parameterstvants官网

CharacterizationofsignalingandtrafficinJoostMajedAlhaisoni&AntonioLiottaReceived:17April2008/Accepted:7October2008/Publishedonline:18December2008#TheAuthor(s)2008.
ThisarticleispublishedwithopenaccessatSpringerlink.
comAbstractPeer-to-Peer(P2P)IPTVapplicationshaveincreasinglybeenconsideredasapotentialapproachtoonlinebroadcasting.
Recently,manyapplicationssuchasPPlive,PPStream,andSopcasthavebeendeployedtodeliverlivestreamingviaP2P.
OneofthelatestsystemsisJoost,whichcandeliverbothVideo-on-DemandandReal-Timeservices.
Measuringandcharacterizingthisapplica-tionintermsofsignalingoverheadsandtrafficprofileshelpstobetterunderstandthekeylimitationsofcurrentP2PIPTVsystems.
Therefore,themainpurposeofthispaperisfirstlytostudytheimpactofJoostonthenetwork.
Secondly,wewishtodeterminetheunderlyingmecha-nismsofJoost,distinguishingbetweentheVideo-on-DemandandtheReal-timeservices.
OurstudyiscarriedoutthroughacloseinvestigationandanalysisonthetrafficofJoostintwotypesofstreaming.
Baseduponthedatatracingandcollection,manydifferentstatisticshavebeenderived.
Ourstudyunveilsstrengths(e.
g.
goodresiliencetoend-to-enddelayandjitter)andshortcomings(e.
g.
poorlocality)andyieldsrecommendationsforfutureP2PIPTVsystems.
KeywordsP2Pstreaming.
Trafficcharacterisation.
Joost1IntroductionPeer-to-Peer(P2P)streamingrepresentsaneconomical,robust,andscalablealternativetothemoreconventionalclient–server(CS)approach[11].
Thebasicideaisthat,ratherthanstreamingmediafromdedicatedservers,anapplication-leveloverlayisformedbytheuserterminals,whichcooperateinthedistributionofthestreamitself[11].
Whilereceivingastream,terminalssimultaneouslyactasdistributionhubsforit.
Inthisway,thebottlenecksandfailurepointstraditionallyassociatedwithserversarevirtuallyeliminated,sinceindividualserversarereplacedbyamultitudeofuserterminals.
Also,asthenumberofusersconnectingtoacertainstreamgrows,soisthenumberofdistributionpoints.
HencethesystemscalesmuchbetterthananyCScounterpart[11].
TheP2PstreamingconcepthasnowleadtoanumberoftrialP2PIPTVsystemssuchasPPlive[1],Joost[2],PPStream[3],andSopcast[4].
Thereisnowclearcommercialinterestinthesenewtechnologieswhicharerevolutionizingtheonlinebroadcastingarena.
DespitethenumerousadvantagesofP2PstreamingingeneralandP2PIPTVinparticular,theircharacteristicsintermsofsignalingoverheadsandnetworkefficiencyarenotwellknown.
Mostsystemsmakeuseofproprietaryprotocolsandarenotopentotheresearchcommunity.
Thismakesitvirtuallyimpossibletocarryoutsimulationstudiesaimedatdeterminingwhetheralarge-scaledeploy-mentofP2Pstreamingwillbesustainableintermsofnetworkresources,operation,andmanagement.
Ontheotherhand,trialplatformsarealreadyinuse,whichcreatestheopportunitytogetabetterinsightintomechanismsofP2Pandtheireffectsonthenetwork.
Ourworkpursuesthisavenue.
WelookatsignalingandtrafficcharacteristicsinJoost,oneofthemostrecentapplicationsPeer-to-PeerNetwAppl(2009)2:75–83DOI10.
1007/s12083-008-0015-5M.
AlhaisoniUniversityofEssex,WivenhoePark,Colchester,CO43SQ,UKe-mail:malhai@essex.
ac.
ukA.
Liotta(*)DepartmentofElectricalEngineering,EindhovenUniversityofTechnology,DenDolech2,P.
O.
Box513,PT11.
29,5600MBEindhoven,TheNetherlandse-mail:A.
Liotta@tue.
nlsupportingbothVideo-on-Demand(VoD)andReal-Time(RT)streamingservices.
Datatracinganddeeppacketinspectionunveiledawealthofinterestingproperties,someofwhichwerenotentirelyexpected.
Ourstudyrevealsstrength(e.
g.
goodresiliencetoend-to-enddelayandjitter)andshortcomings(e.
g.
poorlocality)andyieldsrecom-mendationsforfutureP2PIPTVsystems.
2P2PstreamingarchitecturesPeertoPeerstreamingarchitecturescanbecategorizedbasedupontheirdistributionmechanisms.
ThevariousapproachestoP2PstreaminghavebeensurveyedbyLiuetal.
[11].
Ouranalysisisparticularlyconcernedwithtwomainmethods,whicharebrieflyoutlinedbelow.
2.
1Tree-basedarchitectureInthetree-basedarchitecture,thepeersareorderedhierarchicallybythesource,knownastheparent.
Theparentnode,inturn,sendsdatapacketstointermediatenodes,andthesenodesrelaythemiterativelyuntilleafnodesarereached.
Despiteintroducingagoodlevelofparallelismanddistribution,thisapproachsuffersfromanumberoflimi-tations.
Theroot,ordatasource,isasinglepointoffailure,whichlimitstherobustnessofthesystem.
Anotherproblemisthatifpeersjoinandleavefrequentlythetreehastoberebuilttoooften,whichhasanegativeimpactonsignalingoverheads,latency,andstability.
Anexampleoftree-basedstreamingapplicationisPeercast[12],anopen-sourcesoftwareforstreamingbothaudioandvideo.
ApeculiarityofPeercastisthatanynodecanspecifythemaximumnumberofincomingconnectionsallowed.
2.
2Mesh-basedarchitectureInthisarchitecture,theoverlaynetworksupportingthestreamdistributionisamesh.
Dataisdividedinchucksinsuchawaywhichallowsapeertoreceiveportionsofthestreamfromdifferentpeersandassemblethemlocally.
Thisapproachismorerobustthanthetree-basedarchitec-ture,sincewhenastreamcomesfromvarioussourcescommunicationdoesnotbreakwhenonlyasubsetofpeersdisconnect.
Anotherbenefitisthatthistransportmethodreflectswelltheasynchronousnatureofmanyaccessnetworktechnologies(e.
g.
ADSL).
Infact,peerscandownloadastreamatfullqualitywhilstuploadingonlyafractionofit.
ThesamefeatureisexploitedinP2PfilesharingapplicationBitTorrent[14].
Commonstreamingapplicationsinthemesh-basedcategoryareGNuStream[13],PPlive[5],Coolstreaming[9],andJoost[2].
Thismethodisbecomingmorewidespreadthanthetree-basedone.
ThatiswhyinourstudywehavechosentoscrutinizeJoost.
3RelatedworkVeryfewresearchershavestudiedP2Pstreamingbasedoncommercialplatforms.
ApplicationsthatcanbetrulycategorizedasP2Phaveappearedonlyrecentlyandtheirunderlyingalgorithmsare,inmanycases,proprietaryandnotreadilyavailableforacademicscrutiny.
AstudywhichissimilartotheonepresentedinthisarticlehasbeencarriedoutbyHeietal.
[1,6].
TheycharacterizethetrafficandbehaviorofPPlive,whereaswehavefocusedonJoostwhichismorerecentandprovidesbothVoDandRTstreaming.
OtherstudiesfocusedonlyontheVoDservicearereportedin[7,8],and[9].
Giventoitsoriginalconception,JoosthasseveralsimilaritiestoSkype,whichishoweveraP2Pconferencingtool.
FromtheseveralworksassessingSkypewecanlearnaboutmechanismsthataresupposedlyanalogousinJoost.
Examplesincludethebootstrappingprocess,thesupern-odes'electionalgorithm,andthemanagementoftheoverlaywhennodesjoinordepart[10].
Atthetimeofwriting,onlytwostudiesofJoostareavailableintheformoftechnicalreports.
Leietal.
explainthekeyunderlyingmechanismsofJoost[11].
Leietal.
focusonJoostcomponentsandarchitecture[5].
ThesestudieswereconductedonearlierversionsofJoost(v.
0.
9.
2andv.
1.
0,respectively).
Ourstudyisbasedonv.
1.
1.
4whichismorestableandincludesnewfunctionality,mostnotablyRTstreaming.
Becauseofthis,wehavebeenabletoidentifykeydifferencesbetweenVoDandRT,inadditiontocarryingoutfurtherstatisticsonpackettraces.
AlsowecharacterizebothUDPandTCPtraffic,bothinuploadanddownloadmodes.
Ourworklooksspecificallyatnetworklocality,theabilityofJoosttoselectnodeswhichareclosetoeachother.
Wecharacterizetherelativedistributionofnodesandthemappingbetweenlogicalandphysicaloverlaysatcityandcountrylevels.
TobiasHofeldetal.
[15],havedoneacomparativestudyofpopularP2PIPTVsystemssuchasJoost,PPlive,andZattoo.
TheirstudywasconcernedabouthowtheuserperceivestheseP2PIPTVsystems.
Inaddition,theyhaveindicatedkeyfeaturesoftheseapplications,includingtopologymanagement,distributionprotocols,andband-widthutilization.
Similarlytoourstudy,theyconductedexperimentsbasedonpassivemeasurementsandconsid-eredbothdownloadanduploadtraffic.
Ourstudyconsidersadditionalparameters,studiesbothRTandVoDstreamingandlooksspecificallyatnetworklocalityandgeographicloaddistribution.
76Peer-to-PeerNetwAppl(2009)2:75–83AnotherstudyofP2Pstreamingsystem,focusedonCoolstreaming,isbySusuXieetal.
[16].
Theyusedrealtracesandtriedtodrawsometheoreticalbasistodemon-stratethatselectingpeersrandomlyhasthepossibilitytoscalewell.
TheydevelopedsomefundamentalconceptsaboutCoolstreaming,showinghowproblemsrelatingtoheterogeneitycanbeaddressedbysomeadvancedbuffer-ingtechniques.
TheyalsolookedatimportantperformancefactorsincludingbootstraptimeandtherateofjoinandleaveintheP2Pnetwork.
SimilarlytoXieetal.
,westudyP2PIPTVintermsofchurnandbootstrappingdelay.
However,wealsolookatotherparameterssuchastrafficdistributionandnetworklocality.
WeconsidereduploadanddownloadtrafficincurredbyRTandVoD.
Mostimportantly,ourworkcanbeconsideredcomplimentarytoXie'ssincewedrawconclusionsbasedonrealmeasurements,ratherthanrelyingonsimulations.
AnotherworkthatfocusesontheissueofnetworklocalityisbyAleksandraKovacevicetal.
[17].
Theyconcludethatlocationawarenessdecreasestransmissiondelay,oneofthemostimportantfactorsinmediastreaming.
Despiteworkingontheassessmentofdifferentsystems,ourstudyleadstosimilarconclusionsabouttheimportanceofdesigninglocation-awareoverlaystopursuebandwidthconservation.
AnothereminentworkisbyThomasSilverstonetal.
[18]whohaveperformedameasurementstudyduringthelastFIFAWorldCup,comparingfourP2Pstreamingapplications(PPStream,PPlive,TVAnts,andSopcast).
Theirstudywasconcernedwiththetrafficstatisticsandthechurnofpeers.
Inaddition,theywereconcernedabouttheimpactofthetrafficgeneratedonthenetworkfromtheseapplications.
Finally,theyshowedtheuserbehaviorintothesesystems.
Likewise,wepresentourresultsbasedonpassivemeasurementsoflivestreams.
However,ourworkfocusesonJoostwhichisamorerecentapplicationandsupportsbothRTandVoDservices.
Additionally,westudynetworklocality.
AlexandroSentinellietal.
[19]reportontheirmeasure-mentsbasedonSopcast.
Theyfiguredoutanumberofimportantparametersliketheaveragenumberofpeersthatanodeconnectsto,thetypicalstart-updelay,thecontinuityindexing,andtheamountofdatabufferedattheclientside.
Theirworkisparticularlyvaluableasitisbasedonalarge-scaletestbed,PLANETLAB.
Again,ourworkcanbeconsideredtobecomplementaryaswelookatadditionalfeaturesandstudyJoost.
4Experimentalmethod4.
1OverviewofJoostJoostisapeertopeerstreamingapplicationthatdeliverstelevision-qualityVoDandRTstreamingservicesviaaP2Pnetworkoverlay.
ItwascreatedbythefoundersofSkypeandKaZaAandiscurrentlyinBetaversion(v.
1.
1.
4atthetimeofwriting)[5].
Joostsupportsmorethan15,000TVprogramsthroughmorethan200channels[2].
KeycomponentsandfunctionsofJoostaredescribedbelow.
Atsystemstart-up,Joostperformsanumberofoperations:&PortselectionDuringtheinitialbootstrapping,JoostselectsaspecificporttoconnecttoandcommunicatewithotherpeersviaUDP.
TheInternetAssignedNumbersAuthorityhasrecentlyassignedport4166forthispurpose.
&LocalCacheEachJoostClientstoresallthemediadataas"anthillcache"intheCdrive,usingthefollowingdirectory:C:\Documentsandsettings\ApplicationData\Joost\"anthillcache".
Anthill[20]isanagentbasedsupportingthemediadistributionservices.
Thesizeofthelocalcachedependsonhowandforhowlongtheprogramhasbeenlaunched,soitincreaseswiththesizeoftheprograms.
However,incaseofwatchingthewholechannel,allthemediadatahasbeenstoredinthelocalcache,sointhesecondwatching,thechannelwillberunningfromthelocalcacheinsteadofconnectingtotheserverexceptforsomeCodec.
Inourexperiment,itcachesmorethan2GB.
However,thiswillaffecttheuserresourceswhenmorechannelsarewatched.
&InstallationOneofthefunctionsofJoostistheInstallation,sointhisphasetheJoostclientconnectstotheserversendinganHTTPrequesttoretrievetheavailablechannellistanddownloadtheSQLitefile[21]whichgivestheinitialavailablechannellists.
Moreover,SQLiteisusedformanagingthedatabaseofJoostchannels.
&BootstrappingInJoost,therearethreeserversandtwosupernodes.
Initially,theJC(JoostClient)connectstotheserverlux-www-lo-2.
joost.
netoverHTTP.
Then,JCwillreceivesomeavailablesupernodes.
Afterthat,anHTTPrequestwillbesenttolux-www-lo4.
joost.
nettogettheupdatedversionofthesoftware.
Lastly,theJCwillconnecttothesupernodessuchaslid-snode-1-eth0.
joost.
nettogettheavailablelistofchannelsandtheavailablepeersthatarewatchingthechannels.
Beforethat,JChasalreadystartedcommunicat-ingwithotherserversandpeers.
AschematicviewofJoost'sarchitectureisdepictedinFig.
1.
4.
2ExperimentalsetupOurexperimentswereconductedintheUnitedKingdom.
WecollectedJoostpacketsindifferenttypesofstreamingPeer-to-PeerNetwAppl(2009)2:75–837777mode(VoDandRT).
WeusedthecurrentbetaversionofJoost,v.
1.
1.
4.
Figure2,outlinesthesetupofthetwomachinesusedforthecollectionoftraces,whichwerethenusedforpacketanalysis.
Themachineswereconnectedtoa100MbpsEthernet,connectedtothecampusInternetleasedline.
ThisensuredthatbothinboundandoutboundbandwidthwereconsiderablyhigherthatthenminimumrequiredforthecorrectfunctioningofJoost.
ThespecificationsofthemachineusedtogatherVoDdatawereasfollows:WindowsXPwithIntelcore2CPU6420@2.
13GHz,3.
25GBofRAM.
Thespecificationsofthemachineusedtogatherlivestreamingdatawereasfollows:WindowsXPwithAMDAthlonTM643400+processor(2.
10GHz),1GBofRAM.
EachPCranWiresharkv1.
0.
0(anopensourcenetworkprotocolanalyzerwhichisknownasEthereal)andNetpeekerv3.
10,anetworkmonitorusedtocaptureallinboundandoutboundtrafficincurredbyJoost.
4.
3DatacollectionVoDtraceswerecollectedfromoneofthemostpopularJoostchannelsi.
e.
,selectedfromthe"what'spopular"section.
Tracesincludedalleventsbasedona2-hobser-vationwindow.
Theoverallsizeofourtracefileswas277Mb.
RTtraceswerebasedonthefourlivechannelswhichbroadcastsportclips.
Ourtracesincludealleventsbasedona1-hobservationwindowrelatingtoasinglelivechannel.
Broadcastingiscurrentlyfairlylimitedtojustafewhoursperday,whichlimitedtheamountofdatathatcouldbecollectedwithinourexperimentaltimeframe.
Theoverallsizeofourtracefileswas104Mb.
Datatraceswerefilteredatdifferentgranularitiesi.
e.
,atbyte,packet,andsessionlevels.
WedistinguishedbetweenUDP,TCP,upload,download,data(VoDandRT),andsignalingpackets.
Signalingpacketsweredifferentiatedfromdatapacketssincetheformerareintheorderof60–70kb,whereasthelatterarebetween800and1,000bytes.
5Trafficcharacterizationandunderlyingmechanisms5.
1Start-updelayStart-updelayisoneoftheimportantfactorsinonlinebroadcasting.
Highvalueswillgivealowuserexperience,TCPDUMPTCPDUMPFig.
2ExperimentalenvironmentFig.
1Joostarchitecture[5]78Peer-to-PeerNetwAppl(2009)2:75–83sincetheuserisusedtoverylowstart-upsassociatedtotelevisionsets.
InP2Pstreaming,astart-updelayisunavoidable.
Thetimeincurredbetweentherequesttoconnecttoachannelandtheactualstartofplayback,ismainlydueto:(1)theunderlyingpeerdiscoverymechanisms,whiletheP2Poverlayselectssuitablepeersthatcanactasuploaders(ordatasources);(2)bufferingtime,requiredtodealwithnetworkjitter,packetloss,andtemporarycongestion.
ForVoD,wefoundanaveragestart-updelayof25s,rangingfromaminimumof16toamaximumof35s.
Ontheotherhand,RTstreamingincurredastart-updelayof5s,rangingfromaminimumof4toamaximumof8s.
ThereasonsforthisconsiderabledifferencebetweenVoDandRTcanbefoundbyanalyzingsomeofthedatadescribedlaterinthisarticle.
Inessence,bufferingtimesinRTaresmallerthantheirVoDcounterpartsincetheformeristransmittedatalowerbitrateandgiveshigherprioritytoresponsivenessratherthanquality.
Alsothediscoverytimeincurredwhilstdeterminingthesourceswillbelower,sinceRTconnectstoanaverageofthreesourceswhereasVoDrequiresanaverageoffivesources(Figs.
7and6,respectively).
5.
2UDPvs.
TCPutilizationJoostusesbothUDPandTCP,althoughouranalysisshowsawiderandmorefrequentuseoftheformer.
SupernodesperiodicallyexchangesmallUDPandTCPsignalingpackets(64bytes)withotherpeersforoverlaymanagementpurposes(e.
g.
,tocheckwhetherrelevantsimplepeersarestillreachable).
Furthermore,everytimeauserswitcheschannel,thepeerneedstogetintouchwithitssupernode,whichhandlesthestreamre-directionprocess.
However,ourdatashowsthatUDPpacketsareusedmorefrequentlyforthispurpose.
Data(video)packetsaretransportedviaUDP(ofapproximately1kb).
ThedifferencesbetweenUDPandTCPprotocolsutilizationforthecasesofVoDandRTarevisiblefromFigs.
3and4.
ThefacttheRTpacketsareonlyobservedforabout20minisbecausethisiscurrentlytheaveragelengthoflivebroadcastingsessioninJoost.
Figure3,showsthatUDPincursahigherrateatthebeginningwhichisduetotheneedtobufferasfastaspossible,arequirementwhichislessimportantinVoD.
However,lookingatthelong-termaverageoftransferrate,wecanseethatRTisencodedatalowerratethanVoD.
Clearlytheextraconstraintsintro-ducedbyRTrequirerelaxingsomeoftherequirementsonthequalityofthestream.
Figure4,givesanaccountofthedifferentpatternobservedinRTandVoDintermsofsignalling.
Infact,TCPisnotusedfordatatransferinJoost.
WecanseetheRTrequiresahigherpeakofsignallingatthebeginning,butthelong-termaveragegoesbelowitsVoDcounterpart.
Wehavenotbeenabletodefinereasonsforthis.
However,thisbehaviorsuggeststhatRTinitiatesamoreaggressivesearchforsources,sinceitrequiresconnectingquickly.
AndthisisalsoinagreementwiththelowerconnectiontimesseeninRT(5sasopposedtothe25sofVoD),asobservedinSection5.
1.
5.
3DownloadvideotrafficInthissectionweevaluatethetrafficcomingintoapeer(downloadmode).
Throughput,asdepictedinFig.
5,ismeasuredbyaddingallvideocomponents(orsub-streams)whichconstituteanindividualsession(eitherVoDorRT).
Bitrateisobtainedbyaveragingovera10-swindow.
Fig.
3UDPutilizationFig.
4TCPutilizationPeer-to-PeerNetwAppl(2009)2:75–837979ThemostnotableresultfromFig.
5isthatRTisencoded(thustransmitted)atalowerbitratethanVoD.
WehavealreadycommentedonthisaspectbeforeinSection5.
2,however,thisbehaviorsuggeststhatRTinitiatesamoreaggressivesearchforsources,sinceitrequiresconnectingquickly.
AndthisisalsoinagreementwiththelowerconnectiontimesseeninRT(5sasopposedtothe25sofVoD),asobservedinSection5.
1.
WealsoobservethatthevariabilityinVoDtransmissionrateismuchhigherandseveralintervalsdonotseeanypackettransmissionatall.
InVoDthisisacceptablesinceplaybackdeadlinesarelessstringentthaninthecaseofRT.
AgreaterinsightisachievedbylookingatFigs.
6and7.
Thenumberofsourcesisonaverage5inVoD(Fig.
6),bycontrasttotheaverageof3inRT(Fig.
7).
WealsostudiedhowfrequentlysourceschangedduringthetransmissionandfoundconfirmationthatthisisconsiderablyhigherinVoD(datanotincludedforbrevity).
ThismeansthattheVoDoverlaymanagementalgorithmprioritizesloadbal-ancing(thisresultsfromthefrequentconnectionsanddisconnectionsbetweenpeers,andespeciallyinVoDforfindinggoodqualityoftransmission),whereasRTpriori-tizesasmoothertransmission.
Soonceasetofsourcesisidentified,RTtriesnottochangethemduringtransmission,topreventlossofqualityortransmissionintermittency.
5.
4UploadvideotrafficUploadtrafficisincurredwhenotherpeersconnecttothepeerunderscrutiny(inourtestbed).
ThisdataisveryusefulinestablishinghowJoosthandlestheasymmetrytypicalofmanyaccessnetworks(e.
g.
,ADSLallowsforhigherdownloadratesthanuploadones).
Figure8,showsthatthisrequirementis,infact,satisfiedbyJoost.
BothRTFig.
8UploadthroughputFig.
7Numberofactivesources(RT)Fig.
6Numberofactivesources(VoD)Fig.
5Downloadthroughput80Peer-to-PeerNetwAppl(2009)2:75–83andVoDfiguresareconsiderablylowerthantheirdownloadcounterparts.
TheaverageratefortheuploadingfortheRTwas0.
68MbpsandforVoDwas0.
46Mbps.
ConsistentlywithFig.
8areourobservationsofinboundconnections.
Joostisconfiguredinawayastoallowonaverage1inboundconnection,bothinRTandVoD.
Thepatternofconnectivityis,however,different.
RTinboundconnectionsaremainlyconcentratedintheinitialpartoftheoverallbroadcastperiodbutarerelativelymorestableandincurahigherratethaninVoD.
ThisiscoherentwiththeneedforgreaterstabilityinRT,wherehandoveramongdifferentsourcesiskepttotheminimum.
Ontheotherhand,VoDgetsinboundconnectionrequestduringalongerperiod,althoughtherearelongerperiodshavingnoconnections.
Thisreflectsthecontinuoushand-overamongalternativepeersforloadbalancingpurposes(asalreadynotedabove).
Worthmentioningisthatwhenwestudiedthecorrela-tionbetweeninboundandoutboundconnectionsamongpeers,wenoticedthatthereisnoreciprocity.
Thatistosaythat,contrarytomanyP2PfilesharingsystemssuchasBitTorrent,Joostdoesnotemployatit-for-tatpolicy.
Inotherwords,nodesdonottrytoactasuploadersforthosepeerswhohavepreviouslyprovideddata.
5.
5P2PstreamingapplicationstaxonomyTable1showshowJoostperformsandworkscomparedtootherpopularP2Pstreamingapplicationswithrespecttodifferentparameters.
6NetworklocalityandgeographicdistributionNetworklocalityistheabilitytomaintaintheP2Poverlayinsuchawayastocreatelogicalconnectionsamongpeerswhoarephysicallyclosetoeachother.
Thewayinter-communicatingpeersaregeographicallydistributedhassignificantimplicationsintermsofnetworkefficiency—theidealconditionbeingwhenthemostintensivedataexchangeshappenamongnearbypeers.
InordertodeterminewhetherJoostislocationaware,wehaveconductedanumberofexperiments,monitoringtheIPaddressesofallcommunicatingpeers.
WethenusedIPNETinfov1.
10tomapIPaddressestophysical(geo-graphical)addresses—atcityandcountrylevels.
Asbefore,weconsideredVoD,RT,upload,anddownloadtraffic.
However,whileintheexperimentsdescribedabovewewerestudyingtrafficandconnectionsoverthetime;inthefollowingpie-chartswedisplaytheaveragegeographicaldistributionofinboundandoutboundconnections.
Aselectionofresultsispresentedbelow.
6.
1DownloadtrafficLookingatFigs.
9and10,itisclearthatJoostdoesnotprovideagoodimplementationoftheconceptofnetworklocality.
Infact,connectionswithourUK-basedtestbedarescatteredacrosstheglobe.
ItisironicthatonlyasmallfractionofdatacomesfromtheUK(26%).
IfwethenbreakdownthedistributionofthisUKinboundtraffic,mostofitoriginatesfromcitiesthatarefartheraway.
Unexpectedly,mostoftheVoDtrafficoriginatesfromCanadawhereasmostoftheRTtrafficcomesfromtheUSA.
ThisseemstoinferthatJoostloadbalancescomputingresourcesbutneglectsnetworkresourceoptimization.
6.
2UploadtrafficTheanalysisofuploadtrafficallowsstudyingthedistributionofpeersconnectingto(streamingfrom)ourtestbed(Figs.
11and12).
Again,wenoticeanetwork-unfriendlybehavior.
However,neitherVoDnorRTtrafficgotothosecountriesthatwereprovidingthehighestpercentageofdownloaddata,thatisCanadaforVoD(Fig.
9)andtheUSAforRT(Fig.
10).
AlsotheportionofoutboundtrafficgoingtotheUK(11%)wasalldirectedtoauniquelocation(Watford),whichisfarawayfromourtest-bedsite(Fig.
12).
TheseresultsconfirmourpreviouscommentabouttheintrinsicunfairnessofJoostwhichdoesnotimplementatit-for-tatpolicy.
ItisthereforehardtodrawanydefiniteconclusionastowhetherJoostimplementsanexplicitcomputationalloadbalancingalgorithm.
Itispossiblethatahighdegreeofrandomnessisembeddedinoverlaymanagement.
7ConcludingremarksInthispaperwehaveconductedanexperimentalstudyononeofthemostrecentP2Pstreamingapplications,whichfeaturesgenuineP2PmechanismsandsupportsbothRTTable1P2PstreamingapplicationsP2PapplicationServiceStart-updelayDatarateArchitectureProtocolSopcastLive/VOD1–5min300–350kbpsMeshUDPPPliveLive/VOD20s–2min500kbpsMeshTCPJoostVOD25s500kbpsMeshUDPZattooLive6.
2s560kbpsTree/ForestTCPPeer-to-PeerNetwAppl(2009)2:75–838181andVoDservices.
Throughpassivemeasurementsonthoseserviceswehavecharacterizedsomeunderlyingmecha-nisms,includingprotocolbehaviorandsignalingover-heads.
ThishasleadtotheidentificationofimportantfactorsandissuesthatarecommontothemoregeneralareaofP2Pbroadcasting,P2PIPTV(includingVoDandRTservices),andP2Pconferencing.
Lessonslearnedfromthisstudy,whichcanyieldinterestingfurtherdevelopmentsandinvestigations,canbesummarizedasfollows:&Loadbalancing:P2Psystemsareknownfortheirabilitytoautonomicallybalancecomputingresources.
ThisisachievedwellinP2Pfilesharingapplicationswheretimeconstraintsarenotsostringent,whichallowsforbetterresourceoptimizationandoverlaymanagement.
However,P2Pstreamingposestimelimitswhichmakethistaskharder.
Methodswhichcaterfornear-optimalloadbalancingofcomputingresourceswithinthelimitsimposedbyRTstreamingwillacquireimportancesincemobile,thinterminalswilldemandmoreeffectiveoff-loadingmechanisms.
&Networklocality:streamingwithoutconsideringwaysforoptimizingtheuseofnetworkresourcesisboundtoposeserioushurdles,sincethisclasheswiththemodusoperandiofnetworkoperatorsandISPs.
Methodswhichallowprioritizingconnectionsbasedongeo-graphicalproximityaswellasmobilitypatternshaveconsiderablepotentialintermsofnetworkefficiency.
&Fairnessandfreeriding:thisisoneoftheissuesincommontoalltypeofP2Papplications,notmerelyP2Pstreaming.
Theremustbeawaytoprioritizeconnectionsbasedonmutualresourcesharing.
Thereis,ontheotherhand,astrongtrendtowardsfreeriding(i.
e.
,peerwhogetresourceswithoutsharingtheirown),aproblemwhichdramaticallydegradestheperformanceofP2P.
InthecaseofP2Pstreaming,itseemseasiertotacklethisissue.
Appropriatemechanisms,coupledwitheconomicmodels(e.
g.
,basedonincentives)willberequired.
Belgium14%France14%Poland12%Spain26%UK11%Switzerland9%Bulgaria7%Estonia7%Fig.
12Upload(RT)Australia23%Finland22%Denmark11%USA11%Italy16%Germany5%France5%Japan4%Austria3%Fig.
11Upload(VoD)USA77%UK23%Fig.
10Download(RT)Canada20%Sweden14%UK14%SouthAfrica13%UK12%Portugal9%Romania6%Norway5%Portugal4%Croatia3%Fig.
9DownloadVoD82Peer-to-PeerNetwAppl(2009)2:75–83&Start-updelay:thisisexcessiveincurrentsystems,whichmakestheuserexperiencepoorifcomparedtoconventionalTVbroadcasting.
Switchingchannelsisalsotooslow,whichmakeszappingimpossible.
Thisseemsahardproblemsincebufferingtimecannotbereducedincurrentbest-effortnetworks.
&Mobileuser:currentP2Pstreamingsystemsimposerequirements(intermsofcomputationalandaccessnetworkcapability)thatarebeyondthereachofstate-of-the-artmobilephonesorPDAs.
MobileP2Pstream-ingposessomechallengingresearchissuesthatareboundtoattractvividattention.
OpenAccessThisarticleisdistributedunderthetermsoftheCreativeCommonsAttributionNoncommercialLicensewhichper-mitsanynoncommercialuse,distribution,andreproductioninanymedium,providedtheoriginalauthor(s)andsourcearecredited.
References1.
PPlive,www.
pplive.
com.
LastAccesswason17thofApril20082.
Joost,www.
Joost.
com.
LastAccesswason18thofApril20083.
PPStream,www.
ppstream.
com.
LastAccesswasonFebruary20084.
Sopcast,www.
sopcast.
com.
LastAccesswasonJanuary20085.
LeiJetal(2007)AnexperimentalanalysisofJoostpeertopeerVODservices.
TechnicalReport,ComputerNetworksGroup,UniversityofGottingen,Germany6.
HeiX,LiangC,LiuY,RossK(2007)AmeasurementstudyofalargescaleP2PIPTVsystem.
IEEETransMultimed9(8):1672–16877.
DoT,HuaK,TantaouiM(2004)P2vod:providingfaulttolerantvideo-on-demandstreaminginpeer-to-peerenvironment.
ProcIEEEICC3:1467–14728.
GuoY,SuhK,KuroseJ,TowsleyD(2007)P2cast:peer-to-peerpatch-ingforvideoondemandservice.
MultimediaToolsAppl33(2):109–1299.
HuangC,LiJ,RossK(2007)Peer-assistedVoD:makinginternetvideodistributioncheap.
InProc.
IPTPS2007,Feb26–2710.
ZhangX,LiuJ,LiB,YumT-SP(2005)DONet/CoolStreaming:adata-drivenoverlaynetworkforpeer-to-peerlivemediastreaming.
ProcIEEEINFOCOM3:2102–211111.
LiuY,GuoY,LiangC(2008)Asurveyonpeer-to-peervideostreamingsystems.
Peer-to-PeerNetwAppl1(1):18–28March12.
DiotC,LevineBN,LylesB,KassemH,BalensiefenD(2000)DeploymentissuesfortheIPmulticastserviceandarchitecture.
IEEENetw14(1):78–8813.
JiangX,DongY,XuD,BhargavaB(2003)GnuStreamAP2Pmediastreamingsystemprototype.
Proc2003ICME2:325–32814.
QiuD,SrikantR(2004)ModelingandperformanceanalysisofBitTorrent-likepeer-to-peernetworks.
SIGCOMM34(4):367–37815.
HofeldTetal(2008)Aqualitativemeasurementsurveyofpopularinternet-basedIPTVsystems.
AppearsintheInternationalConferenceonCommunicationandElectronics,4–6June2008,pp156–16116.
XieSetal(2007)Coolstreaming:design,theory,andpractice.
IEEETransMultimed9(8):1661–167117.
KovacevicAetal(2008)Locationawareness—improvingdistributedmultimediacommunication.
ProcIEEE96(1):131–14218.
SilverstonT,FourmauxO(2007)P2PIPTVmeasurement:acomparisonstudy.
Submittedon23Oct2006(v1),lastrevised19Apr2007(v4)19.
SentinelliAetal(2007)WillIPTVridethepeer-to-peerstream.
IEEECommMag45(6):86–9220.
BabaogluO,MelingH,MontresorA(2002)Anthill:aframeworkforthedevelopmentofagent-basedpeer-to-peersystems.
InProc.
22ndIEEEInternationalConferenceonDistributedComputingSystems2002,Vienna,Austria21.
HippR.
SQLite.
http://www.
sqlite.
org/MajedAlhaisoniearnedhisBScinComputerSciencefromQassimUniversity,SaudiArabia(2005)andhisMScinComputerandInformationNetworksfromtheUniversityofEssex,UK(2007).
HeisnowaPhDcandidateattheDepartmentofComputingandElectronicSystems,UniversityofEssex,UK.
Heresearchinterestsencompassheterogeneouswirelesssystems,peer-to-peernetworkingandapplica-tions,anddistributedcomputing.
HisPhDisintheareaofP2Pstreaming.
DrAntonioLiottaisaProfessorofCommunicationNetworksattheEindhovenUniversityofTechnologyinTheNetherlands.
HeisaFellowoftheU.
K.
HigherEducationAcademyandservesanumberofsenioradvisoryboardsincluding:thePeerReviewCollegeofEPSRC(theUKEngineeringandPhysicalSciencesResearchCouncil);thescientificadvisorypaneloftheIWT(theBelgianResearchCouncil);theBoardofEditorsoftheJournalofNetworkandSystemManagement(Springer);andtheadvisoryboardofeditorsoftheInternationalJournalofNetworkManagement(Wiley).
Heisanactivememberofthenetworkingresearchcommunity.
Hehasco-organizedandco-chairedinternationalconferences;hasservedtheTechnicalProgrammeCom-mitteeofover70conferences;andhasalsocontributedaskeynoteandtutorialspeaker.
AttheUniversityofEssex,DrLiottawasleadingthePervasiveServicesteam,knownforitspioneeringworkonubiquitouscomputing,advancedservicemanagement,andsystemsengineering.
Heaprolificandenthusiasticwriter,withover100scientificpublicationstohiscreditintheareasoftelecommunicationservices,distributedcomputing,andadvancednetworking.
Recentarticleshavecontributedtotopicalthemesincluding:operator-mediatedpeer-to-peer;mobilegrids;qualityofexperiencemanagementincontext-awareservices;andthelegalissuesofpervasivesystems.
Peer-to-PeerNetwAppl(2009)2:75–838383