redistributedtvants官网

OntheScalabilityofP2P-BasedPush-DrivenLiveStreamingSystemsCyrilCassagnesUniversityofBordeauxBordeaux,Francecassagnes@labri.
frDamienMagoniUniversityofBordeauxBordeaux,Francemagoni@ieee.
orgHyunseokChangBellLabsHolmdel,NJ,USAhyunseok@ieee.
orgWenjieWangUniversityofMichiganAnnArbor,MI,USAwenjiew@eecs.
umich.
eduSugihJaminUniversityofMichiganAnnArbor,MI,USAjamin@eecs.
umich.
eduAbstract—TelevisiontransmittedoverIP(IPTV)presentsnu-merousopportunitiesforusersaswellasserviceproviders,andhasattractedsignicantinterestfrombusinessandresearchcommunitiesinrecentyears.
AmongtheemergingIPTVdeliveryarchitectures,thepeer-to-peerbaseddeliverymechanismiscon-sideredattractiveduetotherelativeeaseofservicedeployment.
However,thequestionofhowwellP2PTVapplicationswouldsupportagrowingnumberofusershasnotbeenfullyinvestigatedsofar.
Inthispaper,wetrytoaddressthisquestionbystudyingscalabilityandefciencyfactorsinatypicalP2Pbasedlivestreamingnetwork.
ThroughtheuseofthedataprovidedbyaproductionP2PTVsystem,wecarryoutsimulationswhoseresultsshowthattherearestillhurdlestoovercomebeforeP2Pbasedlivestreamingcouldbecomewidelyused.
I.
INTRODUCTIONWiththeincreasingbroadbandspeedandcontinuedim-provementinvideocompressiontechnologies,Internet-basedtelevisionservices(IPTV)havebeenexperiencingsustainedgrowthlately.
WhenitcomestorealizingIPTVservicesintoday'sInternet,peer-to-peer(P2P)baseddeliverymechanismisconsideredanattractiveoptionbecauseoftheeaseofdeploymentandpotentialbandwidthsavings.
InatypicalP2PbasedIPTVnetwork,clientsretrievevideostreamsbyconnectingtothebroadcastserverorotherexistingclientsthatarealreadyconnectedtothenetwork.
ThebroadcastservergeneratespacketizedvideostreamsbyencodingliveTVsignalscapturedfromsatellite.
Afterjoiningthenetwork,clientscancontributetheiruplinkbandwidthbyforwardingtheincomingvideostreamstolaterjoiningclients.
Toallowmoreefcientutilizationofclient'suplinkbandwidth,thevideostreamsaretypicallydistributedviatheoverlayintheunitof"chunks"(e.
g.
,[1])or"sub-streams"(e.
g.
,[2],[3]).
Chunksaretime-dividedsegmentsofpacketizedstreams,whilesub-streamsarespace-dividedsubsetsoftheoriginalstreams(e.
g.
,layersinH.
264SVC).
Thechunksorsub-streamsareeitherpushedbyforwardingclients,orpulledbyreceivingclientsdependingontheP2Psharingprotocolused.
Inthepull-drivendelivery,clientssearchandpullindividualstreamunitsinanopportunisticway,whileinthepush-drivenapproach,aclientestablishesa"virtual"connectiontoaforwardingclient,andcontinuestoreceivedatapushedfromtheforwarderuntileitherendterminatestheconnection.
Push-drivendeliverydesignwasshowntobemoreefcientthanpull-basedcounterpartinrecentwork[4].
ComparedtotraditionalP2Pdatasharingorprogressivestreamingofvideoondemand,optimizingend-userexperienceintheP2Pbasedlivestreamingenvironmentisanon-trivialtaskbecauseofitsmorestringentdelayconstraintandlimitedsharedbufferspace.
Inaddition,uploadcapacityconstraintsandinherentchurningbehaviorofparticipatingclientscanaddtothedifcultyinrealizingafullyscalabledeliverysystem.
MotivatedbyourearlierstudyontheoperationalscalabilityofP2Pbasedlivestreaming[5],weexploretheimpactofuplinkbandwidthdistributionandpeerselectionalgorithmsonthesystemperformance.
Inourstudy,wefocusonthepush-driven,sub-streambasedstreamingarchitecture,andperformsimulationsinstantiatedwiththedatacontributedbyaproductionsystememployingsuchanarchitecture.
Theremainderofthispaperisorganizedasfollows.
Sec-tionIIoutlinesthepreviousworkdoneonP2Pbasedlivestreamingsystems.
SectionIIIdiscussesthescalabilityoftheredistributioninP2Pbasedlivestreamingnetworks.
Sec-tionIVproposesmechanismsforincreasingefciencyofthepeerselectionalgorithminP2Pbasedlivestreamingnetworks.
SectionVpresentsevaluationresultsobtainedbysimulatingthearchitectureofoneofthelargestproductionP2PTVprovidersinEurope.
Weconcludethepaperbypresentingfutureresearchdirectionsinthelastsection.
II.
RELATEDWORKManymultimediastreamingsystemshavebeenproposedandevaluatedintheresearchcommunityinrecentyears[6],[7],[8],[9],[10],[11].
Besidestheseresearchprototypes,anumberoffullyoperationalP2PTVsystemshaveemerged(e.
g.
,PPLive,PPStream,SopCast,TVAnts,LiveStation,Joost,andZattoo).
WiththeP2PTV'sgrowingpopularity,alargenumberofmeasurementpapershavebeenpublishedontheseP2Plivestreamingsystems,tacklingworkloads[12],topolo-gies[13]orspecicsystems[14],[15],[16],[17],[18],[19].
NumerousproposalsforimprovingexistingP2P-basedlivestreamingsystemshavealsobeenpresented,rangingfromfeasibility[20]anduploadcapacity[21]tolocalityaware-ness[22]andevenstochasticuidtheoryforP2Pstreamingsystems[23].
Manyrecentproposalsbasedonrobustincen-tives[24],altruism[25],contributionawareness[26]andsub-streamtrading[27]aimatavoidingfree-ridersinlargescalesystems.
III.
REDISTRIBUTIONINP2PLIVESTREAMINGOVERLAYSInthissection,westudytheimpactofclient'suplinkbandwidthonthescalabilityofaP2Psystem(i.
e.
,maximumnumberofusersthatcanbesupportedbythesystem).
Inthissection,wedonotconsiderpeerdynamics(e.
g.
,join-ing/leavingbehavior).
Wewillincorporatepeerdynamicsinlatersections.
Wedenetheratiooftheincomingstreamthatcanberedistributedtootherpeersasaredistributionfactor,andlabelitwithkthroughoutthispaper.
Theredistributionfactorkcanvarybetween0toinnity,andcanbefractional,dependingonpeer'suplinkcapacity.
Ifk=1,itmeansthatthepeercanredistributethefullstream.
Ifk=2,itmeansthatthepeerredistributestwocopiesofthestream.
Ifk=0.
5,itmeansthatthepeerredistributesonlyhalfofthestreamduetoitsuplinkbandwidthconstraint.
Fractionalvaluesarepossiblebecauseafullstreamcanbedividedintomultiplesub-streams,whichallowsapeertoredistributeonlyasubsetofthestreamtootherpeers.
Technically,theredistributionfactorofapeercanbeestimatedasthetotaluploadbytesdividedbythetotaldownloadbytesofthepeer.
Inthispaper,weassumethattheredistributionfactorremainsconstantovertimeforanypeer.
A.
TheoreticalanalysisFirst,westudytheimpactoftheredistributionfactoronsystemscalabilityinatheoreticalperspective.
Astheoverlayisadirectedacyclicgraph,wecandenethedepthofapeerintheoverlayasthenumberofoverlaylinksbetweenitselfandpeerconnectedtothesource.
Ifweassumethatallpeershavethesameredistributionfactork,wecandeterminethemaximumnumberofpeersintheoverlaybyusingasimplegeometricseriesexpressedink,thegraphdepthn,andthesourcecapacityC.
Thatis,ifUnrepresentsthenumberofpeersatdepthnandifU0=C,thenwehaveUn=k*Un1.
Thetotalnumberofpeersisthusequalto:Sn=ni=0Ui=C*1kn+11k(1)Themaximumnumberofpeersabletoconnecttothesystemwilldependonthevalueofk:Ifk1thenthenumberofpeerswillexponentiallydiverge;Un→+∞,andthusthesystemwillscaleinsize.
Consideringamoregeneralcasewhereatagivenlevell,eachpeerihasanupload(uli)anddownload(dli)capacityrate.
Then,thekfactorforthislevellcontainingppeerswouldbe:kl=pi=1ulipi=1dli(2)Wedeneanidlepeerasonenotdownloadingthefullstream.
Wedeneanhyperactivepeerasonehavingk>1.
Inthemoregeneralcase,theredistributionfactorkisexpressedas:k=Ua*(1RiRh)+Ui*Ri+Uh*RhDa*(1RiRh)+Di*Ri+Da*Rh(3)Da:numberofdownloadsub-streamsperactivepeerUa:numberofuploadsub-streamsperactivepeerDi:numberofdownloadsub-streamsperidlepeerUi:numberofuploadsub-streamsperidlepeerUh:numberofuploadsub-streamsperhyperactivepeerRi:ratioofidlepeersvs.
totalpeersRh:ratioofhyperactivepeersvs.
totalpeersInanidealworld,wewouldhavek≥1forallthepeersandthushavenooverlaysizelimit(althoughwecouldhaveanundesirablydeepoverlayifkisnotmuchlargerthan1).
Inreality,wehaveamuchdiversesetofredistributionfactors.
kcanbelowerthan1–Freeriders(k=0)–Securityblockedpeers(k=0)–Peerswithlowuplinkcapacity(k>1)–Peerswithhighuplinkcapacity(k>1)–Idlepeers,i.
e.
,peersthatdonotwatchthestreambutredistributeasubsetofthesub-streamsofthestream(k>1forsomesub-streams)B.
IdlepeersInordertoovercomealowkfactor,onesolutionistotakeadvantageofidlepeerswhocouldredistributethestreaminthebackground.
Notwatchingthestreamactively,idlepeerscouldreceiveonlyafractionofthefullstream,andredistributecopiesofthisfraction,thusactingasamultiplier.
Fig.
1showstheP2Plivestreamingnetworkcapacityasafunctionofthepercentageofidlepeers.
Thenumberofsub-streamscontainedinthefullstreamissetto16,sameastheZattoooverlay[5].
Itisassumedthatactivepeershavek=0.
5,andthatidlepeersredistributeonesub-stream(1/16thofthestream)fourtimes(thusk=4).
Toallowequalavailabilityofdifferentsub-streamsinthenetwork,eachactive/idlepeerrandomlychoosessub-stream(s)toredistribute.
Theplotsinthegureareobtainedbyusingformula3.
Thenetworkcapacityisequaltothemaximumnumberofpeersintheoverlaydividedbythemaximumnumberofpeersthatcanconnecttothesourcedirectly.
Ifthenetworkcapacityequalsto1,itmeansthatapeercanonlyconnecttothesource,butnottoanyotherpeers.
Ifthenetworkcapacityequalsto10,itmeansthattheoverlaynetworkcansupporttentimesmorepeersthanthecapacityofthesource.
Thegureshowsboththeactivenetworkandtheidlenetworkcapacities.
Theplotlabeled"active"isthenumberofpeersinthenetworkwithk=0.
5.
Theplotlabeled"total"isthesumofthenumberofidlepeersandthenumberofactivepeers.
Itshowsthatidlepeershelpverylittleinincreasingthetotalnetworkcapacitywhenidlepeerpercentageremainsrealisticallylow(10).
Thustheuseofidlepeersmaynothelpmakeasystemscaleintherealworld.
C.
HyperpeersAnothersolutiontoovercomealowkfactoristousehyperactivepeersforredistributingmultiplestreams.
Anhy-peractivepeerisapeerwhoiswatchingthestreamandcanredistributemultiplecopiesofthestream,oratleastonefullstreamplussomeadditionalsub-streams.
Fig.
2showsthestreamingnetworkcapacityasafunctionofthepercentageofhyperactivepeers.
Similartothepreviouscase,afullstreamconsistsof16sub-streams.
Itisassumedthatregularactivepeershaveak=0.
5,andthathyperactivepeershaveak=1.
5.
Theplotsingure2arealsoobtainedbyusingformula3.
Theplotlabeled"hyperactive"isthenumberofpeersinthenetworkwithk=1.
5.
Theplotlabeled"total"isthesumofthenumberofhyperactivepeersandthenumberofactivepeers.
Wecanseethatthereisdivergencearoundat60%ofhyperactivepeers.
Iftherearelesshyperactivepeers,thentheymusthavehigherkvaluesthan1.
5inordertoreachsuchdivergence.
IV.
EFFICIENCYOFTHEPEERSELECTIONALGORITHMINP2PLIVESTREAMINGWhenanewpeerjoinsanoverlay,itperformsitsownpeerselectionalgorithmtochoosethetargetpeerstoconnectto.
Ifthepeerselectionisdonerandomly,theresultingoverlaycouldbecomequiteinefcientintwoways.
First,theoverlaycouldbecometoodeepandnotwideenough,thusincurringlargeplaybacklags.
Secondly,theoverlaycouldexperiencealotofchurns,thusincurringmanyinterruptionsforusers.
A.
EffectsfortheoverlayWewouldliketoseewhetheritispossibletoimprovetheefciencyofastreamingoverlaybyusingpeerselectionalgorithmsthatincorporatedynamicparameterssuchasup-loadcapacities,sessionlengths,distancesbetweenpeersandFig.
2.
Networkcapacityvs.
amountofhyperactivepeers.
overlaydepthpositions.
Theideabehindthisproposalisthatbyplacingmorestableandhighbandwidthpeersclosertothesource,itcouldmaketheoverlaymoreefcient.
Thepeerselectionalgorithmisagoodplacetoinuencetheevolutionoftheoverlayaswecanmoreorlesscontrolwherethepeerswillplacethemselvesintheoverlay.
Ifapeerselectionalgorithmmanagestoputstablepeersclosetothesource,thisshouldreducetheoverallchurnintheoverlay.
Also,ifthisalgorithmmanagestoputhighbandwidthpeersclosetothesource,thisshouldincreasethecapacityoftheoverlaywhilekeepingareasonabledepthfortheoverlay.
WewillstudythisapproachbysimulationinSectionV.
B.
EffectsforapeerInreality,peerstypicallyzapthroughchannelsandredis-tributeonlyaminimumofthestreaminordertosavetheirlimiteduploadbandwidth.
Suchbehaviorisdetrimentaltothescalablegrowthofanoverlay.
SowhatcanbetheincentiveforaP2PlivestreamingpeertobehaveotherwiseOnesolutionistoemployatitfortatmechanismwithlayeredqualityvideoasproposedin[28].
Iflayeredqualityvideoisnotavailable,anothersolutionthatweproposeistonegotiatethedepthpositionintheoverlayIfthepeerhasahighuplinkcapacityandremainsinthechannel,itwillgraduallygetclosertothesource.
Intheend,thestableandhighbandwidthpeerswouldbeexposedtolesschurnsintheirupstreamconnectivityandenjoynearrealtimeTV,andthereforewouldbeproperlyincentivizedtostayinthechannel.
Ithasbeenshownin[16]thatplaybacklaginalargescaleP2PTVsystemistypicallybetween30to90seconds.
Thus,thepeerpositioncouldbeonefeasibleincentiveforastableandpowerfulpeertostayinthechannel.
InthesimulationresultsshowninSectionV,weusethosecriteria(e.
g.
,sessionlengthanduplinkcapacity)toseeiftheoverlaywouldperformbetterwhenthepeersareselectedbylookingatsuchcriteria.
V.
SIMULATIONOFAP2PLIVESTREAMINGOVERLAYAfterhavingstudiedsometheoreticalaspectsofaP2Pstreamingoverlayandhavingproposedenhancementmech-anisms,wenowpresentresultsobtainedbysimulationandhighlighttheimpactofpeerselectionalgorithmsontheperformanceofatypicalP2Plivestreamingsystem.
A.
SimulationparametersandmetricsOursimulationcodeimplementsZattoo'spush-drivensub-streambasedstreamingarchitecture[3],andwasrunontopofthenetworkmanipulator(nem)software[29].
AnInternetmapof4.
2k-nodewasusedastheunderlyingtopology[30].
WeassumethatagivenoverlaydistributesoneTVchannel,anddonottakedailychannelsizevariationsintoaccount.
Eachrunlastsfor12hoursandonlythelast6hoursareanalyzed;afterthe6thhour,weareinasteadystateregime.
Eachresultvalueistheaverageof30runs,andthestandarddeviationvaluesareprovided.
Foralloursimulations,weusethesameparametersusedinourpreviouswork[5],whichisbasedontheanalysisof9.
6MsessionsrecordedbytheZattoo'sP2PTVduringa2-weekcampaign.
Accordingtothesessiondata,thecumulativedis-tributionfunction(CDF)oftheredistributionfactorkfollowsanexponentialdistribution.
50%ofthepeerscanredistributelessthan50%ofthefullstream(i.
e.
,k<0.
5).
82%ofthepeerscanredistributelessthanthefullstream(i.
e.
,k<1).
Theuplinkcapacityofindividualpeersisassignedsothattheresultingdistributionbecomesthesameastheempiricaldistribution.
TheNATtypeofapeer,whichdeterminesitsreachabilityintheoverlay,isalsotakenfromtheempiricaldistributionofNATtypesreportedin[5].
Finally,session'sinter-arrivaltimeandsessionlengthareallinstantiatedfromthecorrespondingexponentialdistributionsreportedin[5].
Wevarythechannelsizefromscarcemode(closetosourcecapacity)toheavilycrowdedmode(severaltimeshigherthansourcecapacity).
TableIshowstheremaininginputparameters.
TABLEISIMULATIONPARAMETERSParametersValuesNumberofsub-streamsperstream)16Sourcecapacity50clientsSearchperiod2secMaximumsearchattempts2Jointimeoutperiod0.
25secMaximumsizeofcandidatepeerlist40peersNumberofsimulationrunsperscenario30InordertoassesstheperformancesoftheP2Plivestreamingsystem,westudythefollowingoutputmetrics:1)Viewingtimeratio(in%=100*peerviewtime/peerlifetime):averageviewingtimeofthepeersthatendedduringthisperiod(thehigherthebetter).
2)Percentageofkickedoutpeers(in%):100xtotalnumberofpeersthatcouldnotconnecttotheP2Plivestreamingoverlaynetworkduringthegivenperioddividedbythenumberofnewpeersperperiod(thelesserthebetter).
3)Averagenumberofinterruptionsperpeer:totalnumberofvideoviewinginterruptionsforallpeersinthegivenperioddividedbythenumberofnewpeersperperiod(thelesserthebetter).
Weanalyzetheseoutputmetricsbyvaryingthenumberofnewlyarrivingpeersperhour,whichdenesthestressputonthesourceanditsP2Pnetwork.
B.
SimulationresultsThissectionpresentsoursimulationresults.
Apeertryingtoconnecttootherpeerstogetallnecessarysub-streamsiscalledanorphanpeer.
Itsendssearchmessagestodiscoverotherpeers,sendsjoinmessagestoconnecttoavailablepeers,andnallyobtainsthefullstreamfromthem.
Apeerwhoisableandwillingtoofferapartoforallrequestedsub-streamsforanorphanpeeriscalledanadoptivepeer.
Anadoptivepeerhaspositivelyansweredtothesearchmessageofaorphan.
Oncehavingmultiplepositiveanswersfromcandidateadoptivepeers,anorphanhastochoosetowhichpeeritshouldsendajoinmessage.
WeevaluatethefollowingpeerselectionalgorithmsdescribedinSectionIV:Random:anorphanpeertriestoconnecttoarandomlyselectedadoptivepeer.
Local:anorphanpeertriestoconnecttoitsclosestadoptivepeer(thedistancebeingmeasuredinhops).
Upload:anorphanpeertriestoconnecttotheadoptivepeerproposingthehighestuploadamount(measuredinnumberofsub-streams).
Uptime:anorphanpeertriestoconnecttotheadoptivepeerhavingthehighestsessionlengthasithasahigherprobabilityofstayinglongerintheoverlay(i.
e.
notswitchingchannels).
Figure3showstheaveragepeerviewingtimeasafunc-tionofthetrafcload.
Wecanseethattheeffectsofthevariousalgorithmsontheviewingtimedonotmakemuchdifferencecomparedtoarandomselection.
Althoughthereis7%differencebetweentheworstandthebestalgorithmsatatrafcloadof2000newpeersperhour,and5%differenceataloadof4000,thesevaluesarenottremendous.
Thissomewhatunexpectedresultimpliestherelativeimportanceoftheuserlevelparameterssuchasuploadcapacityandsessionlengthoverthesystemparameterssuchasthepeerselectionalgorithms.
Fig.
4showstheaveragepercentageofpeerskickedoutasafunctionofthetrafcload.
Bycomparingwiththepreviousresults,thedecreaseinviewingtimeismainlyduetopeersbeingkickedoutofthenetwork.
Onlyasmallpercentageiscausedbytheinterruptionsduetopeerde-connectionsandreconnections.
Fig.
5showstheaveragenumberofinterruptionsperpeerasafunctionofthetrafcload.
WeobservethattheaverageFig.
3.
Averageviewingtimevs.
trafcload.
Fig.
4.
Averagepercentageofpeerskickedoutvs.
trafcload.
numberofinterruptionsperpeergraduallyincreaseswhenthenumberofpeersincreasesuntilreachingaplateaufor2000newpeersperhourandabove.
Whenthenumberofnewpeersincreases,theoverlaygrowsandtheaveragechurnratebecomehigher,andthusleadingtomoreconnectionsandreconnections.
However,whentheoverlayisgettingsaturatedbythenewpeers,thosenewpeerscannotmanagetojointheoverlayandarekickedout.
Thus,thenumberofconnectionsandreconnectionsdoesnotgrowasmuch,becausethosekickedoutpeersdonotsignicantlycontributetothisnumber.
However,thetotalnumberofpeersstillincreases,andthustheratiodoesnotincreaseanymore.
Theseresultswerenotexpectedwhenwedevisedthesealgorithms.
Tworemarkscanexplainthedifcultiesforthealternativealgorithmstomakeadifferencecomparedtotherandomselectionalgorithm.
First,thesessionsaretypicallyshort-lived.
Roughly50%ofthesessionsareshorterthanFig.
5.
Averagenumberofinterruptionsperpeervs.
trafcload.
1.
5minutes.
Thiscreatesalotofchurnsthatrendertheevolutionoftheoverlayhardtocontrolovertime.
Second,theredistributionfactordistributionisheavilylopsidedtowardssmallvalues.
50%ofthepeershaveanuploadcapacitylowerthan50%.
Thus,peerswithlongsessionsmayhavealowuploadcapacityandnotbesouseful.
Allinall,thesetwofactorsweighmuchmoreheavilyontheviewingtimethanthevariousselectionalgorithms.
Whentrafcloadishigh,thealgorithmthatperformsbetterthanrandomselectionisthe"local"algorithm.
VI.
CONCLUSIONInthispaperwehavestudiedthebehaviorofapush-driven,sub-streambasedlivestreamingsystem.
Wehaveconductedatheoreticalstudyconcerningtheimpactoftheredistributionfactor,andhaveshownthatthescalabilityofthesystemstronglydependsontheavailableuploadcapacityofitspeers.
Wealsohaveshownthattheuseofidlepeersaswellashyperactivepeersisnotafundamentalsolutiontomakethesystemscalable.
Wehaveproposedpossibleimprovementofthepeerselectionprocessastheselectionprocesshasadirectimpactonthestructureandefciencyoftheoverlay.
Toevaluateitspotentialimpact,wehavecarriedoutsimulationexperimentsinordertomeasuretheefciencyofvariouspeerselectionalgorithmsunderaheavilyloadedlivestreamingsystem.
Wehaveinstantiatedoursimulationwithrealisticparametersderivedfromthedatafrom9.
8Msessionscollectedbytheprofessional-gradeZattooP2PTVsystem.
OurresultsshowthattheredistributionfactorandthesessionlengthhaveaprofoundeffectonthemaximumcapacityoftheP2Poverlay,andthatthevariousselectionalgorithmsplayarelativelymarginalroleinimprovingsystemscalability.
FutureworkwillbeaimedatstudyingotherimprovedpeerselectionalgorithmsaswellasstudyingtheimpactofthebuffersizeandthepeersearchparametersontheoverallefciencyofaP2P-basedlivestreamingsystem.
REFERENCES[1]L.
Bracciale,F.
L.
Piccolo,D.
Luzzi,S.
Salsano,G.
Bianchi,andN.
Blefari-Melazzi,"APush-basedSchedulingAlgorithmforLargeScaleP2PLiveStreaming,"inProceedingsofthe4thInternationalTelecommunicationNetworkingWorkshoponQoSinMultiserviceIPNetworks,2008,pp.
1–7.
[2]S.
Xie,B.
Li,G.
Y.
Keung,andX.
Zhang,"CoolStreaming:Design,Theory,andPractice,"IEEETrans.
onMultimedia,vol.
9,no.
8,December2007.
[3]H.
Chang,S.
Jamin,andW.
Wang,"LiveStreamingPerformanceoftheZattooNetwork,"inProceedingsofInternetMeasurementConference,November2009.
[4]M.
Zhang,Q.
Zhang,andS.
-Q.
Yang,"UnderstandingthePowerofPull-basedStreamingProtocol:CanWeDoBetter"IEEEJSAC,vol.
25,no.
9,pp.
1678–1694,2007.
[5]K.
Shami,D.
Magoni,H.
Chang,W.
Wang,andS.
Jamin,"ImpactsofPeerCharacteristicsonP2PTVNetworksScalability,"inProceedingsofthe28thIEEEConferenceonComputerCommunications–Mini-Conference,April2009.
[6]D.
Tran,K.
Hua,andT.
Do,"ZIGZAG:AnEfcientPeer-to-PeerSchemeforMediaStreaming,"inProceedingsoftheIEEEInfocom,2003.
[7]R.
RejaieandS.
Stafford,"AFrameworkforArchitectingPeer-to-PeerReceiver-DrivenOverlays,"inProceedingsoftheACMNOSSDAV,2004,pp.
42–47.
[8]X.
Liao,H.
Jin,Y.
Liu,L.
Ni,andD.
Deng,"Anysee:Peer-to-PeerLiveStreaming,"inProceedingsoftheIEEEInfocom,2006.
[9]X.
Zhang,J.
Liu,B.
Li,andT.
-S.
Yum,"CoolStreaming/DONet:AData-drivenOverlayNetworkforLiveMediaStreaming,"inProceedingsofthe24thIEEEInfocom,2005,p.
21022111.
[10]F.
Pianese,"PULSE:AFlexibleP2PLiveStreamingSystem,"inProceedingsofthe9thIEEEGlobalInternetSymposium,2006.
[11]F.
Pianese,D.
Perino,J.
Keller,andE.
Biersack,"PULSE:AnAdap-tive,Incentive-Based,UnstructuredP2PLiveStreamingSystem,"IEEETransactionsonMultimedia,vol.
9,no.
6,2007.
[12]K.
Sripanidkulchai,B.
Maggs,andH.
Zhang,"AnAnalysisofLiveStreamingWorkloadsontheInternet,"inProceedingsoftheACMIMC,2004,p.
4154.
[13]C.
Wu,B.
Li,andS.
Zhao,"Magellan:ChartingLarge-ScalePeer-to-PeerLiveStreamingTopologies,"inProceedingsofthe27thInterna-tionalConferenceonDistributedComputingSystems(ICDCS'07),2007,p.
62.
[14]M.
Cha,P.
Rodriguez,S.
Moon,andJ.
Crowcroft,"OnNext-GenerationTelco-ManagedP2PTVArchitectures,"inProceedingsofthe7thInternationalWorkshoponPeer-to-PeerSystems,2008.
[15]S.
Xie,G.
Keung,andB.
Li,"AMeasurementofaLarge-ScalePeer-to-PeerLiveVideoStreamingSystem,"inProceedingsoftheInternationalConferenceonParallelProcessingWorkshops(ICPPW'07),2007,p.
57.
[16]X.
Hei,C.
Liang,J.
Liang,Y.
Liu,andK.
Ross,"Ameasurementstudyofalarge-scalep2piptvsystem,"IEEETransactionsonMultimedia,vol.
9,no.
8,2007.
[17]L.
Vu,I.
Gupta,J.
Liang,andK.
Nahrstedt,"MeasurementandModelingofaLarge-scaleOverlayforMultimediaStreaming,"inTheFourthInternationalConferenceonHeterogeneousNetworkingforQuality,Reliability,SecurityandRobustness,2007.
[18]S.
Ali,A.
Mathur,andH.
Zhang,"MeasurementofCommercialPeer-to-PeerLiveVideoStreaming,"inProceedingsoftheICSTWorkshoponRecentAdvancesinPeer-to-PeerStreaming,2006.
[19]G.
Mara,G.
Pau,P.
D.
Ricoy,andM.
Gerla,"P2PStreamingSystems:ASurveyandExperiments,"inProceedingsoftheSTreamingDay,2007.
[20]K.
Sripanidkulchai,A.
Ganjam,B.
Maggs,andH.
Zhang,"TheFeasi-bilityofSupportingLarge-ScaleLiveStreamingApplicationswithDy-namicApplicationEnd-points,"inProceedingsoftheACMSIGCOMM,2004.
[21]J.
Douceur,J.
Lorch,andT.
Moscibroda,"MaximizingTotalUploadinLatency-SensitiveP2PApplications,"inProceedingsofthe19thACMSPAA,2007,pp.
270–279.
[22]F.
PianeseandD.
Perino,"ResourceandLocalityAwarenessinanIncentive-basedP2PLiveStreamingSystem,"inProceedingsoftheACMP2P-TV,2007.
[23]R.
Kumar,Y.
Liu,andK.
Ross,"StochasticFluidTheoryforP2PStreamingSystems,"inProceedingsoftheIEEEInfocom,2007.
[24]M.
Feldman,K.
Lai,I.
Stoica,andJ.
Chuang,"RobustIncentiveTechniquesforPeer-to-PeerNetworks,"inProceedingsofthe5thACMConferenceonE-Commerce,2004,pp.
102–111.
[25]Y.
ChuandH.
Zhang,"ConsideringAltruisminPeer-to-PeerInternetStreamingBroadcast,"inProceedingsoftheACMNOSSDAV,2004.
[26]Y.
-W.
Sung,M.
Bishop,andS.
Rao,"EnablingContributionAwarenessinanOverlayBroadcastingSystem,"inProceedingsoftheACMSIGCOMM,2006.
[27]Z.
Liu,Y.
Shen,K.
Ross,S.
Panwar,andY.
Wang,"SubstreamTrading:TowardsanOpenP2PLiveStreamingSystem,"inProceedingsoftheIEEEInternationalConferenceonNetworkProtocols,2008,pp.
94–103.
[28]Z.
Liu,Y.
Shen,S.
S.
Panwar,K.
W.
Ross,andY.
Wang,"UsingLayeredVideotoProvideIncentivesinP2PLiveStreaming,"inProceedingsoftheACMWorkshoponPeer-to-peerstreamingandIP-TV,2007,pp.
311–316.
[29]D.
Magoni,"NetworkTopologyAnalysisandInternetModellingwithNem,"InternationalJournalofComputersandApplications,vol.
27,no.
4,pp.
252–259,2005.
[30]D.
MagoniandM.
Hoerdt,"InternetCoreTopologyMappingandAnalysis,"ComputerCommunications,pp.
494–506,2005.