pursuedp2pover

AnEvaluationCriterionforAdaptiveNeighborSelectioninHeterogeneousPeer-to-PeerNetworksMarcoPicone,MicheleAmoretti,andFrancescoZanichelliDepartmentofInformationEngineering,Univ.
ofParma,Italypicone@ce.
unipr.
ithttp://dsg.
ce.
unipr.
itAbstract.
Thepeer-to-peerparadigmpotentiallyenableslow-costandhighlyscalabledistributedsystemswhereusernodeshaveatthesametimetherolesofconsumerandproviderofresources.
Nowadays,thecomputationalpowerandstoragecapacityofmobiledeviceshasnotablyincreased,forwhichtheirinclusioninthephysicalnetworkthatsupportsP2Poverlayscanberealisticallyconsidered.
Forthoseapplicationswherethequalityofserviceisafundamentalconstraint,suchaslivestreaming,itisimportanttohaveoverlaynetworklevelstrategiestodynamicallyre-conguretheactiveconnectionsamongpeers.
Inthispaperweproposeasimplebuteectivestrategyfortheevaluationofneighbors,takingintoaccounttheircapabilities,tosupportthedynamicselectionofre-sourceproviders.
WeapplytheformalframeworktoP2PInternetLiveStreaming,simulatingthreerealisticscenariosanddiscussingtheresults.
1IntroductionPeer-to-Peer(P2P)applicationshaverecentlyemergedasaneectivesolutionforlarge-scalecontentdistributionovertheInternet,fromtheearlyNapster(1999)totoday'swidespreadfullydecentralizeddeliveryofmultimediastreams,withoutrelyingonthetraditionalclient-serverparadigm.
Ingeneral,P2Papproachespotentiallyenablelow-costandhighlyscalabledistributedsystemswhereusernodeshaveatthesametimetheconsumerandproviderroles.
Sharingtheirownresources,peerscontributetotheachievementofaglobaltaskwhichneedsonlyalimited(ifany)infrastructuresinceitex-ploitsthelargepoolofuserresources.
P2PInternetStreamingisbecomingoneofmostpopularservicesontheglobalnetworkandbymeansofapplication-levelmulticastingtechniquesachieveslarge-scaledistributionofmassiveamountsofdatawithstrongtemporalconstraintswithoutcompromisingtheprovidedqual-ityofservice[5,7,8].
Theseapplicationsaremostlytargetedtoasetofwirednodeswhichareconsideredheterogeneousonlywithrespecttotheconnectiontypeandspeed.
Recentyearshavealsoseentherelentlessmarketsuccessofaplethoraofmobiledevices(PDAs,smartphones,MIDs,PMPs,netbooks,whoseeverT.
PfeiferandP.
Bellavista(Eds.
):MMNS2009,LNCS5842,pp.
144–156,2009.
cIFIPInternationalFederationforInformationProcessing2009AnEvaluationCriterionforAdaptiveNeighborSelection145increasingcapabilitiesmakethemattractivetoagrowingnumberofnetworkapplicationsinbusinessandinfotainmentdomainstobefullyexperiencedinmobility.
Althoughthecomputing,storageandcommunicationresourcesavail-ableonmanymodernmobiledevicesareoftennotfarfromthoseavailableoncommonPCs,thespecicissuesofreducedconnectionstability(becauseofbe-ingwirelessandmovingacrossdierentaccessnetworks)andlimitedautonomy(becauseofbeingbattery-powered)shouldbeaddressedinanydistributedap-plicationwhichincludessupportformobilenodes.
Thisworkpresentsanadaptivepeerselectionstrategyforheterogeneouspeer-to-peernetworks,withparticularreferencetoInternetlivestreamingap-plications.
TheprioritizationstrategysupportsextremelyheterogeneousP2Pnetworkswherenodesexhibitverydiversenatureandperformance.
Giventhatneighborselection,i.
e.
onethecriticalfunctionsofanyP2Psystem,isgenerallybasedonlyupontheavailabilityofrequiredresourcesandtheevaluationofthebandwidthcapabilitiesofprospectivepartners,amoregeneralselectioncriterionisrequiredtocopewithnodeheterogeneity.
Theadaptiveselectionstrategyinourapproachcomparesnodesbytakingintoaccountacertainnumberoftheirfeatures(asrequiredbytheapplicationdomain)tothepurposeofincreasingtheexperiencedqualityofservicebyperforminguniformevaluationofhighlyheterogenouspeers.
Thepaperisorganizedasfollows.
Section2outlinesabriefreviewofthemainchallengestobefacedwhendesigningP2Papplicationssupportingmobiledevices.
Insection3somerelatedworkisdiscussed.
Section4providesaformalcharacterizationoftheevaluationcriterionforpeerselectionandintroducesasetofnodefeaturesappropriateforP2Plivestreamingapplications.
Insection5theproposedselectionstrategyispreliminarlyevaluatedbymeansofsimulationonanumberofheterogenousP2Plivestreamingscenarios.
2DesignChallengesSofar,P2Papplicationshavebeenmostlydesignedandimplementedconsideringpersonalcomputersashostsandresourceproviders,ratherthanmobiledevices.
Nowadays,thecomputationalpowerandstoragecapacityofmobiledeviceshasnotablyincreased,forwhichtheirinclusioninaphysicalnetworkthatsupportsaP2Poverlaycanberealisticallyconsidered.
Inthissectionwerecallsomegeneralconsiderationstobetterdenetheissuesthatarisewhenmobiledevicesparticipateinadistributedsystem[6].
Dierentinterfacesfordataconnectivity.
Inparticular,mobilephonesmayhavedierentkindsofdataconnections(e.
g.
CDMA/GPRS/3G)orinaddition,asinthecaseofmostmoderndevices,theycanaccessWiFinetworks.
ApplicationslikeP2Pmediastreamingareverysensitivetodelaysandlossofsegments.
Forallthesereasons,thedierentnatureofconnectionsandthepossibilityofsuddenswitchesfromanaccessnetworktoanother,arefactorstotakeintoaccountinordertodynamicallyadapttheoverlaynetworktothecharacteristicsofindividualnodes.
146M.
Picone,M.
Amoretti,andF.
ZanichelliIntrinsicmobilityofdevices.
Physicalmobilitymayinvolvechangesinaccessspeedtothenetwork,butcanalsocauseprolongedormomentarydisconnectionsofpeers,morefrequentlywithrespecttopeerhostedbypersonalcomputers.
Redundancyofresourcesandquickdiscoveryofnewprovidersisthegeneralsolutiontothisproblem,butapurelyreactiveapproachisnotsucientforap-plicationslikeP2Pmultimediastreaming,forwhichproactivestrategiesshouldpursuedinstead(e.
g.
eachpeershouldhavealistofsuitableneighbors,inthesensethatifanewsegmentproviderisselectedwhenthecurrentonedisconnectsfromthenetworkoroerspoorperformance,packetlossesareminimized).
Limitedresources.
Asmentionedabove,withrespecttopersonalcomputers,mobiledevicesarecharacterizedbylimitedbatteryautonomylowmemoryandreducedcomputationalpower.
Thisaspectisveryimportantandrequiresseveraloptimizationsforthesoftwaretoberunonthedevice,inordertoguaranteeagoodqualityservice.
Comparedtothepreviousissue,thisoneisfortunatelybecominglessandlessrelevant,sinceanincreasingnumberofmobiledeviceswithveryhighautonomyandgenerouscomputationalandstoragecapacitiesisenteringthemarket.
Thesenewmobiledeviceswillallowthedeveloperstocreatericherapplications,improvingtheexperiencefortheuser.
3RelatedWorkDynamicneighborevaluationandselectionisusedinseveralP2Psystemsde-scribedintheliterature.
InKaZaA[3]andGnutella2[2],themostpowerfulnodes(e.
g.
thosewithhigherbandwidth)aresetassuper-nodestoformthebackboneoftheP2Pover-laynetwork,whichmanagesmostofthemessagingloadforresourcediscovery.
InBitTorrent[1]systems,centralserversstoreinformationabouttrackers,i.
e.
agentsthatareresponsibleforhelpingpeersndeachother.
TheBitTorrentprotocolfocusesonhighdatatransferspeedratherthanonsearchcapabilities.
Whenjoiningthetorrent,thepeeraskstothetrackeralistofIPaddressofpeerstobuilditsinitialpeerset,i.
e.
thelistofotherpeersitknowsabout.
Apeercanonlysenddatatoasubsetofitspeerset,calledactivepeerset.
The"chokealgorithm"determinesthepeersbeingpartoftheactivepeerset.
Eachdownloaderreportstoallitspeerswhatpiecesithas,thuseachpeerknowsthedistributionofthepiecesforeachpeerinitspeerset.
ThepieceselectionstrategyusedinBitTorrentisbasedonrarestrststrategy.
Kwongetal.
[9]proposeasimplebutinterestingprotocolforbuildinghet-erogeneousunstructuredP2Pnetworks.
Thebasicideaistotakeintoaccountnodecapacityduringthejoiningandrebuildingprocess.
Inarststep,incom-ingnewpeersselectasneighborsthosepeersthataremoresuitableintermsofcapacityandconnectivity,withthepurposeofachievinggoodload-balancing.
Therebuildingprocessdescribeshowthenodesacttore-establishanecienttopologywhensomeoftheirlinksarebroken.
Tothesamepurpose,i.
e.
dynam-icallyre-shapingtheoverlaynetworktopology,interestingstrategiesbasedongeneticalgorithmshavebenproposed[10,11].
AnEvaluationCriterionforAdaptiveNeighborSelection147InapplicationslikeP2PInternetlivestreaming,wherethequalityofservice(QoS)playsaveryimportantrole,neighborselectionmustbecarefullyperformed.
Thebasicsolutionisrandomneighborselectionfromthelistreceivedbytheboot-strapnodeofthesystemorbyadecentralizedmechanism.
AbettersolutionisadoptedbyotherprotocolslikeCoolStreaming[5],wherenodesperiodicallyex-changeinformationabouttheirstatus,e.
g.
aboutthestreamsegmentsthattheyownandtheycanshare,andabouttheiractualbandwidth.
Theseinformationareusedbytheproviderselectionstrategyduringthestart-upprocessofeachsinglenode,andformanagingchangesofproviderduringthelifeofthepeer.
TocreateaP2Pstreamingprotocolthatcanbeappliedtoanetworkofheterogeneoushosts,includingmobiledevices,itisimportanttointroduceapeerselectioncriterionthatdynamicallyevaluatestheQoSoftheneighbors,inordertopersonalizethebehaviorofeachpeer.
AworkofNematietal.
[7]triestoanalyzeandmodelanetworkwithmobileandheterogeneousdevicesintermsofmobilityandQoS.
Theprotocoldenestwokindsofmobilepeers:indirectlymobilepeerswhosehostsaremobile,anddirectlymobilepeersthatarerealizedasmobileagents.
Amobilepeercanmanipulateamultimediacontentinahost.
QoSsupportedbythepeerdependsonthehost.
MovementsofmobilepeersaremodeledintermsofchangesofQoS,whichisdescribedandcharacterizedbybandwidth,delayandpacket-loss-ratiounderoverlaylevelandattheoverlaybyframe-rate,resolution,numberofcolors,qualityofsound,etc.
Withthismodelthepeersofthenetworkcananalyzetheparametersofothernodestoselectthebestprovider,tryingtomaximizetheobtainedQoSinadynamicway.
4FormalStatementoftheEvaluationCriterionWeconsideranetworkwithNpeersasagraphG=(V,E),whereVisthesetofverticesandEisthesetofedges.
Eachpeerhasanumberofresources,whosedomainisdenedasR.
Moreover,eachpeercandistinguishprincipalresources,inasetRp,fromsecondaryresources,inasetRs,suchthatR=Rp∪Rs.
Thestatusofbothmainandsecondaryresourcesshouldbetakenintoaccountforapeerthathastobeevaluatedbyanotherpeer.
Butinaheterogeneousnetworksomekindsofpeercouldbelackingoneormorekindsofresources,withrespecttootherricherpeers.
Forthisreasonweintroduceauniqueparametertocharacterizeanykindofpeer,i.
e.
thetness.
Inotherwords,westatethatforeachpeerinVfi>0denedas:fi=f(gi,hi)=gi(rs)K0+hi(rm)K0>0rs∈Rmisrm∈Rnip(1)whereg:Rms→Rm∈N(2)andh:Rnm→Rn∈N(3)148M.
Picone,M.
Amoretti,andF.
ZanichelliAsillustratedingure1,thetnessisafamilyofcurvesthatdependonthevalueoffunctionshandg.
Foragivenpeeri,withniprincipalresourcesandmisecondaryresources,thetnessvalueandhiareinverselyproportional,whilethetnessvalueandgiaredirectlyproportional.
Thecurveprogressoffunctionsg,h,aswellasthevalueofK0mustbedenedbasedontheapplication,alwaysavoidingthefollowingsituation:f1(h1,g1)>f2(h2,g2)h1>h2(4)inwhichthenodewithlowesthhasthehighesttnessvalue.
Fig.
1.
FitnesscurveprogresswithrespecttohandgThisevaluationcriterioncanbeusedinawiderangeofP2Papplications.
ForexampleweconsideraP2PInternetLiveStreamingapplicationbasedonthedata-drivenapproach,wherethestreamisdistributeddependingontheavail-abilityorneedofdatachunks,unliketraditionalInternetstreamingsystems(e.
g.
thosebasedonthesource-drivenparadigm)thatperformexplicitsearchoftheoptimalstreamingpathsfromthesource(s)tothesinks.
Thetness-basedapproachintroducesadaptivenessinthedata-drivensystem,tomaintainahighdegreeofeciencyeveninpresenceofheterogeneousnodeswithdier-entavailabilityproles(e.
g.
mobilenodesversusstablenodes).
Eachnodeisperiodicallyevaluatedaccordingtothestatusofitsresources,i.
e.
itsdynamicfeatures(residualbattery,typeofnetworkconnection,availablebandwidth,on-linepermanencetime),withatnessvaluebeingcomputed.
Bycomparingthetnessofitsneighbors(i.
e.
knownpeers),eachpeercanselectthelocallyoptimalmultimediastreamproviders.
Thisallowtoreducetheimpactonthesystemofleastperformingnodes(e.
g.
nodeswithreducedbandwidthandcomputationalcapacity),andtoimprovetheoverallqualityofservice.
Thetnessvalueofasinglenodeiscomputedconsideringthefollowingpa-rameters(whichrepresentresourcescharacterizingeverynode):AnEvaluationCriterionforAdaptiveNeighborSelection149–actualuploadbandwidth(U):thenominaltransmissionrateofanode,overthenumberofneighborsbeingservedbythatnode–batterypercentage(B):incaseofPCnode,itsvalueisalways100;foramobilenode,thevaluedecreasesdependingonnodeusage–on-linepermanencetime(To):itsvaluedeneshowlongthepeerhasbeenconnectedtothenetworkAmongtheseresources,weconsideredUasprincipal,whileBandToaresec-ondary.
Sinceallnodesarecharacterizedbythesamenumberofresources(i.
e.
mi=mandni=nforeachi=1,.
.
,N),thehandgfunctionsarethesameforeverynode.
Supposinganodeievaluatesanothernodej,thehfunctionappliedtothelatterisdenedash(Uj)=min{Di,Ujkj+1}S1(5)whereDiisthedownloadrateofnodei,Sisthestreamrateofthemultimediasourceassumedtobeconstant,andkj=Nl=1elj(6)whereelj∈{0,1}hasnon-zerovalueifnodelisbeingservicedbynodej.
In(5),thenumeratorrepresentsthetransmissionratethatwillbedevotedtonodeifromnodej,ifselected.
Ujisevaluatedwithrespecttoalreadyactiveoutgoingconnections(whichingeneralarelessthanthenumberofknownnodesbyeachpeer),plustheenvisionedconnection(forwhichneighborsarebeingevaluatedbynodei).
Thegfunctionappliedtonodejisdenedasg(Bj,Toj)=K1BjBMax+K2TojToMaxK1>0K2>0(7)Thusthetnessfunctionappliedtonodejisfj=f(gj,hj)=g(Bj,Toj)K0+hj(Uj)(8)Fitness-basedevaluationcanbeappliedtothefollowingphasesofagenericdata-drivenprotocol(seealsogure2):1.
Neighborlistupdate:tolimitthenegativeeectsofrandomdisconnectionsandtomonitorthequalityofserviceofthesystem,eachpeerperiodicallyupdatesitsneighborlist;thisisdonebysendingmessagestoallknownnodes,whichinturnrespondwiththeirupdatedresourcevalues;ifacontactednodedoesnotrespondbeforeagivendeadline,itisremovedfromtheneighborlist,andreplacedwithanothernode.
150M.
Picone,M.
Amoretti,andF.
ZanichelliFig.
2.
Exampleofneighborlistupdate,performedbynodeA2.
Providerselection:oncethepeerhasupdateditsneighborlist,itselectstheneighborwiththehighertnessvaluetobetheproviderofthemultimediastream.
5SimulationResultsTheadaptivepeerselectionstrategyappliedtolivestreaminghasbeenimple-mentedandpreliminaryevaluatedusingtheDiscreteEventUniversalSimulator(DEUS)[4],whichisanopensourcetoolthatprovidesasimpleJavaAPIfortheimplementationofnodes,eventsandprocesses,andastraightforwardbutpowerfulXMLschematoconguretheexperiments.
DEUSaimstobeoneofthereferencetoolsintheeldofcomplexsystemsimulation,thankstoitshighexibilityandperformance.
ThreenodetypeshavebeenmodelledwiththeparametersillustratedinTable1:PCwithhighperformanceintermsofcomputationalcapacityandbandwidth,Mobile-WiFinodeswithlimitedresourcesbutaccessingthenet-workwithaWirelessLanconnection,and3GMobilenodesusingaconnectionwithlowerbandwidthandsubjecttothevariabilityofcellularnetworks.
Forthebehavioralmodelingofmobiledevicestwotypesofspeciceventshavebeenconsidered:Table1.
ParametersofthesimulatednodesNodetypeUplinkDownLinkBatteryMobile3G-Node50Kbit/sec400Kbit/sec[20,100]MobileWiFi-Node100Kbit/sec2Mbit/sec[20,100]PC-Node300Kbit/sec4Mbit/sec100%AnEvaluationCriterionforAdaptiveNeighborSelection151–Thetransitionfroma3Gtoa2Gconnection,i.
e.
aconnectionwithlowperformancethathasareduceduplinkbandwidthofonly25Kbit/secandadownlinkbandwidthof100Kbit/sec.
–Batteryconsumption;eachpeerentersthenetworkwitharandomfractionofbatterycapacity;thisvalue(QB)isperiodicallydecreasedaccordingtoelapsedtimeandtothenumberofoutgoingactiveconnections:QB=f(QB,kout)(9)Ingeneral,weassumedthatwithachargeof100%,amobilepeercanstayactivefor1hourand30minutesofplayback.
Whentheremainingbatteryfallsbelowacertainthreshold,thenodeleavesthesystemwithaconsequentrecongurationofthepeerswhichweredirectlyconnectedtoit.
96.
59797.
59898.
59999.
5100010002000300040005000600070008000ContinuityIndexSimulationTimeContinuityIndexFitnessProtocolNoFitnessProtocol05101520253035404550010002000300040005000600070008000UnstableNodeSimulationTimeUnstableNodeNoFitnessProtocolFitnessProtocolFig.
3.
ComparisonofContinuityIndexandUnstableNodesintherstscenario152M.
Picone,M.
Amoretti,andF.
ZanichelliInitiallyweconsideredasimpledata-drivenprotocol[8]inwhicheachnodehasasinglesegmentprovider.
Wecomparedthepeerselectionstrategybasedontheevaluationofthetnessvalueassociatedtoeachnodewithasimplerandomselectionofthesupplierforthemediastream.
TherstanalyzedscenariohasahigherpercentageofPCnodesandMobile-WiFiandconsequentlyalargeamountofavailableresources.
Applyingthet-nessstrategyallowstoselectthemostecientsuppliers,reducestheloadonthesourcenodeandpreventstheundesirableeectofhavingMobileWi-Fiand3Gnodesthatbecomesuppliersforsubgroupsofnodes.
Obtainedsimulationresults(gure3)showasmallpercentageoflinkstonodeswithpoorperfor-mance,thusincreasingtheContinuityIndex(CI),whichisthepercentageofvideosegmentsreceivedbeforetheplaybackdeadline,andreducingthenumberofunstablenodesinthesystem(i.
e.
nodeswithaCIunder90%).
ThesecondscenarioshowsahighernumberofMobile3Gcomparedtoothertypesofnodesandimpliesareducedavailabilityofresources.
Inthissituation,whennoevaluationofthenodesisperformed,aremarkablereductionoftheCIcanbenoticedalongwithasignicantincreaseofunstablenodesinthesystemasingure4.
Thereasonisthatselectednodesareinecientsuppliers7580859095100010002000300040005000600070008000ContinuityIndexSimulationTimeContinuityIndexFitnessProtocolNoFitnessProtocolFig.
4.
ComparisonofCIobtainedwithahighernumberof3GNodesTable2.
CoolStreamingsimulationsparametersNameValueTotalNodes712Num.
OfPC-Node396Num.
OfMobileWiFi-Node141Num.
OfMobile3G-Node179SimulationMaxTime110000VTRealMaxTime1:30hAnEvaluationCriterionforAdaptiveNeighborSelection153fortheirsubgroupsofpeersinthesystem.
Thischoiceinvolvesareductionofperformanceandaprogressivedeteriorationofthesystem.
Instead,theuseofthetness-basedstrategyreducesthesenegativeeects,tryingtoexploitasmuchaspossibletheabilityofthemostecientpeers.
TheincreaseoftheCIwithouttnesscriterionafter6000VTisaconsequenceofthedisconnectionofmobilenodesthatinthisscenariohavetomanagealotofconnections,whichcauseafastdecreaseofthebatterylevel.
Attheendofthesimulationtherearelessmobilenodescomparedwiththebeginningofthesimulaton,forwhichPCnodescanmanagethedistributionofthemediastreamwithbetterperformance,intermsofcontinuityindex.
Obviouslytheimpactofmobilenodesisveryimportantinasystemwithonlyonesegmentproviderperpeer,wherethedisconnectionorthepresenceoflow-performingintermediateprovidersreadilyaectsthesystem'sperformance.
More95.
59696.
59797.
59898.
59999.
5100020000400006000080000100000120000ContinuityIndexVTContinuityIndexCoolStreamingNoFitnessProtocolFitnessProtocol051015202530354045020000400006000080000100000120000NumberofUnstableNodesVTUnstableNodesNoFitnessProtocolFitnessProtocolFig.
5.
UnstableNodesandContinuityIndexinCoolStreaming,withandwithouttness154M.
Picone,M.
Amoretti,andF.
Zanichellicomplexprotocols,involvingmultiplesuppliersperpeer,aremorerobustagainsttemporarylossofsuppliers,becausetheresponsibilityoftheowisdistributedamongmultiplenodes.
Forthisreason,thesecondphaseofanalysisofthet-nesscriterionhasbeenfocusedtoitsapplicationtothepopularmulti-supplierprotocolforP2Plivestreaming,i.
e.
CoolStreaming[5].
Thisprotocolassignsasetofsegmentproviderstoeachpeer,withanevaluationstrategythatanalyzestheavailablebandwidthandresourcesofnodeswithwhichtheotherpeerscomeintocontact.
Thisvalueisusedfortheselectionoftheappropriateproviderforthemedia-stream.
Forthereasonsexplainedpreviouslyinanetworkthatshowsheterogeneityintermsofdevices,amoredetailedandspecicstrategyisneces-sary,forwhichweaddedthetness-basedstrategytotheCoolStreamingprotocol.
SimulationshavebeencarriedoutusingDEUS,usingthesametypesofnodesandeventsthathavebeenusedfortheanalysisofthesimpledata-drivenprotocol.
TheevaluationisbasedonthecharacteristicslistedinTable2.
Inthissecondanalysisthetnessstrategyreplacesonlythebandwidthstrat-egyofCoolStreaming,notthestreamsegmentsevaluationthatistomuchtiedtogetherwiththeprotocolandcannotbechanged.
TheresultsshowthatalsoincaseofarealsystemlikeCoolStreamingwithahighheterogeneousnetworkintermofkindsofdevices,connectionsandbehaviourscanexperienceanincreaseofperformanceusingthetnessselectionstrategy.
Figure5showsthatduringthesimulatedtimethereisanincreaseofCIandadecreaseinthenumberofunstablenodes.
Neartheendofthesimulation,wherethenumberofdevicesandtheglobalbatterylevelarereduced,thereisstillasmallgroupofhighperfor-mancenodes.
Forthisreasonthereisareducedrangeofgoodpotentialproviders02040608010005101520ContinuityIndexVTMobile3GNodeStreamForwardingOut-Degree-No-FitnessStreamForwardingOut-DegreeDistributionAverage02040608010005101520ContinuityIndexVTMobileWifiNodeStreamForwardingOut-Degree-No-FitnessStreamForwardingOut-DegreeDistributionAverage02040608010005101520ContinuityIndexVTPCNodeStreamForwardingOut-Degree-No-FitnessStreamForwardingOut-DegreeDistributionAverageFig.
6.
Distributionofthenumberofoutgoingactiveconnections(kout)foreachpeer,forthestandardCoolStreamingprotocolAnEvaluationCriterionforAdaptiveNeighborSelection15502040608010005101520ContinuityIndexVTMobile3GNodeStreamForwardingOut-Degree-FitnessStreamForwardingOut-DegreeDistributionAverage02040608010005101520ContinuityIndexVTMobileWifiNodeStreamForwardingOut-Degree-FitnessStreamForwardingOut-DegreeDistributionAverage02040608010005101520ContinuityIndexVTPCNodeStreamForwardingOut-Degree-FitnessStreamForwardingOut-DegreeDistributionAverageFig.
7.
Distributionofthenumberofoutgoingactiveconnections(kout)foreachpeer,forCoolStreamingprotocolenhancedwithtness-basedstrategyandthedierenceofperformancebetweenthetwoapproachesislessevident.
Ingures7and6wecomparethestreamforwardingout-degreedistribution(atsimulation'svirtualtime50000)withoutandwithtness-basedstrategyforeachkindofnode.
Theyshowthatwiththeadaptivestrategythereisalittleadvantageandthatnodeswithhighercapacitiesareusedmorethannodeswithlowercapacities.
6ConclusionsInthisworkwehaveproposedanevaluationcriterionforneighborselectionindistributedsystemsbasedonthepeer-to-peerparadigm,wherepeerprocessesarehostedbyheterogeneousdevices,includingmobiledevices.
Ourapproachisbasedontheconceptoftness,whichallowstocompareresourceproviderswithdierentcharacteristics.
Wehaveshownthatthegeneralapproachcanbespe-cializedtoaparticularlychallengingproblem,i.
e.
P2PInternetLiveStreaming.
Simulationresultsareencouraging,showingthatthetness-basedstrategycanbeappliedtoexistingprotocols,witheectiveperformanceimprovements.
Asfuturework,wewilluseourstrategyinconjunctionwithtechniquesandprotocolsthathavebeenspecicallydenedforconstraineddevices.
Forexam-ple,intheeldofP2PInternetLiveStreamingwewillconsiderlayeredvideocoding,andmultipledescriptioncoding.
Thosetechniquesallowtoadaptmul-timediacontenttodierentlyfeatureddevices.
Moreover,wewillinvestigatetheapplicationofthetness-basedstrategytootherP2Papplications,suchasdata-intensivecomputingandmulti-playeronlinegaming.
156M.
Picone,M.
Amoretti,andF.
ZanichelliReferences1.
BitTorrentocialsite,http://www.
bittorrent.
org2.
Gnutellaocialsite,http://www.
gnutella.
com3.
KaZaAocialsite,http://www.
kazaa.
com4.
Amoretti,M.
,Agosti,M.
,Zanichelli,F.
:DEUS:aDiscreteEventUniversalSim-ulator.
In:Proc.
ofthe2ndICST/ACMInternationalConferenceonSimulationToolsandTechniques(SIMUTools2009),Roma,Italy(March2009)5.
Keung,G.
Y.
,Xie,X.
Z.
S.
,Li,B.
:CoolStreaming:Design,Theory,andPratice,pages.
IEEETransactionsonMultimedia9(8),1661–1671(2007)6.
Ahmed,T.
,Mushtaq,M.
:P2P-basedMobileIPTV:ChallengesandOpportuni-ties.
In:Proc.
ComputerSystemsandApplications,AICCSA2008.
IEEE/ACSInternationalConference,pp.
975–980(2008)7.
Takizawa,M.
,Nemati,A.
G.
,Enokido,T.
:AMulti-SourceStreamingModelforMobilePeer-to-Peer(P2P)OverlayNetworks.
In:Proc.
The28thInternationalConferenceonDistributedComputingSystemsWorkshops,pp.
18–23(2008)8.
Fourmaux,O.
,Silverston,T.
:SourcevsData-drivenApproachforLiveP2PStreaming.
In:Proc.
InternationalConferenceonNetworking,InternationCon-ferenceonSystemsandIntenationConferenceonMobileCommunicationsandLearningTechnologies,pp.
99–104(2006)9.
Kwong,K.
W.
,Tsang,D.
H.
K.
:Buildingheterogeneouspeer-to-peernetworks:Pro-tocolandanalysis.
IEEE-ACMTransactionsonNetworking12(2)(April2008)10.
Koo,S.
G.
M.
,Kannan,K.
,GeorgeLee,C.
S.
:Onneighbor-selectionstrategyinhybridpeer-to-peernetworks.
FutureGenerationComputerSystems22,732–741(2006)11.
Abraham,A.
,Yue,B.
,Xian,C.
,Liu,H.
,Pant,M.
:Multi-objectivePeer-to-PeerNeighbor-SelectionStrategyUsingGeneticAlgorithm.
In:Aluru,S.
,Parashar,M.
,Badrinath,R.
,Prasanna,V.
K.
(eds.
)HiPC2007.
LNCS,vol.
4873,pp.
443–451.
Springer,Heidelberg(2007)