavertexinG0insteadofrepresentingroutersbyverticesasinG.Therefore,intheaboveexample,nodeCisrepresentedbytwoverticesinG0correspondingtoitstworadiosinsteadofjustonevertexinG.TheedgesinG0arebetweenthemeshradiosinsteadofthemeshroutersasinG.WethenrepresenteachedgeinG0usingavertexinF0.TheedgesbetweentheverticesinF0arecreatedinthesamewaytheoriginalconictgraphiscreated,i.e.,twoverticesinF0haveanedgebetweeneachotheriftheedgesinG0representedbythetwoverticesinF0interferewitheachother.Asanexample,Figure3(c)showsthemulti-radioconictgraphofthenetworkshowninFigure3(a).Inthegure,eachvertexislabeledusingtheradiosthatmakeupthevertex.Forexample,thevertex(A1:C2)representsthelinkbetweentherstradioonrouterAandthesecondradioonrouterC.WhenusingavertexcoloringalgorithmtocolortheMCG,weimposeanimportantconstraint:oncoloringanyMCGver-tex,alluncoloredverticesintheconictgraphthatcontainanyradiofromthejust-coloredvertexberemoved.Forexample,afterassigningacolortovertex(A1:C2)inFigure3(c),allverticescontainingeitherA1orC2shouldberemovedfromtheconictgraph.Thisisrequiredtoensurethatonlyonechannelisassignedtoeachradiointhemeshnetwork.V.CHANNELASSIGNMENTALGORITHMA.OverviewThechannelassignmentproblemformeshnetworksissimilartothelistcoloringproblem,whichisdenedasfollows:givenagraph,G=(V;E),andforeveryvinV,alistL(v)ofcolors,isitpossibletoconstructavalidvertexcoloringofGsuchthateveryvertexvreceivesacolorfromthelistL(v)?ThelistcoloringproblemisNP-complete[21].Therefore,werelyonanapproximatealgorithmforchannelassignment.Ouralgorithm,calledtheBreadthFirstSearchChannelAssignment(BFS-CA)algorithm,usesabreadthrstsearchtoassignchannelstothemeshradios.Thesearchbeginswithlinksemanatingfromthegatewaynode.Therationalebehindtheuseofbreadthrstsearchisintuitive:byusingbreadthrstsearch,wesatisfyourgoaldescribedinSectionIIIofgivingchannelassignmentprioritytolinksstartingfromthegatewayandthenindecreasinglevelsofprioritytolinksfanningoutwardtowardstheedgeofthenetwork.BeforeusingtheBFS-CAalgorithm,thechannelassignmentserver(CAS)obtainstheinterferenceestimatesfromthemeshrouters.Itthenchoosesachannelforthedefaultradios.Thedefaultchannelischosensuchthatitsuseinthemeshnetworkminimizesinterferencebetweenthemeshnetworkandco-locatedwirelessnetworks.TheCASthencreatestheMCGforthenon-defaultradiosinthemesh.WeuseatwohopinterferencemodeltocreatetheMCG.Inthismodel,twomeshlinksareinterferingiftheyeitherhaveacommonrouterorareseparatedbyahopasdeterminedfromtheneighborinformationsentbyeachmeshroutertotheCAS.Padhyeetal.proposeanempiricaltechniquetodetectinterferingmeshlinks[17].Thistechniquecanbemoreaccuratebecauseofitsempiricalnature,however,takesalongtime(severalhours)tocomplete.Wearecurrentlyinvestigatingapproachestospeedupthistechnique.Inthemeantime,weleveragethetwohopmodelinourwork.AfterconstructingtheMCG,theCASusestheBFS-CAalgorithmtoselectchannelsforthenon-defaultradios.Oncethechannelsareselectedforthemeshradios,theCASinstructstherouterstoconguretheirradiostothenewlyselectedchannels.Tosimplifytheexplanationofthechannelselectionprocedureinthissection,letusassumefornowthatthemeshradiosarereconguredatthesametime.WeaddressthisassumptioninSectionVI-D,whereweprovidedetailsonthespecicprotocolusedtore-assignchannels.ThedefaultchannelselectionprocedureispresentednextfollowedbyadetaileddescriptionoftheBFS-CAalgorithm.TheCASperiodicallyinvokesthechannelselectionproceduresummarizedabovetocopewiththevaryingnatureofinterfer-enceinthemesh.Thissectionendswithadiscussionoftheperiodofinvocationanditsimplications.B.DefaultChannelSelectionTheCASchoosesthedefaultchannelusingtherankofachannel,c,fortheentiremesh,Rc.Rciscomputedasfollows:Rc=Pni=1Rankic nwherenisthenumberofroutersinthemeshandRankicistherankofchannelcatrouteri.ThedefaultchannelisthenchosenasthechannelwiththeleastRcvalue.Theintuitionbehindthismetricistousetheleastinterferedchannelasthedefaultchannelinthemesh.Usingsuchachannelsatisesourgoalofminimizinginterferencebetweenthemeshandco-locatedwirelessnetworks.C.Non-DefaultChannelSelectionInthisphase,theCASusestheneighborinformationcollectedfromallrouterstoconstructtheMCG.Neighborinformationsentbyaroutercontainstheidentityofitsneighbors,delaytoeachneighbor,andinterferenceestimatesforallchannelssupportedbytherouter'sradios.SectionVI-Cdetailsthecalculationoflinkdelayperformedbymeshrouters.TheCASassociateswitheachvertexintheMCGitscorrespondinglinkdelayvalue.TheCASalsoassociateswitheachvertexachannelrankingderivedbytakingtheaverageoftheindividualchannelrankingsofthetworadiosthatmakeupthevertex.TheaverageisimportantbecausetheassignmentofachanneltoavertexintheMCGshouldtakeintoaccountthepreferencesofbothend-pointradiosthatmakeupthevertex.ForallverticesintheMCG,theCASthencomputestheirdistancesfromthegateway.ThedistanceofanMCGvertexistheaverageofthedistancesfromthegatewayofthetworadiosthatmakeupthevertex.Thedistanceofaradioisobtainedfrombeaconsinitiatedbythegateway.Abeaconisagatewayadvertisementbroadcastedhop-by-hopthroughoutthemesh. Algorithm1BFS-CAAlgorithm 1:LetV=fvjv2MCGg2:whilenotAllVerticesVisitedfVgdo3:Leth=smallestHopCount(V)4:Q=fvjv2VandnotVisited(v)andhopcount(v)==hg5:sort(Q)6:whilesize(Q)�0do7:vcurrent=removeHead(Q)8:ifvisited(vcurrent)then9:continue10:endif11:visit(vcurrent)12:Vn=fuju2MCGandedgeInMCG(u;vcurrent)==TRUEg13:permanentlyassignhighestrankedchannelcfromvcurrent'schan-nelrankingthatdoesnotconictwithui,fui2Vnand0isize(Vn)g14:ifcdoesnotexistthen15:permanentlyassignrandomchanneltovcurrent16:endif17:L=fvjv2MCGandvcontainseitherradiofromvcurrentg18:removeVerticesInListFromMCG(L)19:tentativelyassignctoradiosinLthatarenotpartofvcurrent20:Letrfberouterwithinterfaceinvcurrentthatisfarthestawayfromgateway21:LetTail=listofallactivev(v2MCG)suchthatvcontainsaninterfacefromrf22:sort(T)23:addToQueue(Q,Tail)24:endwhile25:permanentlyassignchannelstoradiosthatareunassignedapermanentchannel.26:endwhile Eachbeaconcontainsahop-counteldthatisincrementedateachhopduringitsbroadcast.Thedistanceofarouterfromthegatewayistheshortestpathlengthofasinglebeaconinstancereceivedbytherouteroverallpaths.TheroutercommunicatesthedistancetotheCASviaperiodicheartbeatmessagessenteveryminuteinourimplementation.Algorithm:Oncetheaveragedistancesarecomputed,theCASusestheBFS-CAalgorithmtoassignchannelstothemeshradios.ThealgorithmissummarizedinAlgorithm1.ThealgorithmstartsbyaddingallverticesfromtheMCGtoalist,V(Line1).ItdoesabreadthrstsearchoftheMCGtovisitallverticesandassignthemchannels.Thesearchstartsfromverticesthatcorrespondtolinksemanatingfromthegateway(Lines3,4).Inline3,thesmallesthopcountvertexisdeterminedofallverticesintheMCG.Inline4,allverticeswithdistanceequaltothesmallesthopcountareaddedtoaqueue,Q.Ifverticescorrespondtonetworklinksemanatingfromthegateway,theirhopcountis0:5.Theseverticesarethensortedbyincreasingdelayvalues(Line5).Thissortisperformedinordertogivehigherprioritytothebetterlinksemanatingfromtheshortesthopcountrouter(thegatewayfortherstBFSiteration).ThealgorithmthenvisitseachvertexinQ(Line11)andpermanentlyassignsthemthehighestrankedchannelthatdoesnotconictwiththechannelassignmentsofitsneighbors(Line13).Ifanon-conictingchannelisnotavailable,aran-domlychosenchannelispermanentlyassignedtothevertex.Note,however,thatthedefaultchannelisneverassigned.Onceavertexisassignedachannel,allverticesthatcontaineitherradiofromthejust-assignedvertexareplacedinalist,L(Line17).Inline18,allverticesfromLareremovedfromtheMCG.Thisstepisneededtosatisfytheconstraintthatonlyonechannelisassignedtoeachradio.Theradiosinthelistofverticesthatdonotbelongtothejust-assignedvertexaretentativelyassignedthelatter'schannel(Line19).Inlines20-21,verticesatthenextlevelofthebreadthrstsearchareaddedtoQ.Theseverticescorrespondtolinksthatfan-outfromthegatewaytowardstheperiphery.TondsuchlinksintheMCG,twostepsareperformed.Intherststep(Line20),therouterfromthejust-assignedvertexthatisfarthestawayfromthegatewayischosen;thefarthestrouteristherouterwiththehigherhop-countofthetworoutersthatmakeupthejust-assignedvertex.Inthesecondstep(Line20),allunvisitedMCGverticesthatcontainaradiobelongingtothefarthestrouterareaddedtothelist,Tail.Thislistissorted(Line22)byincreasingvalueofthedelaymetrictogivehigherprioritytobetterlinksthatemanatefromthefarthestrouter.Finally,inline23,theverticesfromTailareaddedtoQ.TheabovedescribedalgorithmcontinuesuntilallverticesintheMCGarevisited.Inline25,anyradiothatisnotassignedapermanentchannelduringthesearch,becauseverticescontainingitweredeletedinline18,ispermanentlyassignedoneofthechannelstentativelyassignedtoitinline19.Oncechannelassignmentsaredecided,theCASnotiesthemeshrouterstore-assigntheirradiostothechosenchannels.TheexactprotocolisdescribedinSectionVI-D.D.ChannelRe-assignmentStrategyToadapttothechanginginterferencecharacteristics,theCASperiodicallyre-assignschannels.Theperiodicitydependsultimatelyonhowfrequentlyinterferencelevelsinthemeshnetworkareexpectedtochange.Ifalargenumberofinterfer-ingdevicesinthevicinityofthemeshnetworkareexpectedtobeshort-lived,theinvocationrateshouldbeincreased.Ontheotherhand,ifamajorityoftheinterferingdevicesarelikelytobelong-lived,theinvocationratecanbedecreased.Inourimplementation,wehavesettheratetotenminutes.Webelievethisrateresultsinagoodtradeoffbetweeninterferenceadaptationandmeshradioreconguration.Nevertheless,weexpectthenetworkoperatorofameshnetworktochoosearatetobestsuitthetargetdeployment.VI.IMPLEMENTATIONCONSIDERATIONSIntheearliersections,weomitteddetailsonseveralim-plementationspecicsteps,suchasthechannelestimationprocedure,linkdelayestimation,andthechannelassignmentprocess.ThissectiondescribestheimplementationaspectsofthesestepsastheybecomecriticalwhenBFS-CAisusedinpractice.A.InterferenceEstimationEachmeshrouterperformsinterferenceestimationbycap-turingpacketsfromthemedium.Tocapturepackets,weuseaspecialoperatingmodesupportedbytypicalradiohardwarecalledtheRFMonmode.ThismodeallowsIEEE802.11managementframesandregulardataframesinthemediumto becaptured.Themanagementframesareperiodicallytrans-mitted(typicallyevery100msec)bydevicesthatimplementtheIEEE802.11specication.WedonotusethetypicalpromiscuousmodesupportedbytheradiosbecausethatmodeonlysniffsdataframesanddoesnotsniffIEEEmanagementframes.Consequently,devicesthatdonottransmitanydatapacketsduringthecaptureperiodwillnotbediscovered.TheRFMonmodeprovidesinformationabouteachpacket,suchastherateatwhichitwassentandthesizeofthepacket.Itderivesthisinformationbyinterpretingphysicallayerinformationassociatedwiththepacketandthereforecanprovidetheinformationeveniflayer2encryptionisusedtoconcealthecontentsofthepackets.Thecapturedpacketsarethenusedtomeasurethenumberofinterferingradiosandthebandwidthconsumedbythoseradios.Radios,whenplacedinRFMonmode,cannottransmitdatapacketsforthedurationofthesnifng.Thisisbecausesomecommercialradios,suchasthoseutilizingthePrism2/2.5chipset,cannotperformpackettransmissioninRFMonmode.EvenifradioscantransmitinRFMonmode,suchasonesuti-lizingtheAtheroschipset,becausetheinterferenceestimationateachrouteroccursindependentlyofitsneighbors,thereisnoguaranteethattworadioswillbeonthesamechannelinordertocommunicatesuccessfully.Therefore,anyowthatusesaradiotemporarilyinRFMonmodewillbedisrupted.Topreventowdisruption,weuselinkredirection,describednext.B.LinkRedirectionLinkredirectionisachievedwhenaowintendedforarouter'snon-defaultradioisredirectedtotherouter'sdefaultradioinstead.Linkredirectionisinvokedintwocases:(1)whenarouter'sintendedtransmitterisincapableofdeliveringpackets,and(2)whentheintendedreceiverontheneighboringrouterisincapableofreceivingpackets.Redirectionispossiblebecausethedefaultradiosonallroutersoperateonthedefaultchannel.Ourlinkredirectionprotocolisasfollows:wheneveraradiohastochangeitsstatetoinactive,itbroadcastsanINTERFACE-INACTIVEmessageeverysecondforthreesec-ondsbeforeitchangesstate.Themultiplebroadcastsal-lowforanymessagelosses.AnyneighborthatreceivestheINTERFACE-INACTIVEmessagedeletestheaddressofthesoon-to-becomeinactiveradiofromitsroutingtables.Oncethedeletionoccurs,linkredirectionisinvokeduntiltheradiobecomesactiveagain.Wesimplyrelyontheroutingprotocoltonotifythataradiohasbecomeactive.C.LinkDelayEstimationWemeasurelinkdelayusingtheExpectedTransmissionTime(ETT)metric[9].ETTofalinkisderivedfromthelink'sbandwidthandlossrate.Duetospaceconstraints,wespecifybelowonlythevaluesofvariousparametersusedtocalculateETT.Amoredetaileddescriptionofthemetriccanbefoundin[9].ETTofalinkisgivenas(etxs=b)whereetxistheexpectednumberoftransmissionsnecessarytosendapacketonthelink;sisthesizeofthepacket(setto1024bytesinourimplementation);andbisthebandwidthofthelink.Fortheetxmetric,wesetthebroadcastprobesizeto1024bytesandproberatetoonesecond.etxiscalculatedevery10seconds.Thelinkdatarate,b,isdeterminedusingpacketpairprobing[15].Thepacketpairisofsize134bytesand1200bytesrespectivelyandissenteveryminute.Wechoosetheminimumoftensamplestoestimateb.D.ChannelAssignmentProtocolChannelassignmentisdifferentfornon-defaultradiosanddefaultradios.Assignmentofnon-defaultradioscanhappenwithoutowdisruptionbyinvokinglinkredirection.Toper-formtheassignment,theCASsimplysendsamessagetotherouterwhoseradiosneedtoberecongured.TheCASwaitsforapositiveacknowledgmentfromtherouterforasmalltimeinterval(5secondsinourimplementation).Iftherouterfailstoacknowledgethemessage,theCASsendsthemessageup-tovetimesinourimplementation.Assoonasarouterreceivesthechannelassignmentmessage,itinvokeslinkredirectionbyfollowingtheprotocoldescribedinSectionVI-B.Ittheninitiatescongurationofitsradios.Aftertheassignmentcompletesandtheneighboringradioonthelinkisconrmedtobeactive,redirectionstops.Inourimplementation,weusethepingutilitytoconrmthatthelinkisactive.Inordertoassignthedefaultradiochannel,weassumetheavailabilityofareliablebroadcastprotocol[18]thatcanbeutilizedtonotifyallmeshroutersaboutthenewassignment.However,duringcongurationofthedefaultradio,thereisnoguaranteethatexistingowswillnotbedisrupted.Ifowscanreachadestinationentirelyusingnon-defaultradios,thoseowswillnotbedisrupted.Flowsthatmustusethedefaultradiochannel,ontheotherhand,willbedisrupteduntilthecongurationcompletes.However,weexpecttheassignmentofthedefaultradiochanneltooccurlessfrequentlythanforthenon-defaultradios.ThisisbecausetheinterferencelevelofthecurrentlyuseddefaultchannelhastoincreasethroughoutthelargedeploymentareaofthemeshnetworkfortheaverageinterferencemetricpresentedinSectionV-Btoincreaseenoughtorequirere-assignment.VII.SIMULATIONSTheobjectiveofthesimulation-basedevaluationistounder-standthebehaviorofourBFS-CAalgorithmandprotocolinlarge-scalenetworks.Weconsiderfourlarge-scaletopologies.Wevarythetrafcpatternsandtheamountofexternalinterfer-enceinthreedifferentnetworkscenarios.Inourevaluations,wecompareBFS-CAagainstastaticchannelassignment.Inthissection,werstdescribethesimulationenvironment.Wethenpresentthefournetworktopologiesfollowedbyadescriptionofthethreescenariosusedinourevaluations.Weendthissectionbypresentingtheresults.A.SimulationEnvironmentWeuseQualnetforoursimulations.WeutilizetheOp-timizedLinkStateRouting(OLSR)protocol[8]forrout-ingwithinthemeshnetworkandtheWeightedCumulative ExpectedTransmissionTime(WCETT)metric[9]forrouteselection.OLSRusesanoptimizedoodingmechanismtooodlinkstateinformationinthenetwork.WCETTisaroutingmetricdesignedformulti-radio,multi-hopwirelessnetworks.Themetriccomputesanestimateofthetimetakentotransmitapacketonapathbasedontheestimatedbandwidthandreliabilityofindividuallinksonthepathandthefrequencydiversicationofthepath.WCETTreliesontheETTmetricintroducedinSectionVItomeasuretheexpecteddelayonapath.TosupporttheETTmetric,weimplementedpacketpairprobingandbroadcastoodingasdescribedinSectionVI-C.intheWCETTmetriccanbechangedtovarytheweightgiventothepathdelayandfrequencydiversicationparametersinthemetric.Wesetto0:5whichresultsinagoodtradeoffbetweenthetwoparametersduringpathselection[9].OLSRlinkstateinformationdisseminatedinthemeshisoverloadedtocontainlinkdelay(ETT)valuesandthechannelnumbersonwhichlinksoperate.WealsomodiedOLSRtousesourceroutingbecauseitsimpliesWCETTsupportinOLSR.WesettheTC REDUNDANCYprotocolparameterinOLSRinordertoincreasethenumberofpathsavailableinthemeshnetwork.InordertosupporttheRFMonmodeinoursimulations,aradioinsnifngmodeispreventedfromtransmittinganypackets.Wealsosimulatethelayer2beaconingofIEEEmanagementframesthroughtheimplementationofabea-coningmodulethattransmitsamanagementframeevery100milliseconds.AllradiosareIEEE802.11aradios,support12channels,anduseauto-ratefallback.Thetwo-raypropagationmodelisusedwithRayleighfading.Thetransmissionpowerforthe802.11aratesissetto18dBm.RTS/CTSisdisabled.Tomakeoursimulationsasaccurateaspossible,weimple-mentinterferenceestimation,linkredirection,andthechannelassignmentprotocolasdescribedinSectionVI.Interfaceestimationoccursevery5minutes.Thetimespentestimatinginterferenceoneachchannelissetto3seconds.Therefore,thetotaltimespentbyaradioinRFMonmodeisequalto36seconds(12times3).TheCASinvokestheBFS-CAalgorithmevery10minutes.AllnodesinoursimulationaresynchronizedintimealthoughthisisnotrequiredforBFS-CAtooperatecorrectly.Beforedescribingtheresultsfromoursetofsimulations,werstdescriberesultsfromasimpletopologytovalidatethecorrectoperationofouralgorithm.Thetopologyisa“linear”topologyconsistingoffournodes.Node1isequippedwithoneradio,nodes2and4withtworadios,andnode3withthreeradios.Node1sendsan8MbpsCBRtrafcstreamconsistingof1024bytepacketstonode4thatstartsat30secondsandcontinuesuntiltheendofthetestat1760seconds.Node4isdesignatedastheCAS.Atthestartofthesimulation,allthedefaultradiosareconguredtooperateonacommonchannel.However,thenon-defaultradiosareeachconguredtooperateondifferentchannels.Therefore,theonlypathfromnode1tothegatewayatthestartofthesimulationisonthedefaultmesh.Therst 0 300 600 900 1200 1500 1760 0 200 400 600 800 1000 1200 Time (seconds)Packets Received Data Interference A A S Fig.4.CBRThroughputinafournodenetwork.“S”indicatesperiodicinterferenceestimationbythenon-defaultradiosateachnode.“A”indicateschannelassignmentbytheCAS.channelassignmentoccursat600s.Justbefore900s,alowerbitratestreamisstartedbetweentwointerferingnodes.Theinterferingnodesareco-locatedwithnodes3and4andoperateonthesamechannelasusedbythelinkbetween3and4.Figure4showsthethroughputactivityatthedestinationintermsofnumberofpacketsreceivedpersecond.Before600s,node4receivesjustmorethan200packetspersecondfromnode1.However,at600s,theCASassignstheradiosatnodes2,3,and4tonon-overlappingchannels.Consequently,thenumberofpacketsreceivedatnode4increasedtoabout1000packetspersecond.Atthetimeindicatedby“S”inthegure,thenon-defaultradiosperforminterferenceestimation.Therefore,thenumberofpacketsreceivedatnode4attheindicatedtimesdropsfromabout1000packetspersecondtoalittlemorethan200packetspersecond.Notethatthenumberofpacketsdeliveredpersecondneverdropstozero.ThisisbecausethenodesimplementlinkredirectionasdescribedinSectionVI-B.Justbefore900s,theinterferingstreamisintroduced.Consequently,thenumberofpacketsreceivedatnode4decreases.Thenon-defaultradiosatnodes3and4detecttheinterferenceat1018sandsendtheirinterferenceestimatestotheCAS.At1200s,theCASre-assignsthechannelofthelink.Thegureillustratestheresultingimprovementinthroughput.B.NetworkTopologiesWenowdescribethefourtopologiesusedinourlarge-scaleevaluation.Eachtopologyconsistsof30routersdistributedinaterrainof500x500meters.Ourchoicesforthenumberofroutersandtheterrainsizearetypicalinlarge-scale,real-worlddeployments[5].Intopologies1and2,thephysicalterrainisdividedintoanumberofcells.Withineachcell,arouterisplacedrandomly.Ingeneratingthetwotopologies,weuseddifferentseedvalues.Thetwotopologiesreectreal-worlddeploymentswheremeshroutersareuniformlydistributedformaximumcoverage.Topology3isagridtopologywheretheinter-routerspacingis75meters.WeconsideragridtopologyinordertoevaluateBFS-CAinadenselypopulatedtopology.Fortopology4,wechoosearandomlygeneratedtopologytoevaluateBFS-CAperformanceinanunplanneddeploymentofrouters. Therouterapproximatelyinthecenterofeachtopologyisdesignatedasthegateway.Thenumberofradiosperrouterischosensuchthatroutersclosetothegatewayareequippedwithmoreradiosthanonesfartheraway.Thegatewaynodeisequippedwithfourradios.Threerouters,chosenfromtheroutersonehopawayfromthegateway,areeachalsoequippedwithfourradios.Sixrouters,chosenfromtherouterstwohopsawayfromthegateway,areeachequippedwiththreeradios.Sixmoreroutersatthreehopsfromthegatewayareequippedwithtworadios.Remainingroutersaresingle-radiorouters.Theselectionofmulti-radioroutersisdonemanuallyinordertocomplywiththerouterplacementstrategydiscussedinSectionIII.C.NetworkScenariosWeconsiderthreenetworkscenariosinourevaluation.Intherstscenario,weevaluatethethroughputimprovementobtainedbyutilizingmulti-radioroutersinsteadofsingle-radiorouters.TenrandomlychosenroutersattheperipherysenddataintwominuteFTPtransferstothegateway.Thescenariolastsforfortyminutes.Therefore,multiplechannelassignmentsoccurinthenetwork.Thisscenarioisdesignedtobean“ideal”scenarioinwhich:(1)thereisnointer-owinterferenceinthenetwork;and(2)thereisnointerferencefromexternalnetworks.Tosatisfytherstrequirement,eachsourcebeginstransmission30secondsafterthepreviousonehasstopped;therstsourcestarts620secondsintothesimulationtoallowfortherstchannelassignmenttooccur.Tosatisfythesecondrequirement,wedonotco-locateanyexternalnodes.Forscenario2,weagainconsideranetworksettinginwhichtherethereisnoexternalinterference.However,unlikescenario1,weletmultipleowswithinthemeshinterferewitheachother.Thetrafcisgeneratedbythesamesetofsourcesaschosenintherstscenario.Furthermore,thesourcesstartatthesametimesasinscenario1.However,theydonotstoptransmittinguntiltheendofthesimulation(fortyminutes).Forscenario3,weconsiderageneralnetworksettinginwhichthereisinter-owinterferenceaswellasinterferencefromexternalnetworks.Tocreatethisscenario,wetakescenario2andintroduceinterferingnodesinmultiplesof4uptoatotalof28interferingnodesineachofthefourtopologies.Thenodesareorganizedintosender-receiverpairs.Eachpairisrandomlydistributedintheterrainspace.Asenderfromeachpairtransmitsa8MbpsCBRstreamconsistingof1024bytepacketsin10minuteburststothereceiverinthepair.Theintervalbetweeneachburstis50seconds.Thenodesstarttransmittingat630s.Foreach10minuteburst,thenodepairsrandomlyselectachannelforcommunication.Inthescenariospresentedabove,BFS-CAiscomparedagainstastaticassignmentofchannels.Inthestaticcase,calledthe“StaticMulti-Radio”scheme,thedefaultradiosoperateononechannel.Therstnon-defaultradioonallmulti-radioroutersistunedtoanon-overlappingchannel.Oftheremainingroutersthatstillhaveunassignedradios,thesecondnon-defaultradioistunedtoasecondnon-overlappingchannel.Finally,thegatewayandthemulti-radioroutersclosesttothegatewaytunetheirthirdnon-defaultradiotoathirdnon-overlappingchannel.D.ResultsInscenario1,weevaluatethethroughputimprovementgainedbyutilizingmulti-radioroutersinsteadofsingle-radioroutersforeachofthenetworktopologies.Figure5(a)plotsthemeanthroughputofthetenFTPtransfersinasingle-radiomeshnetwork,inamulti-radiomeshnetworkwithBFS-CA,andinamulti-radiomeshnetworkwithstaticassignment.Thethroughputimprovementwithmultipleradiosisgreaterthan200%fortopology1andgreaterthan100%fortopolo-gies2and3.Thethroughputimprovementfortopology4,however,isonlyabout33%withtheBFS-CAschemeandapproximately54%withthestaticmulti-radioscheme.Intopology4,themulti-radioroutersarelesslikelytobeonpathstothegatewaybecausetheyarerandomlydistributedintheterrainspace.Morenotably,Figure5(a)indicatesthatintopolo-gies2,3,and4,thestaticmulti-radioschemeperformsbetterthanBFS-CAbyabout8%,5%,and15%respectively.ThisisbecauseBFS-CAimprovesthroughputbytuningindividuallinkstonon-overlappingfrequencies.Becauseofthisstrategy,theopportunitytondchannel-diversiedpathsislessthanwiththestaticschemewhereregionsofthemeshnetworkaretunedtonon-overlappingfrequencies.Therefore,withthestaticscheme,theopportunitytochoosechannel-diversiedpathsisbetter.Asaresult,thestaticschemeperformsbetter.Inordertoverifytheabovereasoning,wecomputedtheChannelDiversityExtent(CDE)ofallowswiththetwoschemes.TheCDEofaowisobtainedbytakingtheaverageoftheCDEsofallpathstraversedfortheow;aowcantraversemultiplepathsbecauseofvariationsinalink'sETTvalue,whichcanresultinpathchanges.TheCDEofapathisdenedastheratioofthenumberofchannelsusedinthepathtoitshop-count.Forexample,ifa4hoppathmakesuseof4channels,itsCDEis1.Ontheotherhand,ifthepathmakesuseofonly1channel,itsCDEis0.25.ApathwithahighCDEisgenerallypreferredoverapathwithalowCDE.ATCPowisbi-directionalduetoacknowledgmentsandisconsideredastwoowsinourcalculations.Figure5(b)isascatterplotoftheCDEsofallowsforthetwomulti-radioschemes.ThexvalueofeachpointintheplotistheCDEofapathwithBFS-CA,andtheyvalueistheCDEofthesamepathwiththestaticscheme.Thelinex=yindicatespathswithequalCDEvalues.Amajorityofthepointsintheplotareabovethex=yline.ThesepointsindicatepathsthatwiththestaticschemehavehigherCDEvaluesthanwiththeBFS-CAscheme.Consequently,pathswiththestaticschemearemorechanneldiversiedthanwiththeBFS-CAscheme.Figure5(c)isaplotoftheaveragehop-countsforthetwomulti-radioschemesinscenario1forthefourtopologies.Thehop-countwithBFS-CAishigherthanwiththestaticscheme.ThisisbecauseWCETT,theroutingmetricinourevaluations, 1 2 3 4 0 500 1000 1500 TopologyMean Throughput (Kbps) Single Radio BFS-CA Static Multi-Radio (a)MeanThroughput 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 BFS-CAMulti-Radio Static (b)ScatterplotofCDE 1 2 3 4 0 1 2 3 4 TopologyAvg Hop Count BFS-CA Static Multi-Radio (c)Averagehop-countFig.5.Resultsfromscenario1wheretheBFS-CAandstaticschemesarecomparedagainstasingle-radioscheme.Scenario1hasnointer-owinterferenceandnoexternalinterference. 0 1 2 3 4 0 200 400 600 800 TopologyMean Throughput (Kbps) BFS-CA Static Multi-Radio (a)Meanthroughput 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 BFS-CAStatic Multi-Radio (b)ScatterplotofCDE 1 2 3 4 0 1 2 3 4 TopologyAvg Hop Count BFS-CA Static Multi-Radio (c)Averagehop-countFig.6.Resultsfromscenario2whereBFS-CAiscomparedagainstthestaticschemeinthepresenceofinter-owinterferencebutwithoutexternalinterference.prefersafrequencydiversiedpathoveralternativesthatmaybeshorter.Becauseoftheassignmentstrategiesofthetwoschemes,WCETTpickslongerfrequencydiversiedpathswithBFS-CAthanwiththestaticscheme.TheabovenotedperformanceimprovementofthestaticschemeoverBFS-CAisachievedonlyinanidealscenariowheresourcessendtrafctothegatewayoneatatime.Becauseoftheassignmentstrategyinthestaticscheme,whenmultiplesourcestransmit,owstothegatewayeachtraverselinkstunedtothesamechannels,resultinginincreasedin-terferencebetweentheows.Consequently,thethroughputimprovementwiththestaticschemeisexpectedtobelessthanwiththeBFS-CAscheme.Thisintuitionisveriedinscenario2whereinmultiplesourcestransmittothegatewayatthesametime.Figure6(a)plotsthemeanthroughputobtainedwiththetwomulti-radioschemes.Clearly,BFS-CAoutperformsthestaticscheme.Specically,thethroughputimprovementisashighas72%withtopology2and48%,60%,and13%withtopolo-gies1,3,and4,respectively.Figure6(b)showstheCDEsoftheowsforthetwoschemes.Pointslieinapproximatelyequalnumbersaboveandbelowthex=yline.Althoughtheowsinthetwoschemesareequallychanneldiversied,thethroughputwithBFS-CAishigherbecauseofreducedinter-owinterference.Figure6(c)plotstheaveragehop-countwiththetwoschemes.Theaveragehop-countwithBFS-CAisslightlyhigherthanwiththestaticscheme.ThisisbecauseoftheassignmentstrategyusedbyBFS-CA.However,withBFS-CA,theowstraversepathsthatresultinlessinter-owinterferencethanwiththestaticscheme.Therefore,themeanthroughputwithBFS-CAisgreater.Inthethirdscenario,weconsidertheeffectofvaryingamountsofinterferenceontheperformanceofthemeshtopologiesforthetwomulti-radioschemes.Figure7showsthepercentagedifferenceinthemeanthroughputofthetenFTPtransfersachievedbythetwoschemes.Inamajorityofthecases,BFS-CAoutperformsthestaticscheme.Asanaverageoftheperformancedifferenceswithvaryingthenumberofinterferingradios,BFS-CAperformsbetterthanthestaticschemeby42.14%,31.14%,15.75%,and11.85%fortopologies1,2,3,and4respectively.BFS-CA'simprovedperformanceoverthestaticschemeoccursbecausethemeshroutersinthenetworktopologiesareabletodetectthein-creasedinterferenceandarethereforere-assignedbytheCAStochannelsdifferentthanonesusedbytheinterferingnodes.Inthecaseoftopology4,theunplanneddistributionofmeshroutersintheterrainspaceresultsinBFS-CAperformingonlyslightlybetterthanthestaticscheme.MorenoteworthyisBFS-CA'sperformanceintopology3(thegridtopology)inwhichthethroughputgainswithBFS-CAisonlymarginallybetterthanthetopology4case.Thisisbecauseintopology3,themeshroutersandinterferingnetworksaredenselypopulatedintheterrainspace.Consequently,theinterferencefoot-printoftheinterferingnetworksisgreaterinthedenseenvironment.Asaresult,re-assignmentofchannelsyieldsonlyminimalthroughputimprovementwithBFS-CA. 4 8 12 16 20 24 28 -100 -50 0 50 100 Percentage Improvement 4 8 12 16 20 24 28 -100 -50 0 50 100 (a) Topology 1BFS-CABFS-CAStatic Multi-Radio(b) Topology 2(c) Topology 3(d) Topology 4Interfering RadiosInterfering RadiosInterfering RadiosInterfering Radios 4 8 12 16 20 24 28 -100 -50 0 50 100 BFS-CAStatic Multi-Radio 4 8 12 16 20 24 28 -100 -50 0 50 100 Static Multi-Radio Static Multi-Radio BFS-CA BFS-CA Fig.7.PercentageimprovementinthroughputwithBFS-CArelativetoStaticMulti-Radiointhepresenceofinterference.Thenumberofinterferingradiosisshownonthex-axes.VIII.PROTOTYPEIMPLEMENTATIONOurprimarymotivationinimplementingaprototypeistodemonstratethepracticalityofourproposedBFS-CAalgo-rithmandprotocol.WeevaluatetheimplementationinasixnodeIEEE802.11btestbedwiththeLinux2.4.26operatingsystem.TheprototypeimplementationconsistsofaroutermoduleandaCASmodule.TheroutermoduleisinstalledoneachmeshrouterandimplementstheinterferenceestimationandlinkredirectionproceduresasdescribedinSectionVI.ToperformRFMonsnifng,itinvokesLinuxutilitiessuchastheiwcongwirelessutilitiesandthetetherealpacketcapturetool.TheCASmoduleimplementstheBFS-CAalgorithm.Interferenceestimationoccursevery5minutes,andthechan-nelselectionprocedureisinvokedevery10minutesasinthesimulations.BecauseIEEE802.11bsupportsonlythreenon-overlappingchannels(1,6,and11),theperformanceimprovementob-tainedusingtheproposedsolutioninourtestbedsettingisexpectedtobelimited.WeconsideredusingAtherosbasedIEEE802.11aradios.However,ourtestbednodesuseLinuxandad-hocmodesupportinLinuxdriversforIEEE802.11aradiosiscurrentlyfaulty[3].AllnodesinthetopologyareIBMR32Thinkpadlaptops.ThesixnodesareeachequippedwithoneNetgateNL-2511CDPCMCIAradio.NodesBandCareeachequippedwithanadditionalLinksysWUSB12USBradio.WeuseUSBextendercablesfortheLinksysradios.NodesA,B,C,andDformastaticmulti-hopwirelessnetworkinasimplelinetopology,i.e.,AcannotseeCandD,BcannotseeD,andvice-versa.NodesEandFareinterferingnodespositionedroughlyameterawayfromBandCrespectively.NodesEandFuseadifferentESSIDandnetworkIPaddressthanA,B,C,andD.Allradiosareconguredtooperateat11Mbps.RTS/CTSisdisabled.Fortheentiredurationoftheexperiment,thedefaultmeshissettochannel1.Wedidnotallowthere-assignmentofthedefaultmeshchannelduringourexperimentsbecauseourWUSB12radiosdonotsupportdynamicrecongurationinad-hocmode.NodeBisdesignatedastheCAS.TheCASinformsnodesA,C,andDabouttheIPaddressrangethatbelongstothemeshnetwork.Atthestartoftheexperiment,theradiosonEandFandthenon-defaultradiosonBandCaretunedtochannel6.Intheexperiment,Asends1024byteUDPpacketsasfastaspossibletosaturatethepathtoDforadurationof40minutes.TheinterferingnodeEsends1024byteUDPpacketsasfastaspossibletosaturatethelinktonodeFstartingat250suntiltheendoftheexperiment.At1400s,wemanuallyswitchnodesEandFfromchannel6tochannel11inordertomeasuretheresponseofouralgorithmimplementationtoavaryinginterference.Figure8illustratesthethroughputintermsofnumberofpacketsreceivedatDandF.Fromthegure,itcanbeclearlyseenthatthethroughputofthestreamfromAtoDisnotaffectedbytheinterferingstream.ThereasonisthatthelinksA-BandC-Dareininterferingrangeofeachother.Consequently,theselinksformthebottleneck.Asaresult,linkB-CisabletoconsistentlysustainthelownumberofpacketsinjectedbyAinspiteofinterferencefromlinkE-F.Thedropinthroughputevery300secondsisbecauseoftheperiodicinterferenceestimationperformedbythenon-defaultradioonBandC.Theestimationoccursfor9seconds(3secondstimes3non-overlappingchannels).Thethroughputneverdropstozerobecauselinkredirectionisperformedoverthedefaultradiosuntiltheinterferenceestimationcompletes.AtthetimesindicatedbySinthegure,thenon-defaultradiosatBandCdetecttheinterferenceandnotifytheCAS.AtthetimesindicatedbyA,thenon-defaultradiosarere-assignedtoachannelnotusedbytheinterferingstream.Theresultofthere-assignmentcanbeclearlyseeninthethroughputplotoftheinterferingnetwork.Aftertherstre-assignment,thelinkB-CswitchesfromChannel6to11.Consequently,thethroughputoftheinterferingstreamincreases.At1400s,linkE-Fismanuallyassignedtochannel11.Asaresult,thethroughputdrops.ItincreasesagainwhenlinkB-CswitchesfromChannel11backtoChannel6duetothesecondre-assignment.IX.RELATEDWORKThereexistsalargenumberofstudiesthataimtoaddressthecapacityprobleminwirelessmeshnetworks.Wesumma-rizearepresentativesamplebelow.SeveralproposalsfocusonimprovingtheIEEE802.11MACprotocoltosupportmultiplechannels[7],[11],[23].Thekeyadvantageofsuchschemesisthatonlyasingleradioisrequiredtosupportmultiplechannels.Thedisadvantage,however,isthattheyrequirechangestotheMAClayerandthe 0 500 1000 1500 2000 0 200 400 600 Time (seconds)Packets Received S S A A Interfering Stream Mesh Stream Fig.8.Throughputinthepresenceofaninterferingstream.hardwareinordertosupportper-packetchannelswitching.Tothebestofourknowledge,suchhardwareisstillnotavailable.Raniwalaetal.[19],[20]proposeaload-awarealgorithmtodynamicallyassignchannelsinamulti-radiomeshnetwork.Raniwala'sproposalrequiresthatanticipatedtrafcloadsandpathstraversedbyowsbeknownbeforechannelassignmenttakesplace.Thekeydistinguishingaspectinourproposalisthatweassignchannelsbasedentirelyonknowledgeofinterferenceinthemeshnetwork.Kyasanuretal.proposeahybridchannelassignmentsolutionformulti-radiowirelessadhocnetworks[16].Inthisscheme,asubsetofradiosonarouterarestaticallyassignedchannels.Theremainingradiosdynamicallyswitchtothestaticchannelsassignedtoneighboringroutersinordertocommunicatewiththem.Theschemerequiresthatradioscanswitchbetweenchannelsonaper-packetbasis.Tothebestofourknowledge,suchradiosarenotavailableinthemarketplace.Somemeshhardwarevendors[2],[4]offermulti-radiomeshroutersthatutilizeproprietarychannelassignmentschemes.Therefore,weareunabletoprovideafaircompari-son.X.CONCLUSIONSMulti-radiorouterscansignicantlyimprovetheperfor-manceofwirelessmeshnetworks.However,anystaticassign-mentofchannelstothemeshradioscandegradenetworkper-formancebecauseofinterferencefromco-locatedwirelessnet-works.ThispaperpresentedBFS-CA,adynamic,interference-awarechannelassignmentalgorithmandcorrespondingproto-colformulti-radiowirelessmeshnetworks.BFS-CAimprovestheperformanceofwirelessmeshnetworksbyminimizinginterferencebetweenroutersinthemeshnetworkandbetweenthemeshnetworkandco-locatedwirelessnetworks.Theproposedsolutionispracticalandeasilyimplementable.WendthatBFS-CAresultsinsignicantperformanceimprove-mentsinthepresenceofvaryinginterferencelevels,whicharevalidatedthroughempiricalmeasurementsonatestbed.Asfuturework,weplantoevaluateBFS-CAontheUCSBMeshNet[6],athirtynodemulti-radiowirelessmeshtestbedatUCSB.REFERENCES[1]Aeroex6970RFPowerMeters.http://www.aeroex.com.[2]BelAirNetworks.http://www.belairnetworks.com.[3]MADWIFIProject.http://sourceforge.net/projects/madwi/.[4]MeshDynamics.http://www.meshdynamics.com.[5]Metro-ScaleWiFiAsCityServicechaska.net,Chaska,Minnesota.http://www.tropos.com/pdf/chaska casestudy.pdf.[6]UCSBMeshNetProject.http://moment.cs.ucsb.edu/meshnet.[7]P.Bahl,R.Chandra,andJ.Dunagan.SSCH:SlottedSeededChannelHoppingForCapacityImprovementinIEEE802.11AdHocWirelessNetworks.InACMMobiCom,Philadelphia,PA,September2004.[8]T.ClausenandP.Jacquet.OptimizedLinkStateRoutingProtocol.InternetEngineeringTaskForce,RFC3626,October2003.[9]R.Draves,J.Padhye,andB.Zill.RoutinginMulti-radio,Multi-hopWirelessMeshNetworks.InACMMobiCom,Philadelphia,PA,September2004.[10]P.GuptaandP.Kumar.CapacityofWirelessNetworks.InIEEETransactionsonInformationTheory,volume46,pages388–404,March2000.[11]N.Jain,S.Das,andA.Nasipuri.AMultichannelCSMAMACProtocolwithReceiver-BasedChannelSelectionforMultihopWirelessNetworks.InIEEEInternationalConferenceonComputerCommunicationsandNetworks,Scottsdale,AZ,October2001.[12]K.Jamieson,B.Hull,A.Miu,andH.Balakrishnan.UnderstandingtheReal-WorldPerformanceofCarrierSense.InACMSigcommWorkshoponExperimentalApproachestoWirelessNetworkDesignandAnalysis,Philadelphia,PA,August2005.[13]A.Jardosh,K.Ramachandran,K.Almeroth,andE.Belding.Un-derstandingCongestioninIEEE802.11bWirelessNetworks.InACM/USENIXInternationalMeasurementConference,Berkeley,CA,October2005.[14]J.KatzelaandM.Naghshineh.ChannelAssigmentSchemesforCellularMobileTelecommunicationsSystems:AComprehensiveSurvey.InIEEEPersonalCommunications,volume3,pages10–31,June1996.[15]S.Keshav.AControl-TheoreticApproachtoFlowControl.InACMSigcomm,Zurich,Switzerland,September1991.[16]P.KyasanurandN.Vaidya.RoutingandInterfaceAssignmentinMulti-ChannelMulti-InterfaceWirelessNetworks.InIEEEWirelessCommunicationsandNetworkingConference,NewOrleans,LA,March2005.[17]J.Padhye,S.Agarwal,V.Padmanabhan,L.Qiu,A.Rao,andB.Zill.EstimationofLinkInterferenceinStaticMulti-hopWirelessNetworks.InACM/USENIXInternationalMeasurementConference,Berkeley,CA,October2005.[18]E.PaganiandG.Rossi.ReliableBroadcastinMobileMultihopPacketNetworks.InACMMobiCom,Budapest,Hungary,September1997.[19]A.RaniwalaandT.Chiueh.ArchitectureandAlgorithmsforanIEEE802.11-basedMulti-ChannelWirelessMeshNetwork.InIEEEInfocom,Miami,FL,March2005.[20]A.Raniwala,K.Gopalan,andT.Chiueh.CentralizedChannelAssign-mentandRoutingAlgorithmsforMulti-ChannelWirelessMeshNet-works.InMobileComputingandCommunicationsReview,volume8,pages50–65,April2004.[21]K.Rosen.DiscreteMathematicsanditsApplications.McGrawHill,1999.[22]K.Sanzgiri,I.Chakeres,andE.Belding-Royer.TheUtilityofPerceptiveCommunicationbetweenDistantWirelessNodes.InIEEE/Create-NetTridentCom,Barcelona,Spain,March2006.[23]J.SoandN.H.Vaidya.Multi-ChannelMACforAdHocNetworks:HandlingMulti-ChannelHiddenTerminalsUsingaSingleTransceiver.InACMMobiHoc,Tokyo,Japan,May2004.[24]A.Vasan,R.Ramjee,andT.Woo.ECHOS-EnhancedCapacity802.11Hotspots.InIEEEInfocom,Miami,FL,March2005.[25]S.Vasudevan,J.Kurose,andD.Towsley.DesignandAnalysisofaLeaderElectionAlgorithmforMobileAdHocNetworks.InIEEEInternationalConferenceonNetworkProtocols,Berlin,Germany,October2004.