Citation
Enhanced Distributed Multimedia Services Using Networked Storage Systems

Material Information

Title:
Enhanced Distributed Multimedia Services Using Networked Storage Systems
Creator:
KIM, EUNSAM ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Arbitration ( jstor )
Bandwidth ( jstor )
Broadcasting industry ( jstor )
Buffer storage ( jstor )
Caching ( jstor )
Multimedia materials ( jstor )
Scheduling ( jstor )
Simulations ( jstor )
Streaming ( jstor )
Television programs ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Eunsam Kim. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
8/31/2009
Resource Identifier:
439083736 ( OCLC )

Downloads

This item is only available as the following downloads:


Full Text

PAGE 4

IwouldliketothankallpeoplewhoprovidedmewiththeirhelpduringmyPh.Dyears.Firstofall,Iwouldliketothankmyadvisor,Dr.JonathanLiuforhissupportandencouragement.Withouthisconstantinspiration,thisdissertationwouldnothavebeenpossible.Thediscussionwithhimonanytopicalsomademepleasantandrelaxed.Iamalsogratefultomysupervisorycommitteemembers,Dr.ShiganChen,DouglasDankel,YeXia,andLiuqingYangfortheirinvaluablesuggestionsandcomments.Inaddition,IwouldthankallmyKoreanfriendsandfamilieswhosharedhappymemoriesinGainesville.Iappreciatemyparents,HanjeonKimandKyesunBae,whohavesupportedandprayedformeformywholelife.Ialsoappreciatemyparents-in-law,YongbooYooandBonsooKoo,whohaveprovidedmyfamilywithsincerelove.IwouldalsoliketothankallfamilymembersinKorea.Lastly,Iwouldespeciallyliketothankmywife,JungheeYoo,forherpatience.ShehasalwayssupportedmewhateverIdoandwhateverhappenstoher,takingcareoftwochildrenduringmanyyearsinGainesville.Iwouldalsoliketothankmychildren,JiinandJisung,whoalwaysmakemesmile. iv

PAGE 5

page ACKNOWLEDGMENTS ............................. iv LISTOFTABLES ................................. vii LISTOFFIGURES ................................ viii ABSTRACT .................................... xi 1INTRODUCTION .............................. 1 1.1On-demandVideoServicesusingTime-awarePrefetching ...... 2 1.2On-demandVideoServicesusinganIntegratedPrefetching/Caching 5 1.3High-qualityStreamingServicesforTVContentSharing ...... 9 1.4OutlineofthisDissertation ...................... 12 2BACKGROUNDANDRELATEDSTUDY ................ 14 2.1FC-AL-basedMultimediaServers ................... 14 2.2BueringTechniquesforOn-demandVideoServices ......... 17 2.3BroadcastTVContentDistribution .................. 18 3ON-DEMANDVIDEOSERVICESUSINGTIME-AWAREPREFETCH-ING ...................................... 20 3.1PrefetchingScheme ........................... 20 3.1.1ImpactofSCSIUnfairBusArbitration ............ 20 3.1.2DeterminingFC-ALAccessAcquisitionOrderamongDisks . 21 3.1.3Analysis ............................. 23 3.1.3.1Estimatingthenumberofconcurrentstreamswithprefetching ...................... 25 3.1.3.2EstimatingthenumberofblockstransmittedtoFC-ALperarbitration ................ 28 3.1.4CacheReplacementPolicy ................... 30 3.1.5SupportingMultipleLoops ................... 31 3.2PerformanceResultsandValidation .................. 33 3.2.1NumberofConcurrentStreamsvs.NumberofDisks ..... 34 3.2.2DiskCacheUtilizationandHitRatio ............. 37 3.2.3BlockSizevs.CacheSize .................... 38 3.2.4OverheadofFC-ALArbitration ................ 39 3.3Summary ................................ 40 v

PAGE 6

............................ 42 4.1DeterminingDynamicThresholdIntervals .............. 42 4.1.1CacheRequirementforPrefetching .............. 42 4.1.2TradingSavedDiskBandwidthforCacheSize ........ 44 4.1.3DynamicThresholdIntervals .................. 47 4.2IntegratedPrefetching/Caching(IPC)Algorithms .......... 47 4.2.1SchedulingNewStreams .................... 48 4.2.2DeletingExistingStreams ................... 50 4.2.3ReplacementPolicy ....................... 54 4.2.4NetworkTransmissionScheduling ............... 55 4.3PerformanceEvaluation ........................ 56 4.3.1EectoftheNumberofDisks ................. 60 4.3.2EectofCacheSizeperDisk .................. 61 4.3.3EectofDiskBlockSize .................... 63 4.3.4EectofAccessSkew ...................... 67 4.3.5EectofInter-arrivalTime ................... 67 4.4Summary ................................ 69 5HIGH-QUALITYSTREAMINGSERVICESFORTVCONTENTSHAR-ING ...................................... 71 5.1OverallArchitecture .......................... 71 5.1.1StreamingNetworks ....................... 71 5.1.2SystemComponents ....................... 72 5.2SystemModel .............................. 75 5.3SchedulingSchemes ........................... 78 5.3.1MatchingaCommunicationPair ................ 80 5.3.2Switching/ReplicationforLoadBalancing ........... 82 5.4ExtensiveTime-Shifting ........................ 86 5.4.1ChannelAssignmentforTime-Shifting ............ 86 5.4.2BueringforContinuousStorage ................ 87 5.5ExperimentalResults .......................... 92 5.5.1EectofSchedulingSchemes .................. 94 5.5.1.1Matchingacommunicationpair ........... 94 5.5.1.2Switching/replication ................. 96 5.5.2Eectofextensivetimeshifting ................ 99 5.5.2.1Time-shiftinghoursvs.hitratio ........... 100 5.5.2.2Buerrequirementforprogramstoragecontinuity . 101 5.6Summary ................................ 102 6CONCLUSIONSANDFUTUREWORK .................. 104 REFERENCES ................................... 106 BIOGRAPHICALSKETCH ............................ 113 vi

PAGE 7

Table page 3{1Denitionofsymbolsusedforourprefetchingscheme ........... 24 3{2Summaryofsimulationparametersforourprefetchingscheme ...... 33 4{1DenitionofimportantsymbolsforIPC .................. 43 4{2TheimpactofNiandNponMp(Ni;Np)intermsofthenumberofblocks 45 4{3SummaryofvaryingsimulationparametersforIPC ............ 57 4{4SummaryofdiskandvideotitleparametersforIPC ............ 57 5{1Denitionofsymbolsusedforourschedulingschemes ........... 80 5{2SummaryofsimulationparametersforTVcontentsharing ........ 92 vii

PAGE 8

Figure page 1{1Networkedstoragearchitecture ....................... 8 3{1OverallarchitectureoftheFC-AL-basedvideoserversasaresidentialservicegateway. ................................ 21 3{2ExampleofunfairFC-ALaccessorder ................... 22 3{3Anexampleofcacherequirementforholdingprefetchedblocksinthecache:Ni=3;Np=112 .......................... 26 3{4Supportingdualloops.(a)Exampleofalooppriorityassignmentindualloops,(b)Exampleofloopacquisitionorderandintervalinasingleloopconguration,(c)Exampleofloopacquisitionorderandintervalinadualloopconguration ............................ 31 3{5Numberofconcurrentstreamsvs.numberofdisks ............ 35 3{6Disktimeratioanddiskbandwidthutilization ............... 36 3{7Diskcacheutilization ............................. 37 3{8Blocksizevs.cachesizeintermsofthenumberofconcurrentstreams . 38 3{9OverheadofFC-ALarbitrations ....................... 39 4{1Fourpossiblecasesoccurringwhendeletingexistingstreams ....... 53 4{2Networktransmissionorderandtimeofeightdisksaccordingtotheirpriorities. ................................... 55 4{3Numberofconcurrentstreamsvs.numberofdisks. ............ 58 4{4NpsvaluesofIPCandPOvs.numberofdisks. .............. 59 4{5IntervalvaluesofIPCandCOvs.numberofdisks. ............ 59 4{6CacheportionsofIPCvs.numberofdisks. ................. 59 4{7Numberofconcurrentstreamsvs.cachesizeperdisk ........... 61 4{8IntervalvaluesofIPCandCOvs.cachesizeperdisk ........... 62 4{9NpsvaluesofIPCandPOvs.cachesizeperdisk ............. 62 viii

PAGE 9

............... 62 4{11Numberofconcurrentstreamsvs.diskblocksize ............. 64 4{12ImprovementpercentageofIPCvs.diskblocksize ............ 64 4{13Numberofconcurrentstreamsvs.accessskew ............... 66 4{14IntervalvaluesofIPCandCOvs.accessskew ............... 66 4{15CacheportionsofIPCvs.accessskew ................... 66 4{16Numberofconcurrentstreamsvs.inter-arrivaltime ............ 68 4{17IntervalvaluesofIPCandCOvs.inter-arrivaltime ............ 68 4{18CacheportionsofIPCvs.inter-arrivaltime ................ 68 5{1OverallTVcontentdistributionarchitecture. ............... 73 5{2Astatediagramforabasicstand-alongPVR. ............... 76 5{3Representationofaspecicprogramposition. ............... 78 5{4Anexampleofmatchingacommunicationpair. .............. 81 5{5AnexampleofswitchingtransmissionresponsibilitybetweentwoPVRs. 83 5{6Maximumdelayinswitchingtransmissionresponsibilityofatime-shiftedprogrambetweentwoPVRs. ......................... 88 5{7Eectofmatchingacommunicationpair. .................. 93 5{8Thedataportiontransmittedinfull-duplexmodeusingourmatchingscheme. .................................... 94 5{9PercentageofmatchingsformedindividuallyfromRDiandRIi. ..... 95 5{10Eectofswitchingplaybackrequests. .................... 96 5{11Improvementinnumberofreplicationsbyswitching. ........... 97 5{12AverageFC-ALbandwidthsavedbyswitching. .............. 97 5{13Numberofswitchingsvs.replications. ................... 98 5{14TheaccumulatedaccessfrequencyaccordingtoZipfdistribution. .... 99 5{15Time-shiftinghoursvs.hitratio. ...................... 100 5{16DelayinswitchingtransmissionresponsibilityofaprogrambyFC-ALtransmissionsize. ............................... 101 ix

PAGE 10

.. 102 x

PAGE 11

Recentadvancesinstorageandnetworkingtechnologymakeavarietyofenhanceddistributedmultimediaservices,suchashigh-qualityvideo-on-demandandP2Pcontentsharing,realistic.Toprovidethoseservices,itisessentialtosupportbothsucientI/Obandwidthandhigh-qualitystreamingforretrievingmultimediadataanddeliveringtothenetworkontime.Inthisdissertation,weemploynetworkedstoragesystems,especiallyFibreChannelArbitratedLoop(FC-AL),toconnectthesystemcomponents.Duetoitsseveraladvantages,theFC-ALissuitableforprovidinghigh-qualitymultimediastreamingservices. Theprimaryfocusofthisdissertationishowtosupportthefollowingen-hanceddistributedmultimediaservicesusingnetworkedstoragesystems:on-demandvideoservicesandhigh-qualitystreamingservicesforTVcontentsharing.Intherstpartofthisdissertation,weproposeanecientprefetchingschemebyfurtherutilizingnetworkeddiskcaches.Todevelopthescheme,weinvestigatethecharacteristicsofFC-AL'sfairnessarbitrationalgorithm.Theproposedschemeisnotonlyabletoincreasethetotalnumberofconcurrentstreamssignicantlyby xi

PAGE 12

Inthesecondpartofthisdissertation,wefurtherextendthebueringschemebytakingadvantageofbothprefetchingandcaching.Thebueringscheme,calledIntegratedPrefetching/Caching(IPC),determineswhethertoprefetchortocacheeachstreambasedondynamicthresholdvalues.Wealsopresentthealgorithmstoschedulenewstreamsanddeleteexistingstreams.TheexperimentalresultsshowthatIPCincreasesthenumberofconcurrentstreamssignicantly,comparedtowheneitherprefetchingorcachingisemployedseparately. Inthethirdpartofthisdissertation,wedevelopahigh-qualitycontentsharingarchitectureusingnetworkedstoragesystems.EventhoughPeer-to-Peer(P2P)networkingtechnologyprovidesanalternativeTVcontentdistributionchanneltounidirectionalbroadcasting,itisdiculttosupportHD-qualityTVcontentstreamingontheInternetduetothecurrentlimitedbroadbandbandwidth.WethuspresentanovelHD-qualitystreamingarchitectureforreal-timebroadcastcontentsharing,combinedwithnetworkedPVRsandhigh-speednetworks.Wealsoproposetwoecientschedulingschemestofurtherutilizenetworkanddiskbandwidth. xii

PAGE 13

Withrecentadvancesincomputationalpowers,videocompression,storage,andnetworkingtechnology,avarietyofenhanceddistributedmultimediaservices,includinghigh-qualityvideo-on-demand,P2Pcontentsharing,interactiveTV,andvideo-conferencing,becomerealistic.Therapidlyincreasingnumberofhigh-speednetworkedhomesalsoenablestheInternetserviceproviderstooersuchenhancedservices,notlimitedtohigh-speedbroadbandconnections.Toprovidethoseservices,itisnecessarytosupportbothsucientI/Obandwidthandhigh-qualitystreamingforretrievingmultimediadataanddeliveringtothenetworkontime. Tostablysupportsuchhigh-qualitymultimediaservices,weemploynetworkedstoragesystems,especiallyFibreChannelArbitratedLoop(FC-AL),toconnectsystemcomponentssuchasstoragedevicesandconsumerdevicesinsmallareas[1].TheFC-ALisoneoftheleadingnetwork-attachedstoragesystemsforStorageAreaNetworks(SANs)[2{4].TheFC-ALisgainingpopularityasanalternativetoexpensiveswitches.Duetoitsseveraladvantagesincludingfull-duplextransmis-sion,highbandwidth,andafairnessarbitrationalgorithm,FC-ALissuitableforprovidinghigh-qualitymultimediastreamingservicesatlowcost[5,6]. Inthisdissertation,weshowhowtoecientlysupportenhancedmultimediaservicesusingnetworkedstoragesystems.Theprimaryfocusofthisdissertationisthreefold:(1)todevelopaprefetchingschemetofurtherutilizethediskcacheusingthecharacteristicsofFC-AL'sfairnessarbitrationalgorithm,(2)todevelopanintegratedprefetching/cachingscheme(IPC)tosimultaneouslytakebenetofbothprefetchingandcachingusingdynamicthresholdvaluessothattheutilizationofbothcachespaceanddiskbandwidthcanbemaximized,and(3)tosuggesta 1

PAGE 14

novelHD-qualitystreamingarchitectureforreal-timebroadcastcontentsharingbetweenconsumerdevicessuchasPersonalVideoRecorders(PVRs),anddevelopecientschedulingschemesandtime-shiftingpolicy. Thestoragesystemisparticularlyimportantsincevideoblocksneedtoberetrievedontimeforallconcurrentstreams.OursystemarchitectureintherstpartofthisdissertationisbasedontheFC-ALthatdirectlyconnectsseveralstreamingserversandmanydisks.Thestreamingserversareresponsibleforend-to-end,real-timestreamingservices.EachdiskconsistsofanembeddedcontrollerforadiskI/Oprocessorandalargecachememoryforperformanceimprovement.SincethestoragesystemsaredirectlyaccessibletotheFC-AL,theavailabilityofthenetworkedstoragesystemisnotdependentontheserversconnectedwiththeexclusiveSCSIinterface.Ideally,anyhostssuchasstreamingserversattachedtotheFC-ALcanaccessallFC-ALdisks.Thus,thoughsomeofthehostsmaybedown(e.g.,duetoDenialofServiceattacks),therestofthehosts/serverscanstillaccessthestoragesystem. Byexpandinghomegateways,aresidentialservicegatewayenablesenhancedservicesforthousandsofclientsinasmallarea.ThestreamingserversbasedontheFC-ALallowtheservicegatewaytoconnectexternalserviceproviderstointernalclients.TheFC-ALcanstablyprovidegoodqualityvideostreamingservicesto

PAGE 15

thousandsofclientsintheinternalnetworks.Itsscalabilityalsomakesitfeasibletoservetheservicegatewaybyfrequentlyupdatingthevideotitlesfromtheserviceprovidersaccordingtoclientdemands.Indenselypopulatedareassuchasmetropolitanareas(e.g.,thehigh-risingapartmentbuildingsinNewYorkCityinUSA,SeoulinKoreaandTaipeiinTaiwan),itcanbemuchmorecostecient. Avideostreamingserverisacomplexsystembecausemanydesignfactorsaecttheperformanceinvariousdegrees.Forinstance,tosupportmanyconcur-rentaccesses,theloadsamongdisksshouldbedistributedasevenlyaspossiblewithleallocationmethods.ThiskindofdesignisessentialnomatterwhetherdisksarebasedonSCSIorFC-AL.However,veryfewpreviousstudiesactuallyaddressedtheperformancegainobtainingfromanend-to-end,on-demandvideostreamingsystemthatincludestheclients,network,streamingservers,FC-ALs,anddisks.Usingourend-to-endsimulationstocombineallthecomponents,wehaveobservedthatthissystemperformsbetterthanSCSI-basedsystemsduetoseveraladvantagesincludinghighdatatransferrate,fairnessarbitrationalgorithm,andfaulttolerance[8],butthereisstillroomforperformanceimprovementtoreachthepossibleoptimalityofthesystemdesign. TheSCSIemploysaprioritizedarbitrationprotocolforbusaccess.Whenmultipledevicescompeteforbusaccess,thedevicewiththehighestpriorityalwayswinsthearbitration.Thisunfairarbitrationcausessignicantvariationoflatenciesdependingondisks'priorities,resultinginperformancedegradation.TheFC-ALalsoemploysaprioritizedarbitrationprotocolfortheloopaccess.However,topreventasituationwherethelowerprioritydevicesmayexperiencestarvation,theFC-ALdenesafairnessarbitrationalgorithmthatenablesallthedevicesparticipatinginthearbitrationtohaveanequalopportunitytohaveloopaccess.TheFC-ALfairnessarbitrationalgorithmreducesthevariationofthelatenciesforloopaccessamongdisksduetotheequalloopaccessopportunity.Asaresult,

PAGE 16

thefairnessalgorithmenablesallthediskstoutilizetheircachestoasimilardegree(inthisdissertation,diskcacheutilizationdenotestheaverageratioofdiskcacheblockstobetransmittedtotheFC-ALatleastoncebeforebeingreplacedinthediskcache).Furthermore,thecacheutilizationofeachdiskisverylowfortworeasons.Oneisthatthediskstransferblocksonebyoneoverthehigh-speedFC-AL.Thus,therateofsendingablock Foron-demandvideostreamingservers,prefetchingisabetterschemetoutilizethediskcachethantheconventionalcachingforthereasonmentionedabove.Moreover,prefetchingcanreducediskseektimesignicantlybyreadingblockscontiguouslystoredtogether,whichhasbeenacriticalsystemperformancebottleneck.Itisclearthatasystemcanmaximizethetotalnumberofconcurrentstreamswhenthenumberofprefetchedblocksismaximal.However,thenumberofprefetchedblocksshouldvaryaccordingtothediskblocksize,constrainedbythediskcachesize.Weanalyzehowtodeterminethemaximumnumberofprefetched

PAGE 17

blocksdependingondiskblockandcachesize.Inaddition,toguaranteethediskcachespaceforprefetchedblocks,weadjusttheincomingrateofsomeblocksfromthedisksandtheoutgoingratetotheFC-AL.Inotherwords,weshowhowtondtheoptimalnumberofblockstransmittedovertheFC-ALperarbitrationwhilereadingseveralblocksbelongingtoonestream,includingonenon-prefetchedblockandsubsequentprefetchedblocks.WealsodescribehowtoestimatetheFC-ALaccessacquisitiontimeofeachdiskdependingonthenumberofdiskswithoutmodifyingthefairnessarbitrationalgorithm.Bydoingso,wecanmaximizediskcacheutilizationsincethediskscanpredictwhencacheblockswillbereplacementcandidatesaccordingtoourcachereplacementpolicy.Inaddition,wedescribethecachereplacementpolicytotakefulladvantageofthesequentialaccesspatternofvideoles,andexplainhowtosupportmultipleloops.Byanalysisandsimulationexperiments,weshowthatourprefetchingschemecannotonlyincreasethetotalnumberofconcurrentstreamssignicantlybyreducingthediskseektime,butitcanalsofurtherutilizetheFC-ALbyreducingtheoverheadofarbitration. Diskaccesstimerequiredforreadingonediskblockinvolvesseektime,rotationallatency,anddatatransfertime.Withprefetching,datablockscanbereadaheadintothecachebeforetheyareactuallyrequested.Thus,prefetchingcanreduceseektimeandrotationallatencybyreadingtogetherseveralblockscontiguouslystoredonadiskatatime.Moreover,cachingcaneliminatethewholediskaccesstime,includingdatatransfertime,byreusingblocksalreadykeptinmemoryinsteadofreadingablockfromadisk.Theintervalcachingpolicyis

PAGE 18

knowntobemoreeectiveformultimediadatathantraditionalcachingpoliciessuchasLRU[12].Thispolicyimprovessystemperformancebycachingintervalsbetweensuccessivestreamsforthesamevideotile,takingadvantageofsequentialaccesspatternsofmultimediadata.Inthisdissertation,ourworkisbasedonthediskcacheandintervalcachingisemployedasadiskcachingpolicy. Theeortstocombinetwobueringtechniques,prefetchingandcaching,havefocusedonreducingthestalltimeofaprocessorcausedbythedierenceintrans-fertimesbetweenprocessorandmemoryfortheprocessingofconventionaldatatypes[15{17].However,thesetwobueringtechniqueshavebeenaddressedsepa-ratelywhenhandlingmultimediadatawiththecharacteristicsofpre-determinedsequentialaccesspatternsandlongduration.Inthisdissertation,weintegratethesetwotechniquesbasedonrequiredcachesizeanddiskbandwidthbyexploitingrelevantcharacteristicsofmultimediadata. Wehaveobservedthattheperformanceofeitherprefetchingorcachingcannotcontinuetoincreaseproportionallytothededicatedcachesize.Theperformanceofprefetchingbecomessaturatedatsomepointwhenincreasingthecachememorysize,eventhoughprefetchingcanincreasethenumberofconcurrentstreamssignicantlybeforethatpoint[6].Thereasonisthat,asthenumberofblockstobeprefetchedperstreamincreaseswithalargercachememory,thecachesizerequiredtoholdtheblocksuntiltheyaretransmittedtonetworksincreasesevenmorerapidly.Theperformanceofcachingalsodependslargelyontheintervallengthsbetweensuccessivestreamsforthesamevideotitle.Asalargercachememoryisemployed,theaveragelengthofintervalsshouldalsoincreasebecauselargerintervalsareadditionallycached.Thus,toutilizethecachespacemoreeciently,itisbettertoprefetchtheblocksthataretoolongtobecached.Accordingly,prefetchingtoomanyblocksorcachingtoolongintervalsmightquicklyexhausttheavailablecachespacethatwouldbeotherwiseusedtosupportmoreconcurrent

PAGE 19

streamswiththeothertechnique.Therefore,ifcarefulconsiderationisnotgiventoprefetchortocacheincomingstreams,eitherofthetwotechniquescandecreasetheeectoftheother,whichcannotmaximizethenumberofconcurrentstreams. Inthisdissertation,weproposeascheme,calledIntegratedPrefetch-ing/Caching(IPC),tosimultaneouslytakeadvantageofbothprefetchingandcachingwhendealingwithmultimediadata.TheIPCdynamicallydeterminesthresholdintervalssothatthesystemcandecidetoeitherprefetchorcacheeachnewstream.Toderivethethresholdintervals,werstanalyzehowmanycacheblocksarerequiredtoprefetchoneadditionalstreambasedonthecurrentnum-bersofprefetchedblocksandimmediateblocks(i.e.,non-prefetchedblocksthatrequireimmediatetransmissiontothenetworkwithinthesamecyclewhentheyarereadfromdisks).Wealsoconsideranotherimportantfactorfordeterminingthethresholdvalues,i.e.,thecachingeectofadditionallysavingdiskdatatransfertimecomparedtoprefetching.Bymeasuringthecachingeectintermsofthecachesize,weshowthatthecachesizeanddiskbandwidthcanbetradedforeachother.ThedynamicthresholdvaluesenableIPCtoecientlyutilizethecachespacedependingonthelengthsoftheincomingintervals.Iftheintervalsshorten,thenumberofcachedblocksincreases.Otherwise,thenumberofprefetchedblocksincreases. Wealsopresentthealgorithmstoschedulenewstreamsanddeleteexistingstreams.Tomaximizeutilizationofbothcachespaceanddiskbandwidthbycomplementingoneanotherwheneitherofthemisexhausted,IPCreduceseithertherequiredcachesizeordiskbandwidthbyadjustingthenumberofimmediateblocksandprefetchedblocks.Asaresult,IPCcancontinuetoincreasethenumberofconcurrentstreamsuntilbothsystemresourcesareusedup.Wealsoemploydierentreplacementpoliciesforprefetchingandcachingtoecientlyfreecachespace.

PAGE 20

Figure1{1. Networkedstoragearchitecture Inaddition,weshowthatitispossibletominimizethediskcacherequirementforschedulingnetworktransmissionofdatablockswithnetworkedstoragesystemsemployingfairnessarbitrationalgorithms.Storagesystemshavebecomeattachabledirectlytoanetwork.Storageareanetworks(SAN)[8,18,19]andnetwork-attachedstorages(NAS)[20]areleadingnetworkedstoragesystems[1].Theserversanddiskscanbeconnecteddirectlythroughhigh-speednetworks,therebyprovidingdatasharingandsystemreliabilitybetweensystemcomponentsattachedtoanetwork.Oncerequestsareadmitted,datacanalsobetransferreddirectlybetweendisksandclientswithouttheinterventionofserversasshowninFigure 1{1 .Sincethereisnocentralmainmemorytobuerdatablocksbeforetransmittingdatablockstoclients,however,eachdiskshouldbuerdatablocksusingitsowncachememory.ItisassumedthatoursystemisbasedonnetworkedstoragesystemssupportingfairnessarbitrationalgorithmssuchasFiberChannelArbitratedLoop(FC-AL)forSANs.Weshowthatthefairnessarbitrationalgorithmenablesdiskstoacquirenetworkaccessesinadeterminedordereveryxedtransmissioninterval.Asaresult,fewercacheblocksarerequiredtoschedulenetworktransmissionofdatablocksduringeachcycle.Infact,itmaybeverydiculttopredictwheneachdiskcanacquiretheaccessrightswithEthernetusingrandomaccessprotocolandSCSIusingprioritizedarbitration[8].

PAGE 21

Bysimulationexperiments,weshowthatIPConaveragesupports14.1,35.2,and80.4%moreconcurrentuserscomparedtothreeothercases:prefetchingonly,intervalcachingonly,andnobueringtechniqueused,respectively.Ascachesizeincreases,IPCneithersaturatestheoverallperformancenorrapidlyincreasestheaveragecachedintervallengths.Inotherwords,IPCcontinuestoincreasethenumberofconcurrentstreamsproportionallytothededicatedsystemresources.Thisispossiblebecauseitcanoerallthebenetsofbothprefetchingandcaching. UntilrecentlyusershavebeenquitepassiveaboutTVviewingaccordingtobroadcastingschedules.Theadventofset-topboxes(STBs)withlocalstoragedevicescalledPersonalVideoRecorders(PVRs),however,israpidlychangingusers'TVviewingpatterns,enablingpeopletorecordandtime-shiftliveTVprogramsintothesestoragedevices.Asaresult,peoplecanwatchwhattheywantwithouthavingtositdowninfrontofaTVataspecicbroadcasttime.Theycanthenretrieveprogramseasilybyrandomlyaccessinganyprogrampositionthroughnon-linearVCR-likeoperations.Inaddition,thenumberofcableandsatellitechannelshascontinuedtoincreaseandpayTVserviceshavealsoextendedtheopportunitiestochoosewhatpeoplewanttowatch.Nevertheless,thewayTVcontentsaredistributedhasstillbeenlimitedtounidirectionalbroadcasting. Ontheotherhand,Peer-to-Peer(P2P)networkingapplicationsthatcanshareavarietyofcontents,includingbroadcastTVprograms,aregainingpopularity.

PAGE 22

ItisclearthatPVRscaneasilyshareTVcontentsoverP2Pnetworksoncetheyareequippedwithbroadbandconnections.ThisisimportantbecausetheP2Pnetworkscanprovideanalternativecontentdistributionchanneltounidirectionalbroadcasting.Inotherwords,peoplecanhaveanotherwaytoobtainTVcontentsthattheywerenotabletorecordatthetimeofbroadcast.Atpresent,giventhelimitedbroadbandbandwidthinhomes,however,eitherhigh-qualityTVcontentsmustbetranscodedtolowerbit-rateformatsforreal-timestreamingornon-real-timeletransfershouldbesupportedasapracticalalternative.Moreover,someserviceprovidersarecurrentlytryingtointegratenetworkedPVRsintoVideo-on-Demand(VOD)servicesasbueredstoragessothattheycanovercomethecurrentbroadbandbandwidthinhomesthatlimitstrueVODservices.ThismeansthatP2Pnetworkingbasedoncurrentbroadbandconnectionscannotsupportconcurrentreal-timecontentdistributionformultipleHighDenition(HD)-quality(amountingto19.4Mbps)programs.Therefore,anovelcontentdistributionnetworkisrequiredtodoso. WeemploytheFC-ALasanHD-qualitystreamingnetworkforreal-timecontentdistribution[21,22].InourarchitectureforTVcontentsharing,theFC-ALdirectlyconnectsmanyPVRs,apoolofnetworkdisks,andacontentmanagingserver.Thepoolofnetworkdisksisusedtoreplicatepopularprogramsforload-balancingofplaybackrequestsinthesystemaswellasstoreallthetime-shiftingprogramssothateachPVRcansharethem. Inthisdissertation,wedescribeageneralsystemmodelfornetworkedPVRswithmultipletunersanddecoders.Basedonthemodel,weproposethefollowingtwoecientschedulingschemestofurtherutilizesystemresources:matchingacommunicationpairandswitching/replicationforsystemload-balancing.OneofimportantadvantagesoftheFC-AListhatdatacanbetransmittedinfull-duplexmode.TomaximallyutilizetheFC-ALbandwidthinbothdirections,wehave

PAGE 23

developedaschemetomatchacommunicationpair.Itisalsoacommonsituationthatmostplaybackrequestsareconcentratedonafewpopularprograms.Fromtheperspectiveofoverallsystemperformance,suchaskewedaccesspatternmaycauseseriousload-unbalancingproblemsbecausesomeoverloadedPVRscanbecomeperformancebottleneckseventhoughtherearestillmanydiskshavingsucientavailablebandwidth.Tosolvethisproblem,weswitchplaybackrequestsaswellasreplicatepopularprogram.ToreducereplicationstoasfewasnecessarysincereplicationsrequireadditionaldiskandFC-ALbandwidth,wemakediskbandwidthavailablefornewrequestsissuedtooverloadedsystemcomponentsbyswitchingthetransmissionresponsibilitiesforexistingrequeststolessloadedsystemcomponents.Thereplicationisnotperformeduntilswitchingbecomesimpossibleduetorequeststhatexceedthetotalprocessingcapabilityoftheentiresystem.Inaddition,wedescribehowtoassignchannelstoeachPVRforextensivetime-shiftinginanecientway.Wealsoevaluatetherequiredbuersizetocontinuestoringtime-shiftedprogramswithoutmissinganybroadcastdataeventhoughtime-shiftingresponsibilitiesfortheprogramsareswitchedbetweenPVRs.Mostimportantly,eachPVRcanappeartohavemanymoretunersthanitactuallydoes,soitcanrecordnumerousprogramsbroadcastondierentchannelsatthesametimeaslongasthenetworkanddiskbandwidtharesucient.ThisispossiblebecauseeachPVRcansimultaneouslyreceivealargenumberofliveprogramsfromthepoolofnetworkdiskswithonlysmalldelaysfromtheiractualbroadcastingtimes. Bysimulationexperiments,weshowtheeectoftheproposedschedulingschemesandextensivetime-shiftingpolicy.Wecanimprovetheperformanceby39%usingourmatchingschemeintermsofthenumberofprogramtransmissionrequestsovertheFC-AL.ItwasalsoapparentthateachPVRcouldprocess18%

PAGE 24

morediskaccessrequestswhenbothswitchingandreplicationwereusedthanwhenonlyreplicationwasused. Chapter3proposesaprefetchingschemetofurtherutilizethediskcachewithanalysis.ThemaximumnumberofprefetchedblocksandtheoptimalnumberoftransmissionblocksovertheFC-ALperarbitrationaredetermined.Thecachereplacementpolicyandhowtosupportmultipleloopsarealsoexplained.Theexperimentalresultsforourprefetchingschemewithdierentsystemparametersarepresented. Chapter4proposesabueringschemetosimultaneouslytakebenetofbothprefetchingandcachingusingdynamicthresholdvalues,calledIntegratedPrefetching/Caching(IPC).TheIPCschedulesnewstreamssothattheutilizationofbothcachespaceanddiskbandwidthcanbemaximized.Wealsodescribehowtoschedulenetworktransmissionofdatablocksbasedonnetworkedstoragesystemsemployingfairnessarbitrationalgorithms.Bysimulationexperiments,weshowthattheIPCincreasesthenumberofconcurrentstreamssignicantly,comparedtowheneitherofprefetchingorcachingisemployed. Chapter5proposesanovelhigh-qualitystreamingarchitectureforTVcontentsharing.Oursystemmodelispresented,includingstreamingnetworksandsystem

PAGE 25

components.Acoupleofschedulingschemes,matchingacommunicationpairforutilizingfull-duplextransmissionandswitching/replicationforsystemload-balancingareproposed.Ourtime-shiftingpolicytoextendtime-shiftinghourssignicantlyisalsosuggested.Theperformanceresultsshowtheeectoftheproposedschedulingschemesandextensivetime-shiftingpolicy. Chapter6describesconclusionsandourfutureworkintheresearcharearelatedtothisdissertation.

PAGE 26

TheFiberChannelisahigh-speedserialstorageinterfacethatisafamilyofANSIinterfacestandards[2,8,19,33].Itdenesapoint-to-point,switchandlooptopology,dependingonwhetherornotitusesaswitch.Itconsistsofasetofhierarchicalfunctions,fromFC-0toFC-4,thatdenesthephysicallayer,transmission,signaling,andframingprotocols,commonservicesacrossmultipleportsofanFCnode,manyapplicationinterfacesincludingSCSI,FDDI,andATM. TheFC-ALstandard[2]extendsFCtransmissionandsignalingfunctionstodenealooptopologythatdoesnotrequireswitches.Thelooptopologywas 14

PAGE 27

developedasanalternativetotheswitchdesigntoreducethecost,anditisbecomingapopularmeanstodirectlyconnecthostsystemstostoragedevices.Inaloop,allattacheddevicessharetheloopbandwidth.Beforetransferringdata,apairofdevicesshouldestablishaconnection.Onceaconnectionisestablished,nootherdevicescanusetheloop(i.e.,onlyonepairofdevicescancommunicateatatime).WedescribethemainfeaturesoftheFC-ALsuchasfull-duplextransmission,highbandwidth,afairnessarbitrationalgorithm,andaowcontrolmechanism. TraditionalI/Ochannelssupportonlyasinglepoint-to-pointtransferatatime.However,theFCcanprovidetwoconnectionsbetweenanytwodevicesonasingleFCcable.Sincetheseconnectionsarephysicallyseparate,theFCsupportsseparatecommunicationsoneachhalfofthecablesimultaneously,whichiscalledfull-duplextransmission.Withthiscapability,theFCcanbemuchmoreecientthantraditionalchannelsintermsofthroughputbypairingdevicessothattheycantransmitdatatoeachotheratthesametime.TheFCalsoallowsdatatobetransferredbetweentwonetworknodesatveryhighspeeds.Currently,fullduplex4Gb/sFCscantransferdataatupto800MB/s.Moreover,sincemostFCdevicesaredualported,theycandoublethebandwidthifadual-loopcongurationisused.Thetechnologyallowsadistanceofupto10kilometersbetweendevices. TheFC-ALdenesanarbitrationschemeasanaccessprotocolamongdevices.Thisisaprioritizedprotocolthatgivesloopaccesstothedevicewiththehighestpriorityport.Eachloopporthasauniqueaddressrepresentingitspriority.Portswithloweraddresseshavehigherpriorities.SomePrimitiveSignals,suchasARBx,OPNy,andCLS,areusedtocontrolaccesstotheloop.Whenaportintendstotransferdata,itmustarbitrateforloopaccessbysendingoutaPrimitiveSignalARBx,wherexrepresentstheaddressofthesendingport.Thesignalwilltraveltheentireloop.Theportwithhighestpriorityobtainsloopaccess.Bysendingout

PAGE 28

anOPNy,itshouldthenestablishaconnectionwithanotherportwhoseaddressisy.Onceaconnectionisestablished,twoconnectedportscantransferdatatoeachother.ThesendingorreceivingportcannishtheconnectionbysendingoutaCLStotheother.Thentheloopbecomesavailableforalltheports.Inaddition,withoneacquisitionofthearbitration,aportcanopenmorethanoneconnectioninarowwithoutreleasingtheloopaccesssothattheoverheadofarbitrationcanbereduced. Topreventtheportswithlowerprioritiesfromstarvationfortheaccesstotheloop,theFC-ALdenesafairnessarbitrationalgorithmthatallowsalltheportstohaveanequalopportunitytowinloopaccessaslongastheyareparticipatinginthearbitration.Anaccesswindowcanthusbedenedasthetimeperiodbetweenwhentherst(highestpriority)deviceamongallthedevicescompetingfortheloopaccesswinsthearbitrationandwhenthelast(lowestpriority)devicenotiesotherdevicesthattheloophasbecomeidle.Whenalltheportsparticipatinginthearbitrationhavehadanopportunitytoaccessthelooponce,anewaccesswindowisstarted.Inotherwords,whenaporthasarbitratedandwonaccesstotheloop,theportwillnotarbitrateagainuntilalltheotherportswinthearbitration.Anaccesswindowmayvaryinsize,dependingonthenumberofarbitratingports.Ingeneral,hostsystemsdonotusethefairnessalgorithmtosendoutI/Ocommandsasquicklyaspossible,butdisksusethefairnessalgorithmtosharetheloopbandwidthasfairlyaspossible.Thisfeatureisimportantinareal-timeenvironmentwherealltheI/Orequeststoanystoragedevicesshouldbeprocessedwithingivendeadlines. TheFC-ALemploysaowcontrolschemetoavoidreceivebueroverow[34].WheneachdevicejoinstheFC-AL,itexchangesthenumberofitsreceivebuerswithotherdevices.Onceacommunicationisestablished,thenumberofavailablereceivebuersisincrementedbyasmanycreditsasreceivedanddecrementedbyas

PAGE 29

manyframesastransmitted.Thus,thesenderscantransmitasmanyframesasthereceivebuerscanaccommodate. Whilememorypricehasbeendecreasingatafastrate,ithasbeennoteasytoimprovediskaccesstimesignicantlyduetothemechanicalaccesstimecomponentsofdiskI/Osuchasseektimeandrotationallatency.Toaddresssuchtrends,databueringtechniqueshavetriedtoreducediskbandwidthrequirementforI/Orequestsbykeepingdatablocksinmemoryforfutureaccesses.Inthisdissertation,weexploitdiskcachewithprefetching[6,9]andintervalcaching[10{14]techniquestotoimproveI/Operformance. Prefetchingisatechniquethatcomputersystemsuseinvarioussituationstoimproveperformance.CPUcacheprefetchingfetchesinstructionsanddatafrommainmemoryintotheCPUcachememorythatprovidesfastaccess[48,49]totheCPUregisters.Theprefetchedblockhasnotbeenreferenced,butisexpectedtohavehighprobabilitythatitwillbereferencedsoon.Thoughdiskcacheprefetchingusesthesimilarprinciple,thedierentreferencebehaviorfromapplicationscausesmuchlargermemorytoprefetchdiskI/O.The\localityofreference"hasbeenusedtoeectivelypredicttheaccesspatternfortheconventionaldata,whichisquiterandom[50,51].Intherstpartofthisdissertation,however,westudytheprefetchingforvideodata,whoseaccesspatternissequentialindividually,butthecomplexityishightosupportconcurrentstreams.

PAGE 30

Moderndiskssuchassmartdisks[52,53]includealargecachememory.[47]hasstudiedtheexploitationofcachememoryinahierarchy,suchashost,diskarraycontroller,anddiskdrives.Itshowedthatprefetchingtechniquesarepreferableforsmall-sizecachessuchasadiskdrivecache,whiletraditionalcachingissuitableforalarge-sizecachesuchasahostcache.Intherstpartofthisdissertation,wefocusontheimpactofdierentstorageinterfacesontheoverallsystemperformance. SomeliteraturedemonstratedthattheintervalcachingpolicyismoreeectivethantraditionalcachingpoliciessuchasLRUwhendealingwithmultimediadatawithsequentialaccesspatterns[12].Thispolicycachesintervalsbetweensuccessivestreamsforthesamevideotile,therebyincreasingthenumberofconcurrentstreams.Inthesecondpartofthisdissertation,theintervalcachingisemployedasadiskcachingpolicy.

PAGE 31

networks.However,theprojectdoesnotfocusonreal-timestreamingofhigh-qualitydata,butratheronprovidinginteroperabilitybetweendierentcontentandserviceprovidersandconsumerdevices. Therehasbeenresearchonthearchitecturefordistributing,storing,anddeliveringhigh-qualitystreamingTVcontentsoverIPnetworksusingportalsorproxies[57,58].However,thesehavearebasedonContentDistributionNetworks(CDNs),whichfocusontheunidirectionaldistributionfromserverstoclientsontheInternet,notonaP2Pcontentsharinginahigh-speednetwork. Inresearchonmultimediaserversystemscomposedofmultipleparalleldisks,theload-balancingproblemamongallthediskshasbeenanimportantperformanceissue.Toaddressthisproblem,somesystemshavestripedsegmentsofeachobjectacrossthediskstoevenlydistributeloads[35,59{64],whileothershavereplicatedpopularobjectstoprovideadditionaldiskbandwidthandmakethesystemscalableinadditiontothestripingscheme[39{41,65,66].Thus,itisassumedthatthesystemiscapableofplacingandreplicatingdataonmultiplediskseciently.However,theseschemescannotbeapplieddirectlytoourarchitectureforTVcontentsharingsinceitistheresponsibilityofeachPVRusertodeterminewhichprogramstostoreandremove.Moreover,eachPVRusuallycancontrolonlyoneoraverysmallnumberofdisksattachedtoit,aquantitythatisnotsucienttodistributeconcentratedloads. Therefore,weexploitthecharacteristicsofFC-ALtechnologyasameanstointerconnectstoragedevicesandconsumerdevicestoprovideadvancedmultimediastreamingservices.

PAGE 32

3{1 .Toprovideon-demandvideoservicesovertheFC-AL,theexistingsystemshaveworkedasfollows.TheserversperiodicallygenerateI/Orequeststodisks.Whenevereachdiskreadsoneblockintothediskcache,itparticipatesinarbitrationforFC-ALaccess.ThediskeventuallyacquiresFC-ALaccessandthentransferstheblocktothestreamservers.Inthisservicescenario,thecacheutilizationamongthedisksaresimilar,butstilllow,eventhoughtheFC-AL-basedsystemsprovidebetterperformancethanSCSI-basedsystems.Inthissection,weproposeandanalyzeaneectiveschemetoexploitavailable,butunuseddiskcachememorybyprefetching. 20

PAGE 33

Figure3{1. OverallarchitectureoftheFC-AL-basedvideoserversasaresidentialservicegateway. arbitrationdoesnotonlymakeitdiculttoutilizethediskcacheeciently,butitalsodegradestheperformanceofreal-timemultimediaservicesrequiringlargeblocksize.Therefore,therehavebeenfeweortstoexploitthediskcacheformultimediasystemsbasedonSCSIinterfaces. Figure 3{2 showsthatthesesituationsarecausedbyasignicantvariationindiskaccesstimethatconsistsofdatatransfertime,seektime,androtationallatency.SupposetherearevedisksattachedtotheFC-AL,whereDirepresents

PAGE 34

Figure3{2. ExampleofunfairFC-ALaccessorder thediskwithpriorityi(D1hasthehighestpriority).AssumethatD2nishesreadingablock,followedbyD3,D1,D5,andD4.TheD2winsthearbitrationrstsincethereisnoconictfortheloopaccessbythetimeitnishesreadingtheblockintothediskcache.WhileD2istransferringtheblock,D3andD1jointhearbitration.Accordingtotheprioritizedarbitrationrule,D1winsthearbitrationearlierthanD3andthenbothofdiskstransfertheblocks.AtthispointtherstaccesswindowendsbecauseanylowerprioritydiskssuchasD4andD5cannotnishreadingblocksandparticipateinthearbitrationbeforethehigherprioritydiskssuchasD1,D2,andD3nishtransferringblocks.Consequently,anotheraccesswindowstartsforD5.Likewise,thereshouldbeonemoreaccesswindowforD4.Finally,thevediskshaveacquiredFC-ALaccessintheorderofD2,D1,D3,D5,andD4,notindecreasingorderofpriority.Inaddition,somedisksmaynotacquiretheFC-ALaccessevenonceforacoupleofaccesswindowswhileothersaresuccessfulmorethanonceforthesamereason.Therefore,itisnoteasytopredicttheexactacquisitionorderandtimeofloopaccessforeachdiskwiththeconventionalapproach. WedescribehowtodeterminetheFC-ALaccessorderofeachdiskwithoutmodifyingtheexistingarbitrationprotocolwhileconsideringprefetchingeect.

PAGE 35

TheFC-ALaccessordercanbedeterminedsimplybyadjustingthetimewheneachdiskparticipatesinarbitrationinthefollowingway.ThestreamserversdonotusethefairnessarbitrationalgorithmsothattheycansendoutdiskI/Ocommandsasfastaspossibleatthebeginningofeachcycle.Assoonaseachdiskreceivesthecommand,itparticipatesinthearbitrationsinceithasalreadytheprefetchedblockstotransmit.Sinceallthedisksparticipateinthearbitrationatthebeginningofthecyclesimultaneously,thediskscanacquireFC-ALaccessindecreasingorderofthelooppriorityintherstaccesswindow.Onceallthediskstransferblocks,anotheraccesswindowstarts.Duringtheremainingaccesswindowsofthecycle,theFC-ALaccessorderwillalsobethesameifallthedisksjoinarbitrationassoonaseachaccesswindowstarts.Thediskscandetecttheinitiationofeachaccesswindowbylisteningtothenoticationmessagefromthelastdiskthathasachievedtheloopaccess.Byfollowingthisprocedure,itispossibletodetermineFC-ALaccessacquisitionorder.Additionally,wecandeterminethetimewheneachdiskacquiresFC-ALaccessbycalculatingthetransmissiontimeoftheblocksthatthediskswithhigherpriorities(i.e.,thedisksthathaveacquiredFC-ALaccessaheadofthedisk)havesentoutovertheFC-AL.Therefore,wecanmaximizediskcacheutilizationsincethediskscanpredictwhencacheblockswillbereplacementcandidatesaccordingtoourcachereplacementpolicy(wedescribereplacementpolicylater).

PAGE 36

Table3{1. Denitionofsymbolsusedforourprefetchingscheme Symbol Denition Unit block block block block block block block cachesizeusedforprefetching,givenNiandNp block KB ms ms ms MB/s disk ms ms ms ms MB/s

PAGE 37

Forthepurposeoftheanalysisofourprefetchingscheme,wehaveassumedthattheserverschedulestherequestswiththeworstseekandlatencytimes.Althoughtheestimatedresultsmaybeconservative,theperformancetrendandanalysisofourproposedprefetchingschemeandconventional(i.e.,non-prefetching)schemesstillcanbeclearlydemonstrated.Inaddition,toensurethatourschemeisofpracticaluse,weverifyinthenextsectionthattheexperimentalresultssupporttheanalyticalresultsbasedonaveragediskaccesstime.Table 3{1 liststheimportantsymbolsusedfortheanalysis. Note,however,thatthetotalnumberofsupportedblockspercycle(Nt)islimitedbythefollowingtworesourceconstraints:diskbandwidthandcachesize[6].LetNiandNpdenotethenumberofimmediateblocksandprefetchedblockstobereadpercycle,respectively.NtisthesumofNiandNpwhenonlyprefetchingisemployed,andNtisrstrestrictedbythediskbandwidth.Todeliverallstreamsontime,thetotaldiskaccesstimerequiredforNiandNpshouldnotexceedonecycletime.Eq. 3{1 describessuchdiskbandwidthconstraints,givenacycletime(C),diskblocksize(B),seektime(Dst),rotationallatency(Drl),anddatatransferrate(Dtr).Notethatthereisneitherseektimenorrotationallatencyfortheprefetchedblockssincetheblocksbelongingtoeachstreamarecontiguouslystored. Dtr+NpB Dtr(3{1)

PAGE 38

Figure3{3. Anexampleofcacherequirementforholdingprefetchedblocksinthecache:Ni=3;Np=112 Thus,wecanndapairofNiandNptomaximizeNtamongallpossiblepairsofNiandNpthatsatisfyEq. 3{1 .However,themaximumNtmayalsoberestrictedbyagivencachesize.SincedisksreadoneimmediateblockforeachofNistreamseverycycle,thenumberofimmediateblocksperstreampercycle(Nis)isone.TodistributetheNpblockstotheNistreamsasevenlyaspossible,weobtainthequotient(q)andtheremnant(r)bydividingNpbyNi.Theqrepresentsthenumberofprefetchedblocksthatareequallydistributedtoeachstreampercycle(Nps).Therrepresentsthenumberofremainingprefetchedblocksadditionallydistributedtosomeofthestreamsuntiltheremainingblocksareall

PAGE 39

distributed.ThesearesummarizedinEq. 3{2 . Figure 3{3 illustrateshowmanycacheblocksarerequiredattheendofi-thcycletoholdtheblocksprefetchedduringthepreviousNps+1cyclesinaconditionwhereNi=3andNp=11.Thediskreadsonlythreestreamseverycycle,butitreadsfourorveblocksperstreamthatarestoredcontiguouslyinthedisk.InFigure 3{3 ,eachrowindicatesthecyclewheneachblockisreadfromthediskandeachdiagonalrepresentsthecyclewheneachblockistransmittedtothenetwork.Itcanbeseenthatimmediateblocksareimmediatelytransmittedtothenetworkinthesamecyclewhenreadfromadisk,whiletheprefetchedblocksareservicedwithoutdiskaccessesinsubsequentcycles.Thus,prefetchedblockswillbeheldinthecacheforsubsequentNps+1cycles.Forexample,thediskreadsblock4,5,6,7,and8forstream13ini-thcycle.Theblock4istransmittedinthesamecycleandblock5,6,7,and8willbetransmittedinsubsequent(i+1),(i+2),(i+3)and(i+4)-thcycles,respectively.Thus,thediskshouldholdtheblocksformultiplecyclesuntiltheyaretransmittedtothenetwork.Ingeneral,thediskcacheshouldhold(NiNps+r)blocksforthe(i+1)-thcycle,(Ni(Nps1)+r)blocksforthe(i+2)-thcycle,andsoonuntilreachingrblocksforthe(i+Nps+1)-thcycle.GivenNiandNp,letdenoteMp(Ni;Np)thetotalnumberofcacheblocksrequiredtoholdNpacrossmultiplecycles.WeshowthecomputationofMp(Ni;Np)inEq. 3{3 . 2+r(Nps+1)(3{3) Toavoidasituationwheresomeblocksarereplacedbeforebeingtransmittedtothenetwork,theoutgoingtransmissionrateofblockstothenetworkfromthediskcacheshouldbeequaltoorhigherthantheincomingratefromthedisk(we

PAGE 40

discussthisissueinnextsubsection).Thus,thetotalnumberofblocksrequiredtobeheldinthecacheislargestatthebeginningofeachcycle.Theblockstobeheldinthecacheincludethefollowingtwokindsofblocks: Eq. 3{4 describessuchcachesizeconstraintwhenemployingonlyprefetchingwhereMtdenotethetotaldiskcachecapacityintermsofthenumberofblocks. Therefore,abovetwosystemdesignconstraintsofEqs. 3{1 and 3{4 canbeusedtodeterminethemaximumNp,therebyobtainingNiandNt.

PAGE 41

Threemessageexchangessuchasthearbitration,theopening,andclosingconnectionrequiretheentirelooptriptime,whichiscomputedasEq. 3{5 .Lar,Lopn,andLclsdenotethearbitrationtime,openingconnectiontimeandclosingconnectiontime,respectively.DisthetotalnumberofdisksattachedtotheFC-AL. TotransmitoneblockovertheFC-AL,theblockshouldbedividedintoframessincetheFC-ALtransmitsdatainunitsofframe.Thus,thenumberofframesbelongingtooneblock(B F)dependsontheblocksize(B)sincetheframesize(F)isxed.Inaddition,denotesaconstantofoverheaddegreecausedbytheowcontroldependingonthenumberofreceivebuers.(Dh)denotesthenumberofhopsbetweentwodisks.Thus,theFC-ALoverheadoccurringtotransmitoneblock(Lfc)is F((DhLnd)+Lpd):(3{6) LetNadenotethenumberofblockstransmittedovertheFC-ALperarbi-tration.TheintervalbetweentwoconsecutiveloopaccesstimesforeachdiskisdeterminedbycalculatingthetransmissiontimeoftheblocksthatallthediskshavesentoutovertheFC-ALwithonearbitration.Duringtheinterval,eachdisktransmitsNablocks.Thus,theoutgoingrateofNablockstotheFC-ALfromthediskcache(Cor)canbecomputedasEq. 3{7 . LtrD(3{7) Likewise,sincewecanevaluatediskaccesstimeofblockstoreadperstream,theincomingrateofoneimmediateblockandsubsequentprefetchedblocksperstream(Nis+Nps)tothediskcachefromthedisk(Cir)canbecomputedasEq.

PAGE 42

3{8 . Dtr+NpsB Dtr(3{8) ToensurethatallthecacheblocksareneverreplacedbeforebeingtransmittedtoFC-ALatleastonce,theoutgoingrateshouldbeequaltoorhigherthantheincomingrate.Otherwise,diskbandwidthshouldbewastedsincetheblocksshouldbereadagainsoon.Therefore,weshouldobtainmaximumNatosatisfyEq. 3{9 .

PAGE 43

Figure3{4. Supportingdualloops.(a)Exampleofalooppriorityassignmentindualloops,(b)Exampleofloopacquisitionorderandintervalinasingleloopconguration,(c)Exampleofloopacquisitionorderandintervalinadualloopconguration blockthatisselectedamongallcandidates.Also,thebitofthecacheblockissettotrue.OncetheblockistransmittedtotheFC-AL,thebitissettofalseandtheblockcanbeacandidateforreplacement.Thus,ifthereisatleastoneblockwithafalsebitinthediskcache,noblockisreplacedbeforebeingtransmitted.Ifthediskcacheisfullofblocksonlywithtruebits,oneofthecacheblocksshouldbereplacedbeforebeingtransmittedatleastonce,andthenreadagainlater.Thus,itisimportanttoadjusttheincomingandoutgoingrateofblockssothatthediskbandwidthmaynotbewasted.Ontheotherhand,ourreplacementpolicycanstillmaintainthecachingbenetwhentherequestsforthesamevideolearrivewithinashortperiod.Asthecachesizeincreases,theeectimprovessincetheblocksstaylongerinthecache.

PAGE 44

distinctdatapathsviadierentloops.Thisfeatureprovidesincreasedbandwidthaswellasfaulttolerance.Weshowedthatallthediskscanacquireaccessforasingleloopinthedecreasingorderoftheirpriorities.Asaresult,theintervalbe-tweentwoloopaccesstimesbecameuniform.However,multipleloopcongurationgeneratesthefollowingtwoissues.Oneisthatitishardtoestimatetheloopaccesstimesincearbitrationaremadeseparatelyinmultipleloops.Theotheristhattheintervalbetweenloopaccessesbecomeslongersincethenumberofdisksattachedtomultipleloopsisexpectedtoincreasetoutilizealltheloops.Accordingly,therequiredcachememoryincreasessincetheblocksshouldstaylongerinthecache.Toavoidthesesituations,weproperlyassignmultipleloopprioritiestoeachdisk.First,weassigntheprioritiesfortherstlooptoallthedisks.Basedonthisas-signment,wedeterminethepriorityofeachdiskforotherloopssothattheintervalbetweeneachloopaccessacrossmultipleloopscanbestillmaintaineduniformly.Wecancomputethepriorityofeachdiskformultipleloopsas LmodDforj>1(3{10) whereiPjrepresentsthepriorityofdiskiforj-thloop,andDandLdenotesthenumberofdisksandloops,respectively. Figure 3{4 (a)illustratesapriorityassignmentinadualloopconguration.InFigure 3{4 (b)and(c),weassumethattheuniformintervalbetweentwoconsecutiveloopaccessesisIwhenusingeightdisksattachedtoasingleloop.Also,thenumberofdisksdoublesinthedualloopconguration.NotethatD1canalwaysobtainloopaccessinthesameintervalinbothloopcongurations.Thus,wecanalsoachievethesameintervalinmultipleloopsaswellasestimatetheloopacquisitionorderandtimeeventhoughthenumberofdiskshasdoubled.

PAGE 45

Table3{2. Summaryofsimulationparametersforourprefetchingscheme Category Parameter TypicalValue Disk capacity 20GB seektime 1.2{15ms rotationallatency 4.17ms datatransferrate 20.9{40.8MB/s FC-AL propagationdelay 4ns/meter pernodedelay 6words(240ns) datatransferrate 200MB/s Videostream displayrate 4Mbps runningtime 100minutes Table 3{2 showsthesummaryofsimulationparameters.ThediskmodelisbasedontheIBMDeskStar60GXPmodel,whichhas20GBcapacityandanaveragedatatransferrateof30.8MB/s.Theaverageseektimeandrotationallatencyare8.5msand4.17ms,respectively.Thedisksaredividedinto18zones.Ourexperimentsemployzonebitrecordingaccordingtothediskspecication.Wevarythediskcachesizefrom4MBto256MBtoseetheimpactofthecachesizeontheperformanceofourprefetchingscheme.TheFC-ALhasadatatransferrateof200MB/s,whichatthistimeisavailable.Theper-nodedelayoftheinterfacetoforwardaframeis6words(i.e.,240nanoseconds)andthepropagationdelaybetweentwodevicesis4nanoseconds/meter.Onaverage,thevideostreamshaveabout4Mbps(i.e.,closetoDVDMPEG-2/4quality)andareassumedtohave

PAGE 46

constantbitrate(CBR)

PAGE 47

Figure3{5. Numberofconcurrentstreamsvs.numberofdisks servers.The4MBdiskcacheand256KBblockareemployedinboththeanalysisandtheexperiments.TheanalyticalresultsarecomputedfromEqs. 3{1 and 3{4 . Figure 3{5 demonstratesthattheexperimentalresultssupportouranalyticalmodel.Thenumberofconcurrentstreamsderivedfromtheanalysisisconsistentwiththeexperimentalresults.Forexample,whenfourdisksareemployed,thedierenceisonlytwostreamsbetweentheanalyticalandexperimentalresults.Aslightdierenceforprefetchingstartstooccuraftertheeightdiskcongurationandcomesfromthenumberofloopsemployed.TheFC-ALbandwidthisgraduallybecomingabottleneckaftereightdiskconguration.Thus,weaddedonemoreloopaftertheeightdiskconguration.EventhoughdisksareassignedtoproperFC-ALprioritiesfordualloopstoachieveuniformintervalbetweenloopaccessesacrossdualloops,wehaveobservedthatmoreoverheadtimeoccurstomanagetheloopaccesstimeindualloopscomparedtoasingleloop. AsshowninFigure 3{5 ,ourprefetchingschemeachievedabout30%improve-mentonaverageintermsofthetotalnumberofconcurrentstreams.Itisobviousthatprefetchingcanincreasetheratioofdatatransfertimeandreduceseektimeandrotationallatencybyreadingseveralblockscontiguouslystored.Wheneight

PAGE 48

Figure3{6. Disktimeratioanddiskbandwidthutilization disksareemployed,ourprefetchingschemesupports267concurrentstreams,whilethenon-prefetchingschemesupportsonly201streams(i.e.,thesignicantincreaseof66videostreams).InFigure 3{6 ,wecanseethedisktimeratioobtainedfromtheanalysisandexperimentswheneightdisksareemployed.Thedisktimeconsistsofdatatransfertime,seektimeandrotationallatency,andidletime.Theexper-imentalresultsrevealthattheratioofdatatransfertimeisapproximately11%greaterthanthatoftheseektimeandrotationallatencyinourprefetchingscheme.Conversely,theratioofseektimeandrotationallatencyisabout15%higherinthenon-prefetchingscheme.Thus,asshowninFigure 3{6 ,ourprefetchingschemeenablesa13%greaterdiskbandwidthutilization,whichisdenedas(totaltransfertime)/(totaltransfertime+totalseekandrotationallatency).Thisisbecauseonaverageprefetchingeliminatestheseektimeandrotationallatencyofabout40%ofthetotalblocks.Wehavealsoobtainedtheanalyticalresultsofdatatransfertimewith(Ni+Np)B Dtr,andseektimeandrotationallatencywithNi(Dst+Drl).Thenwecalculatedthediskbandwidthutilizationasdenedabove.Thecalculatedresultsalsoshowthesametrendastheexperimentsexceptthattheidletimeratioisabout3%lower.Itisduetothetemporarylargeuctuationofdiskaccesstime

PAGE 49

Figure3{7. Diskcacheutilization intheexperimentalresults,whichpreventsthedisksfromreadingmoreblockseventhoughbandwidthisstillavailable. 3{7 showsacomparisonofdiskcacheutilizationforprefetchingandnon-prefetchingschemes.Itcanbeseenthatthenon-prefetchingschemeutilizesaverysmallportionofthecache,amountingtoonly3.12%.Asmentionedabove,thereasonforthisisthatthecachingeectisnegligibleandthehigh-speedFC-ALisemployed.Ontheotherhand,ourprefetchingschemereadsmanyblocksaheadtotransmitinsubsequentcyclesandholdsthemtothelimitofdiskcachecapacity.Thus,diskcacheutilizationapproaches83%,whichindicatesthatthediskcachehassucientblocksreadytobetransferredtothestreamingserversintheverynearfuture.Itcanalsobeseenthatasthenumberofdisksincreasesdiskcacheutilizationgoesup.WhenonlyonediskisattachedtotheFC-AL,italwayswinsarbitrationsincethereisnoconictforloopaccess.Thismeanstheavailablecachespaceincreasesbykeepingtransmittingtheblocksinthecache.However,asdisksareadded,theintervalbetweenloopaccessesgetslonger.Thus,weneedtodecreaseNasothateachdiskcanacquireloopaccessmorequickly.Thecache

PAGE 50

Figure3{8. Blocksizevs.cachesizeintermsofthenumberofconcurrentstreams utilizationisassociatedwithhitratio.Sincethenon-prefetchingschemededicatesthediskcachememorytocachingforfutureusers,thediskcachehitratioshouldbeverylowforon-demandvideoapplications.Ourexperimentalresultsshowedthatthediskcachehitratiocanreachupto41%whenourprefetchingschemeisemployed,whiletheconventionalnon-prefetchingapproachcanrarelyachievethediskcachingbenet. 3{3 ,Npdoesnotincreaselinearlywithincreasedcachesize.IfNpblocksareformorethanonecycle,theblocksshouldoccupythediskspaceuntilthecyclewhentheyaretransmitted.Accordingly,thecachespacerequirementfortheprefetchedblocksincreasesasadegreeofk2asexplainedinEq. 3{3 .InFigure 3{8 ,atarounda64MBcacheformostofblocksizes,thenumberofconcurrentstreamsstartstobesaturated.Eventhougha256MBcacheisemployed,theincreasednumberofconcurrentstreamsisassmallas5to12%,comparedtoa64MBcache.Ontheotherhand,cachesizesignicantlyaectsthesystemperformancebeforereaching

PAGE 51

Figure3{9. OverheadofFC-ALarbitrations thesaturationpoint.Whena64KBblockisused,thesystemsupports224concur-rentstreamswitha4MBcachebut400streamswitha64MB,whichindicatesa78.5%performanceimprovement. Itisworthnotingthattheimprovementgraduallydecreasesastheblocksizeincreases.Fora2048KBblock,416concurrentstreamsaresupportedwitha4MBcachebut448streamswitha64MBcache.Thisisbecausetheratioofseektimeandrotationallatencyisinherentlylowwithalargerblock.Astheblocksizegetslarger,thetotalnumberofblockstoreaddecreaseswhilethevideolesizeisxed.Thismeansthatthenumberofdiskseekingoperationsisreducedtothesamedegree.Thus,thediskbandwidthutilizationbecomesrelativelyhighalthoughthesmallcachesizeisemployed.However,asalargerblockisemployed,bothstreamingserversandclientsshouldhavelargerbuers,andthestartuplatencyshouldalsoincrease. 3{9 showsthattheoverheadofarbitrationismuchlower

PAGE 52

withtheprefetchingschemethanwiththenon-prefetchingscheme.Bytrans-mittingNablockstogetherperarbitration,theFC-ALreducesthenumberofarbitrationsto1

PAGE 53

theservers,itisalsoeasytoupdateavarietyofnewmovietitlesfromthecontentproviderstomeetclientdemands. Theintervalcachingisanotherbueringschemethathasbeenmainlyem-ployedasadatabueringtechniqueinmultimediaserverstoreducestorageI/Obandwidth.Itisobviousthat,comparedtoprefetching,cachingcanalsosavediskdatatransfertimeinadditiontodiskseektimeandrotationallatency.Inthefollowingchapter,wethusextendtheproposedprefetchingschemebycombiningitwiththeintervalcachingtofurtherutilizenetworkeddiskcaches.

PAGE 54

4{1 liststheimportantsymbolsusedthroughoutthischapter. 3{1 and 3{4 inthepreviouschapter.Atthispoint,givenNiandNp,weevaluatehowmanycacheblocksarerequiredtoprefetchoneadditionalstream(Mp1(Ni;Np))usingEqs. 3{2 and 3{3 .Mp1(Ni;Np)canbecomputedsimplybysubtractingMp(Ni;Np)fromMp(Ni;Np+1).Dependingonthecurrentvalueofr,Eq. 3{3 isdividedintothefollowingtwocases:when0rNi2andr=Ni1.Intherstcase,asNpincreasesbyone,onlyrincreasestor+1,whileNpsremainsthesame.Inthesecondcase,NpsincreasestoNps+1whilerbecomeszero.Inbothcases,weobtainthesameresults: 1. 0rNi2 =NiNps(Nps+1) 2+(r+1)(Nps+1)NiNps(Nps+1) 2 42

PAGE 55

Table4{1. DenitionofimportantsymbolsforIPC Symbol Denition Unit block block block block block block block block block block block cachesizeusedforprefetching,givenNiandNp cachesizerequiredtoaddanewprefetchedblock,givenNiandNp tradingvalueofsaveddiskbandwidthforcachesize,givenNiandNp block bits/s bits/s bits/s

PAGE 56

=(r+1)(Nps+1)r(Nps+1) =Nps+1 2. =Ni(Nps+1)(Nps+2) 2NiNps(Nps+1) 2(Ni1) =(Nps+1) 2(Ni(Nps+2)NiNps2(Ni1)) =(Nps+1) 22 =Nps+1 (4{1) Thismeansthat,givenNiandNp,Nps+1cacheblocksareneededtoholdprefetchedblocksbelongingtooneadditionalstreamuntilthestreamends.Thus,ifthelengthoftheintervalbetweenastreamanditsprecedingstream(Ic)isgreaterthanNps+1,prefetchingthestreamwouldbebetterthancachingtheintervalifonlytherequiredcachesizeisconsidered. 1. Cachingisbetterintermsofbothrequiredcachesizeanddiskbandwidthsinceprefetchingrequiresmorecacheblocksandmorediskbandwidththancaching. 2.

PAGE 57

Table4{2. TheimpactofNiandNponMp(Ni;Np)intermsofthenumberofblocks Mp(Ni;Np) 1 14 14 0 105 2 13 6 1 49 3 12 4 0 30 4 11 2 3 21 5 10 2 0 15 6 9 1 3 12 7 8 1 1 9 8 7 0 7 7 9 6 0 6 6 10 5 0 5 5 11 4 0 4 4 12 3 0 3 3 13 2 0 2 2 14 1 0 1 1 Cachingisalsobetterintermsofthediskbandwidthduetothesaveddatatransfertimeeventhoughtherequiredcachesizesarethesameforbothtechniques. 3. Thiscasedependsonhowmuchthesaveddatatransfertimecontributestosystemperformance.Inotherwords,cachingmaybebetterinsomecasesduetothesaveddatatransfertimeeventhoughcachingrequiresmorecacheblocksthanprefetching. TodeterminehowlongtoextendthethresholdintervalsforaboveCase3,wethereforemeasurethecachingeectofadditionallysavingadatatransfertimeintermsofthecachesize.FromEq. 3{3 ,itcanbeseenthatMp(Ni;Np)increaseswithincreasedNpordecreasedNisinceNpsiscomputedbyjNp 3{2 .NotethatMp(Ni;Np)increasesmorerapidlythanNiincreasesorNpdecreases,i.e.,approximatelyasadegreeofNiN2ps.Thereasonforthisisthat,asNps

PAGE 58

increases,thediskcacheshouldholdprefetchedblocksformorecyclesuntiltheyaretransmittedtothenetwork. Table 4{2 illustrateshowMp(Ni;Np)variesaccordingtoNiandNpwhenNtisxedat15.ItshowsthatMp(Ni;Np)canbevariedsignicantlybyadjustingNiandNp.Forexample,Mp(Ni;Np)hasbeenreducedfrom105to9whenNiandNpchangedfrom1to7andfrom14to8,respectively.However,itisalsoclearthat,asMp(Ni;Np)decreases,therequireddiskbandwidthshouldincreaseduetotheincreasedNibecausereadinganimmediateblockrequiresmorediskbandwidthbyonediskseektimeandrotationallatencythanreadingaprefetchedblock.Therefore,thisindicatesthatthecachesizerequiredforprefetchingcanbereducedattheexpenseofthediskbandwidthbyincreasingNianddecreasingNp,andviceversa.Inotherwords,theavailablecachesizeanddiskbandwidthcanbetradedforeachother. Toseehowthediskbandwidthsavedbycachingonestreamcanbeconvertedintothecachesize,werstevaluatebothadditionallyrequireddiskbandwidthandreducedcachesizeforprefetchingwhenincrementingNianddecrementingNpbyone.Theadditionallyrequireddiskbandwidthcanbecomputedintermsofsecondsbysubtractingthetimetoreadoneprefetchblockfromthetimetoreadoneasfollows:Dst+Drl+B DtrB Dtr=Dst+Drl.ThereducedcachesizeforprefetchingcanbeobtainedbyMp(Ni;Np)Mp(Ni+1;Np1). Asmentionedabove,thesaveddiskbandwidthbycachingonestreaminsteadofprefetchingisonedatatransfertime,i.e.,B Dtr.Thus,givenNiandNp,thereducedcachesizebycachingonestream(Mtv(Ni;Np))canbeobtainedusingthefollowingproportionalexpression:Dst+Drl:Mp(Ni;Np)Mp(Ni+1;Np1)=B Dtr:Mtv(Ni;Np).Thus,thecachingeectofsavingtheadditionaldiskbandwidthintermsofthecachesizecanbemeasuredusingEq. 4{2 ,whichrepresentsthe

PAGE 59

tradingvalueofthediskbandwidthforthecachesize,givenNiandNp. Dtr IftheintervalisequaltoorshorterthanIth,wecacheeachstreamtominimizethecacheexpenseforsupportingit.Otherwise,weprefetchthestream.Inotherwords,Ithistheupperlimittodecidetocacheastreamintermsoftherequiredcachesize. ItisalsoworthnotingthatIthisdetermineddynamicallyaccordingtoNiandNp,asseeninEqs. 4{1 and 4{2 .Asincomingintervalsgetshorter,NcincreasesmorethanNp.ThismeansthatIthbecomesshorterbecauseMp1(Ni;Np)increasesproportionallytoNp.Similarly,Ithcanbelongerforlongintervals.

PAGE 60

01Computethelengthofthecurrentinterval(Ic)fromthisstreamtoits precedingstream; 02Computethecurrentthresholdinterval(Ith)usingEq. 4{3 ; 03if(IcIth) 04Cache(streamID;Ic); 05else 06Prefetch(streamID;Ic); Inthisalgorithm,Ntisdeterminedbythesumofthreenumbersofstreams:Ni,Np,andNc.Wheneachnewstreamarrives,thealgorithmcomputesIcandIthusingEq. 4{3 .IfIcisequaltoorsmallerthanIth,thealgorithmtriestocachethestream.Otherwise,ittriestoprefetchthestream.ThemainideaforecientlyutilizingbothcachespaceanddiskbandwidthistomanagetheamountofavailablecachesizeanddiskbandwidthbyadjustingNiandNpwhenevereitherthediskbandwidthorthecachespacebecomesexhausted.Whenavailablecacheblocksarenotsucienttosupportthecurrentstreambyeitherprefetchingorcaching,thealgorithmreducesthecachesizerequiredforprefetchingbyincrementingNianddecrementingNpbyone.Conversely,whenthediskbandwidthisnotadequatetoprefetchthecurrentstream,thealgorithmdecrementsNiandincrementsNpbyonetomakeavailablediskbandwidthforprefetchingit.Asexplainedabove,thisispossiblebecausethediskbandwidthrequiredforreadingaprefetchedblockissmallerbyonediskseektimeandrotationallatencythanthatforreadinganimmediateblock.Therefore,Niisusedonlyforbalancingthedegreeofavailability

PAGE 61

01Computeavailablecachesize(Ma); 02if(MaIc)f//thereexistsucientavailablecacheblocksforIc 06if(Il+MaIc+Mp1(Ni;Np))f//thereexistsucientavailablecache blocksforreplacing 07//trytoreplace 08if(DBaDBp)f//thereexistssucientdiskbandwidth forreplacing 09ReplacethelongestcachedintervalwithIc; 10Np++;//prefetchthereplacedstream 11gelsef//neitherManorDBaissucient 12Rejectthisstream; 13g//endofif 16if(DBaDBi)f//thereexistssucientavailabledisk bandwidthforincrementingNi 18Schedule(streamID); 19gelsef//neitherManorDBaissucient 20Rejectthisstream; 21g//endofif Let'slookintotheFunctionCacherst.WecomputeMausingEq. 4{4 bysubtractingthetotalcachesizerequiredforprefetchingandcachingfromMt,addingatermofthecachesizerequiredforcaching(Mc)toEq. 3{4 . IftheavailablecachesizeissucientforIc,wejustcacheit.Otherwise,wetrytoreplacethelongestcachedintervalwithIcorincreasetheavailablecachesize

PAGE 62

tocacheitbyincrementingNi.Weestimatetheavailablediskbandwidth(DBa)basedonhowmuchdiskbandwidthwasavailableinthepreviouscycle.Ifavailablecachesizeanddiskbandwidthareeachsucientbothtoreplacethelongestcachedinterval(Il)withIcandtoprefetchthereplacedstream,weperformthereplacementandthenprefetchthereplacedstream.Thedecisiononwhetherornottoreplaceisbasedonthefollowingcacherequirement:Il+MaIc+Mp1(Ni;Np)(i.e.,ifthenumberofbothincreasedavailablecacheblocksbyreplacingthelongestoneandcurrentlyavailablecacheblocksisgreaterthanthenumberofcacheblocksrequiredbothtocacheIcandtoprefetchthereplacedstream).Ifthecacherequirementisnotmet,weincreaseMabyincrementingNianddecrementingNpbyoneaslongasDBaissucienttoreadoneimmediateblock(DBi).ThenwecallupthisfunctionagaintoapplymodiedNiandNp.InanysituationwhereneitherManorDBaissucient,werejectthestream. TheFunctionPrefetchcanbedividedintofourcasesaccordingtoMaandDBa.First,ifbothMaandDBaaresucienttoprefetchastream,wejustprefetchthestream.Second,ifneitherManorDBaissucientforprefetchingastream,werejectit.Third,ifMaissucienttoprefetchastreambutDBaisnot,wetrytoincreasetheavailablediskbandwidth.IfMaislargeenoughtodecrementNiandincreaseNpbyone(i.e.,MaMp(Ni1;Np+1)Mp(Ni;Np)),wecallupthisfunctionagainafteradjustingthem.Lastly,ifDBaissucienttoprefetchonestreambutMaisnot,wetrytoreplaceIlwithIcandprefetchthereplacedstreamaslongastheabovecacherequirementismet.Otherwise,wealsotrytoincreaseMabyadjustingNiandNpaslongasDBaisgreaterthanDBi.

PAGE 63

01Computecomputeavailablecachesize(Ma); 02Estimatetheavailablediskbandwidth(DBa); 03if(bothMaandDBaaresucientforoneprefetch)f 05g 08g 10//trytoincreaseavailablediskbandwidth 11if(MaMp(Ni1,Np+1)-Mp(Ni,Np))f//thereexistsucient availablecacheblocks fordecrementingNi 13Schedule(streamID); 14gelsef 16g//endofif 18//trytoreplace 19if(Il+MaIc+Mp1(Ni;Np))f//thereexistsucientavailablecache blocksforreplacing 20ReplacethelongestcachedintervalwithIc; 21Np++;//prefetchthereplacedinterval 22gelsef 24if(DBaDBi)f//thereexistssucientavailabledisk bandwidthforincrementingNi 26Schedule(streamID); 27gelsef 29g//endofif

PAGE 64

01if(Thisstreamisbeingservicedbyadiskoperation)f cached)f//\case2" 04DeletecurrentStreamIDfromDiskServicingStreamlist; 05DeleteIffromCachedIntervallist; 06AddfollowingStreamIDtoDiskServicingStreamlist; 07Nc; 08gelsef//\case1" 10if(Np>0) 11Np; 12else 14g//endofif iscached 16if(theintervalbetweenthisstreamanditsfollowingstream(If)is cached)f//\case4" 18DeleteIpfromCachedIntervallist; 19DeleteIffromCachedIntervallist; 20AddfollowingStreamIDtoDiskServicingStreamlist; 21Nc=2; 22Np++; 23gelsef//\case3" 25Nc; 26g//endofif

PAGE 65

Figure4{1. Fourpossiblecasesoccurringwhendeletingexistingstreams therelationshipbetweenthedeletedstreamanditsfollowing/precedingstreams.Wehaveobservedthathowtodeleteitdependsonthefollowingtwofactors:whetherthestreamisbeingprefetchedandwhethertheintervalsfromitsfollow-ing/precedingstreamsarebeingcached. Toscheduleanddeletethestreamsbeingserviced,thefollowingtwolistsofstreamsaremaintained:DiskServicingStreamlistandCachedIntervallist.TheDiskServicingStreamlistconsistsofthestreamsthatarebeingservicedbyreadingprefetchedandimmediateblocksdirectlyfromthedisk.(Fortheconvenienceofexplanation,weassumethattheyareallbeingprefetched.)TheCachedIntervallistmaintainstheintervalsthatarebeingcachedinthedescendingorder. AlgorithmDeleteandFigure 4{1 illustratefourdierentcasesthatshouldbeconsideredwhendeletingexistingstreams.Incases1and2,thecurrentstreamstobeterminatedarebeingprefetched.Since,incase1,theintervalbetweenthecurrentstreamanditsfollowingstream(If)isnotbeingcached,deletingthestreamdoesnotaecttheplaybackofotherstreams.Thus,wejustdeletethecurrentstreamfromtheDiskServicingStreamlist.Incase2,Ifisbeingcached,

PAGE 66

sowedeleteIffromtheCachedIntervallistaswellasdeletethecurrentstreamfromtheDiskServicingStreamlist.SincethefollowingstreamcannotanylongerbeservicedduetothedeletionofIf,itshouldbeprefetchedlateronbyaddingittotheDiskServicingStreamlist. Incases3and4,theintervalbetweeneachcurrentstreamanditsprecedingstream(Ip)isbeingcached.Inbothcases,itsprecedingstreamcanbesupportedbyeitherbeingcachedorprefetched.Incase3,(If)isnotbeingcached,soonlyIpisdeletedfromtheCachedIntervallist.Incase4,sinceIfisalsobeingcached,wedeletebothIfandIpfromtheCachedIntervallist.Itsfollowingstreamshouldalsobeprefetchedforthesamereasonasincase2.

PAGE 67

Figure4{2. Networktransmissionorderandtimeofeightdisksaccordingtotheirpriorities. 4{2 showsthatdiskscanacquirethenetworkaccessesinadeterminedordereveryxedtransmissioninterval(TI)wheneightdisksareemployed.Wecanseethatthedisksacquirethenetworkaccessesinaround-robinmanneraslongastheyareallparticipatinginarbitration. Inaddition,notethatNi+Np+Ncblocksshouldbetransmittedtothenetworkduringeachcycle,whileonlyNi+Npblocksarebeingreadfromthedisk.ThisisbecausetheNcblocksalreadycachedinpreviouscyclescanbeservicedwithoutactualdiskaccesses.Thismeansthattheoutgoingtransmissionrateofblocksfromthediskcachetothenetworkshouldbeonaveragehigherthantheincomingrateofblocksfromthedisktothediskcache.

PAGE 68

Therefore,eachdisktransmitsthesamenumberofblockstothenetworkeveryxedintervalatafasterspeedthanwhenthediskreadsthemfromthedisk.Accordingly,toschedulenetworktransmissionofallblocksduringtheentirecycle,weneedonlythecachesizetoholdimmediateandprefetchedblocksbelongingtothestreamchosenrstlyatthebeginningofeachcycle,asseeninEq. 3{3 . Topreventtheblocksinthecachefrombeingreplacedbeforebeingtransmit-tedtothenetwork,itisalsoimportanttofreetheblocksasquicklyaspossiblefromthebeginningofeachcycle.Thus,wetransmittheblocksthatwillbere-placementcandidatesinthecurrentcycleearlierthanthosetobeheldinthecacheuntilsubsequentcycles.Ineachcycle,thenetworktransmissionschedulecanbemadeinthefollowingorder:(1)blocksprefetchedinthepreviouscyclesandtobetransmittedinthecurrentcycle,andblockscachedinthepreviouscyclesandtobefreedinthecurrentcycle,(2)blockstobereadfromthediskinthecurrentcycleandtobefreedinthesamecycle,and(3)blockstobereadinthecurrentcycleandtobeheldinthecacheforsubsequentcycles.Itisassumedinthischapterthatthenetworkbandwidthissucienttohandlethetransmissionofalltheblocks. 4{3 .Theyareusedthroughoutoursimulationsunlessotherwiseindicated. ThearrivaloftheplaybackrequestsfollowsaPoissondistributionwiththemeaninter-arrivaltimessettobetween0.5and4.5seconds.TheaccessfrequencyofeachstreamfollowsaZipfdistributionwhoseparametershavearangebetween0.0(highestskewedaccessdistribution)and1.0(uniformaccessdistribution).

PAGE 69

Table4{3. SummaryofvaryingsimulationparametersforIPC Parameter Range DefaultValue Numberofdisks 10{100 50 Numberofstreams 100{1000 500 Cachesizeperdisk 16{1024MB 128MB Diskblocksize 64{512KB 256KB Inter-arrivaltime 0.5{4.5seconds 2.5seconds Zipfdistributionparameter 0.0{1.0 0.271 Table4{4. SummaryofdiskandvideotitleparametersforIPC Category Parameter AverageValue Disk Capacity 40GB Seektime 8.5ms Rotationallatency 4.17ms Datatransferrate 14MB/s Videotitle Playbackrate 4Mbps Duration 2hours Weuseadefaultparameterof0.271,whichistypicallyusedforvideorentaldistribution[43].Thecachesizeandblocksizeperdiskalsovaryfrom16to1024MB,andfrom64to512KB,respectively. Table 4{4 showstheaverageparametervaluesofdisksandvideotitlesusedinthissection.Eachdiskhasa40GBcapacityandanaveragedatatransferrateof14MB/s.Theaverageseektimeandrotationallatencyare8.5msand4.17ms,respectively.Theplaybackrateofvideotitlesis4Mbpsandeachvideotitleis2hourslong.Thus,itisassumedthateachdiskstores10videotitlesandthetotalnumberofvideotitlesstoredonallthedisksincreasesproportionallytothenumberofdisks(i.e.,10numberofdisks).Thesimulationsarecarriedoutusingupto100disksand50disksareusedasadefaultvalue.Wealsoassumethateachstreamhasaconstantbitrate(CBR)[54].

PAGE 70

Figure4{3. Numberofconcurrentstreamsvs.numberofdisks. Thevideotitlesareassignedtodiskssothattheloadsofplaybackrequestsamongthediskscanbedistributedasevenlyaspossibleaccordingtotheirex-pectedaccessfrequencies.Eachtitleissegmentedanddistributedonvedisks,anditisretrievedfromthedisksinparallelanddeliveredtoaclientduringacycle. Todemonstratetheeectoftheintegration,fourcasesarecompared:IPC,PreFetchingOnly(PF)[6,9],IntervalCachingOnly(IC)[11,12],andNoBueringTechniqueUsed(NU).Asstatedabove,intervalcachingisemployedasadiskcachingtechnique.Throughoutthissection,IPCtakesadvantageofprefetchingandcachingsimultaneouslybyexecutingthemoreadvantageousoneofthetwotechniquesfortheperformance.Therefore,IPCoutperformsothertechniquesinallsimulationresults,whichareallbasedonasimulationdurationof100hours. Notethatincreasingthenumberofcachedstreams,ratherthanprefetchedstreams,playsamoresignicantroleonperformanceimprovement,sinceacachedstreamonaveragerequiresfewercacheblocksthanaprefetchedstream.Thus,theoverallperformanceimprovementofIPCisaectedmainlybythedegreeofthecachingeectinoursimulationresults.

PAGE 71

Figure4{4. Figure4{5. IntervalvaluesofIPCandCOvs.numberofdisks. Figure4{6. CacheportionsofIPCvs.numberofdisks.

PAGE 72

4{3 illustratestheeectofthenumberofdisksonsystemperformanceintermsofthenumberofconcurrentstreamswhenfourcasesareemployed.Toinvestigatetheeectivenessofourproposedscheme,werstcomparefourcasesbasedontheaveragenumberofconcurrentstreams.ItcanbeseenthatIPConaveragesupported14.1,35.2,and80.4%moreconcurrentstreamscomparedtoPF,IC,andNU,respectively.Thisispossiblebecause,asimpliedinFigs. 4{4 and 4{5 ,IPCsignicantlyreducedtherequiredcachesizeperstream.Figure 4{4 showsthattheaverageNpsofIPCis36.7%smallerthanthatofPF.ThismeansthatIPCrequiresfewercacheblockstoprefetchoneblockthanPF,asindicatedinEq. 3{3 .Figure 4{5 showsthat,comparedtoIC,IPCreducedtheaverageandmaximumintervallengthsby52.3and56.9%,respectively.Asaresult,IPConaveragerequires14.1and25.3%fewercacheblockstoholdonestreamthanPF,IC,andNU,respectively. ItisalsoevidentfromFigure 4{3 thatthenumbersofdisksandvideotitlesaecttheperformancetosomedegree.Asstatedabove,duetolimitationsofstoragecapacity,eachdiskstores10videotitles.Thus,100videotitlesshouldbeservicedadditionallyaseach10disksareadded.Since,whenschedulingrequests,PFandNUarenotaectedbythelengthsoftheintervalsbetweensuccessivestreamsforthesamevideotitle,theirperformancehaslittletodowiththevaryingaccessfrequencyofeachvideotitleaccordingtothenumberofdisks.Thus,theyincreasethenumberofconcurrentstreamsalmostlinearlywiththenumberofdisks. Ontheotherhand,comparedtoPFandNU,IPCandICincreasethenumberofconcurrentstreamsslightlyslowlywithanincreasednumberofdisks.Forexample,IPCsupports333concurrentstreamswhen10disksareemployed,while2880streamsaresupportedwith100disks.Thereasonisthat,aseachdiskis

PAGE 73

Figure4{7. Numberofconcurrentstreamsvs.cachesizeperdisk added,theaccessfrequencyofeachvideotitleisdecreasedbecause,giventhetotalaccessfrequency,theaccessfrequenciesoftheaddedvideotitleswillbeassignedbydecreasingthoseoftheexistingvideotitles.Thus,thedecreaseinaccessfrequencyofeachvideotitlelengthenstheintervals,therebyslightlyreducingthenumberofconcurrentstreamswithanincreasednumberofvideotitles.Figure 4{5 showsthat,forbothIPCandIC,thecachedintervalsgetlongerasthenumberofdisksincreases.Figure 4{4 showsthatIPCincreasesNpswithanincreasednumberofdiskssinceIPCtendstoprefetchstreamsinsteadofcachingduetothelengthenedintervals.Asaresult,theportionofcacheblocksusedforcachingincreasesandthatforprefetchingdecreaseswithanincreasednumberofdisks,asshowninFigure 4{6 . 4{7 showstheeectofthecachesizeperdiskforeachcasewhenvaryingthecachesizefrom16to1024MB.Onaverage,IPCimprovedperformanceby25.1,32.7,and91.9%comparedtoPF,IC,andNU,respectively.NotethattheimprovementpercentageofIPCincreaseswithanincreasedcachesize.Thenumberofconcurrentstreamsincreasedby2.3,18.2,and31.4%witha16MBcachesize,whileimprovementpercentagewas65.5,40.2,and179.3%witha1024MBcache

PAGE 74

Figure4{8. IntervalvaluesofIPCandCOvs.cachesizeperdisk Figure4{9. Figure4{10. CacheportionsofIPCvs.cachesizeperdisk

PAGE 75

size.ThisisbecauseIPCcanecientlyutilizetheincreasedcachesizebytakingthebenetofPFandICsimultaneouslybasedondynamicIthvalues. TheICconstantlyincreasesthenumberofconcurrentstreamseventhoughtheimprovementpercentageslightlydecreaseswithincreasedcachesize.ThisisbecauseICcanholdevenlongerintervalswithalargercache,asshowninFigure 4{8 .WithPF,however,thenumberofconcurrentstreamsbecomessaturatedatbetween128and256MBcachesize.Thereasonforthisisthat,asshowninFigure 4{9 ,Npsincreasesasthecachesizeincreasesbecausemoreprefetchedblockscanbeheldinalargercache.Asaresult,Mpincreasesmorerapidly,approximatelyNiN2ps,tokeepprefetchedblocksinthecacheformorecycles,asexplainedbefore.Therefore,wehaveobservedthat,eventhoughthetotalnumberofcacheblocksisdoubled,theaveragenumberofcacheblocksrequiredtoholdoneprefetchedblockalsoincreasestoasimilardegree.Consequently,PFcanincreaseonlyafewmoreconcurrentstreamsafterthatpoint.Forexample,theimprovementintermsofthenumberofconcurrentstreamsisassmallas2.9%between256and1024MBcachesize.Ontheotherhand,beforereachingthesaturationpoint,theimpactofcachesizeonperformanceismuchgreater.ThePFsupports1042concurrentstreamswith16MBcachesizebut1285streamswith128MBcachesize,whichindicatesa23.3%performanceimprovement. Figure 4{10 demonstratesthatIPCutilizesthecachespaceeciently.Itprefetchesmorestreamsinsteadofcachinguntilreachingthesaturationpointandthenitcachesmorestreamsastheexpenseforprefetchinggetshighersincethecachingeectissimilarwithallcachesizes,asshowninFigure 4{7 . 4{11 .TheIPCobtainedabetterperformancethantheotherswithallblocksizesthroughtheecientintegrationofprefetchingand

PAGE 76

Figure4{11. Numberofconcurrentstreamsvs.diskblocksize Figure4{12. ImprovementpercentageofIPCvs.diskblocksize

PAGE 77

caching.Onaverage,IPCachieved14.6,37.7,and86.6%improvementcomparedtoPF,IC,andNU,respectively. TheNUcontinuestoincreasethenumberofconcurrentstreamsasblocksizeincreasesbecausemorediskseektimeandrotationallatencycanbereducedwithalargerblock.Forthisreason,multimediaserversusuallyemploylargerblockstoimproveperformance.TheICgeneratesasimilarperformancecurvetothatofNUwithallblocksizes.Thisimpliesthat,asshowninFigure 4{12 ,theimprovementpercentagefromthecachingeectisalmostconstantthroughoutallblocksizes.(Theimprovementpercentagerepresentswhatpercentofstreamscanbeadditionallysupportedwitheacheectofcaching,prefetching,andtheintegrationofbothtechniques,comparedtoNU.)Thisisbecausetheintervalsbetweensuccessivestreamsforthesamevideotitleintermsofabsolutetimedonotchangeeventhoughblocksizevaries.Thus,theincreaseinthenumberofconcurrentstreamswithICismainlyduetotheeectofemployinglargerblocks.Forinstance,ICoutperformsPFwith512KBblocksize,whilePFsupports132%moreconcurrentstreamsthanICwith64KBblocksizeeventhoughthecachingeectsremainsimilarwithbothblocksizes. Itisinterestingthat,withIPCandPF,theperformanceimprovementfromemployinglargerblocksisverysmall.Thishappensbecause,eventhoughsmallblocksareemployed,IPCandPFcanhavethesameeectasemployinglargeblocksbyprefetchingseveralsmallblockscontiguouslystoredonthediskatatime.Thisismeaningfulwhendesigningmultimediaserversinthatwecanrelaxtheblocksizeconstraintforbetterperformance.ForafaircomparisonwithIC,however,weemployalargediskblocksizeof256KBasadefaultvalueeventhoughthedependencyofIPCandPFonthediskblocksizeisverysmall.

PAGE 78

Figure4{13. Numberofconcurrentstreamsvs.accessskew Figure4{14. IntervalvaluesofIPCandCOvs.accessskew Figure4{15. CacheportionsofIPCvs.accessskew

PAGE 79

4{14 .Thus,itcanbeseeninFigure 4{15 that,withthedecreasedskewdegree,theportionofthecachesizeusedforcachingdecreases,whilethatforprefetchingincreasestothesamedegree.Forexample,bothportionsarealmostthesamewiththehighestskewdegree(0.0),buttheportionforprefetchingbecomes2.3timesgreaterthanthatforcachingwiththeuniformaccessdistribution(1.0).Notethatonaveragemanymorecacheblocksarerequiredtoprefetchastreamthantocacheit.Consequently,IPCandICimprovedtheperformancebetweenhighestandlowestskewcasesby30.1and48.3%,respectively,asshowninFigure 4{13 . Whencomparingthefourcases,IPCalsoachieves12.3,36.4,and77.4%improvementintermsofthenumberofconcurrentstreams,comparedtoPF,IC,andNU,respectively.ItcanalsobeseenthatperformanceofPFandNUisnotaectedbytheaccessskewdegreesincetheyperformindependentlyofintervallengthsbetweensuccessivestreamsforthesamevideotitle,asmentionedabove. 4{16 illustratesthenumberofconcurrentstreamsvs.inter-arrivaltime(1=)witharangefrom0.5to4.5seconds.Asexpected,IPCimprovedtheperformanceby17.3,33.1,and85.3%,comparedtoPF,IC,andNU,respectively.WithIPCandIC,thenumberofconcurrentstreamsdecreasesastheinter-arrivaltimeincreases.Tominimizetherequiredcachesize,IPCthereforedecreasesthe

PAGE 80

Figure4{16. Numberofconcurrentstreamsvs.inter-arrivaltime Figure4{17. IntervalvaluesofIPCandCOvs.inter-arrivaltime Figure4{18. CacheportionsofIPCvs.inter-arrivaltime

PAGE 81

numberofcachedblockswithalengthenedinter-arrivaltimewhileincreasingthenumberofprefetchedblocks,asshowninFigure 4{18 .ThePFandNUdonotdependontheinter-arrivaltimeforthesamereasonastheeectofaccessskew. Inaddition,theperformancedierencesofIPCandICbetween0.5and4.5secondcaseswere22.7and34.5%,respectively.Similartotheeectofaccessskew,thisisalsoduetoanincreaseinintervallengthsbetweenstreams,asshowninFigure 4{17 . Ourextensivesimulationresultshavedemonstratedtheeectivenessofourproposedintegrationofprefetchingandcachingwhilevaryingseveralsimulationparameters.TheIPCachievedaconsiderableperformanceimprovementbydecreasingthecachesizerequiredtoprefetchonestreamandthelengthsofcachedintervals.TheIPCincreasedthenumberofconcurrentstreamsproportionalto

PAGE 82

dedicatedresourcessuchasthenumberofdisksandcachesize.WeshowedthattheIPCwasabletorelaxconstraintontheblocksizeindesigningmultimediastoragesystemsbyobtainingthesimilarperformanceirrespectiveofvaryingblocksize.ItwasalsoshowedthattheIPCdeterminedwhethertoprefetchortocachestreamsadaptivelytotheaccessskewofpopularitydistributionandinter-arrivaltimes. Thecapabilityofnetworkedstoragesystems,especiallytheFC-AL,toprovidehigh-speedcommunicationbetweenstoragedevicesimpliesthatconsumerdevicesequippedwithFC-ALinterfacecanalsosharehigh-qualitycontentsinreal-timebetweeneachotherwithouttheinterventionofservers.Inthefollowingchapter,weexploittheFC-ALasahigh-speedstreamingnetworktoconstructanovelarchitectureforahigh-qualitydatasharingservice.

PAGE 83

71

PAGE 84

beforebecomingunavailableagain,andonly20%wereavailableforatleasttwohours.Thisusers'characteristicmakesreliablecontentstreamingdicult. ToovercometheabovelimitationsofP2Pnetworkingtechnique,weemployFC-ALtechnologytoconnectallsystemcomponents[5].Asmentionedinchapter3,theFC-ALcanprovidehighperformancecomparedtotraditionalinterfacesduetosuchseveraladvantagesasfull-duplextransmission,highbandwidth,andafairnessarbitrationalgorithm. 5{1 (a),eachinternalFC-ALnetworkisconnectedtotheglobalP2Pnetworksasapeeranditcanbeconstructedindenselypopulatedareas.Itcanbeseenthatindividualpeerscanalsobeconnectedtotheglobalnetworks.Figure 5{1 (b)illustratesaninternalFC-ALnetworkthatisconstructedinahigh-risingapartment.TheinternalnetworkconsistsofaContentManagingServer(CMS),apoolofnetworkdisks,andPVRs. TheCMScoordinatescontentstreamingandsharingamongPVRs.TheCMSmaintainsalistoftheTVprogramsthatinternalPVRsstore,anditcontinuouslycollectsinformationonthestateofthePVRs.Basedonthisinformation,itexecutesecientschedulingschemesthatisdescribedlater.Inaddition,itbehavesasapeerintheglobalP2PnetworksonbehalfofinternalPVRsusingconventionalP2Pnetworkprotocolsforsearchingandtransferring. ApoolofnetworkdisksstoreTVprogramsforthreepurposes.Onepurposeisforextensivetime-shiftingforallchannels.Inourarchitecture,time-shiftedprogramsarestoredonlyinthepoolofnetworkdiskssothatPVRscansharetime-shiftedprogramswitheachotherinanecientwayaswellastakeadvantageofthediskspacethatshouldbereservedfortime-shiftinginstand-alonePVRs.AnotherpurposeistosharepopularTVprogramsintheinternalnetwork.The

PAGE 85

Figure5{1. OverallTVcontentdistributionarchitecture.

PAGE 86

concentratedloadsonsomePVRscanbedistributedontoseveralnetworkdisksbyswitchingplaybackrequestsandreplicatingpopularprograms.ThethirdpurposeistocreatetemporalstoragebetweentheglobalP2PnetworksandinternalPVRs.ThePVRscanrequesttheCMStosearch,transfer,andstoreprogramsthatarenotavailableontheinternalnetwork. PVRsrecordanddisplaywhatuserswant.Inaddition,whenPVRsarerequestedtotransfersomeliveTVprogramfortime-shifting,theyreceivetheprogramthroughanavailabletunerandtransmitittothepoolofnetworkdisks.Intheinternalnetwork,PVRscancommunicateprogramswitheachotherineitherreal-timestreamingornon-real-timeletransferdependingonauser'spreference. Consequently,ourarchitectureisnotonlyabletostablyprovidestreamingofHD-qualityTVcontentsoverfastnetworkconnections,butitcanalsocontributereliabilitytotheexistingP2PnetworkssincetheCMSparticipatesasapersistentpeer.Furthermore,byextendingstoragecapacitywiththepoolofnetworkdisks,itispossibletoprovidemuchlongertime-shiftingforallthechannelsandtodistributetheconcentratedloads.Itisnotedthatoursystemcomponentscanprovidevariouscontentdistributionchannels,suchaslivebroadcasting,localstoragedevice,thepoolofnetworkdisks,internalPVRs,andglobalP2Pnetworks. Ontheotherhand,inthebroadcastingenvironment,theconditionalaccess(CA)systemsareemployedtoprotectthebroadcastTVcontentsfrompeoplethatdonothavetheaccessrightstothecontents.TheCAsystemsencryptthecontentsatthebroadcastingsidesanddecryptitatreceiversideswithauniqueuserkeyassignedtoeachsubscribertomakesurethatonlyauthorizeduserscandecryptthecontents.TheCAsystemscandeterminewhichchannelseachsubscribercanwatchthroughtheSubscriberManagementSystem(SMS)whichisadatabaseofallsubscribers.

PAGE 87

However,sinceinourarchitecturebroadcastTVcontentscanbesharedbetweenPVRsafterstoredonthem,itisnecessarytoextendtheCAsystemstoprotectouradditionalservicessuchasTVprogramsharingandextensivetime-shifting.WeassumethatPay-TVserviceoperatorsprovidetheTVcontentsharingservicebychargingfortheadditionalfeaturesaswellasinstallingtheCMS,thepoolofnetworkdisksandFC-AL.SincetheoperatorscanaccesstheSMSoftheCAsystemsincludingtheuserkeysofallthePVRs,theoriginalCAinformationcanalsobeappliedtothecontentsharingserviceundertheoperator'scontrol.SupposethataPVRrequestsaprogramstoredonotherPVRstotheCMS.AfterauthenticatingtherequestingPVRwithitsuserkeyandverifyingthattherequestingPVRhastheaccessrightstotheprogram,theCMSdeterminesatargetPVRtotransmittheprogramtotherequestingPVR.TheCMSthenissuesaticketrepresentingthattherequestingPVRhastheaccessrightstotheprogram,alongwithasessionkeyandtheaddressoftargetPVR.However,iftherequestingPVRdoesnothavetheaccessrightstotheprogram,theCMSjustrejectstherequest.AfterreceivingtheticketfromtherequestingPVR,thetargetPVRnoticesthattherequestingPVRhastheaccessrightsandthenstartstransmittingtheprogram.Allthedataarecommunicated,encryptedwithkeyssharedonlybetweentwocorrespondingcomponentssuchasuserandsessionkeys. SinceitmustreceivetheTVsignalthroughthetuneranddecodethesignalfordisplay,aPVRconsistsofatleastonetunerandonedecoder.Inthisdisser-tation,thisiscalledabasicPVR.Figure 5{2 showsastatediagramforabasicPVR.Thereareveprimitivestates:aninitialstate(I),liveTV(L),time-shifting

PAGE 88

Figure5{2. Astatediagramforabasicstand-alongPVR. (T),recordedprogramplayback(P),andnlogicalchannelrecording(Rn)forn=1;2;3;4.Infact,sinceonephysicalchannelfordigitalTVcanincludeuptofourlogicalchannelscarryingfourStandardDenition(SD)-qualityprograms,itispossibletodisplayoneSDprogramwhilesimultaneouslyrecordinguptothreemoreSDprogramswithinthesamephysicalchannel.Thus,thismodelsupportsthreecombinationsoftwoprimitivestatesthatincludeRn,suchasRnL,RnT,andRnPforn=1;2;3.WhenthePVRisturnedon,itshouldbeatstateI.Subse-quently,itcanmovetoanotherstatedependingonwhatfunctionauserselects.ThePVRthencontinuestochangeitsstate,triggeredbythefollowingeighteventsdependingonuserinteractions:channelchange,recordedprogramplaybackstart,recordedprogramplaybackdone,recordingstart,onerecordingdone(notthelastone),lastrecordingdone,VCR-likefunctionssuchasfastforwardandrewind,andreturntoliveTV.ThedetailedstatesandtransitionoperationsareillustratedinFigure 5{2 . OncePVRsarenetworkedwithahigh-speedFC-AL,theycanshareTVprogramsinreal-time.Tosupportnetworkedfunctions,wethusextendthebasicPVRmodelbyaddingseveralprimitivestates.First,astatePcanbedividedintothreedierentstatestosupportthestreamingfromanotherPVR(statePp)

PAGE 89

andthepoolofnetworkdisks(statePd),aswellastheexistinglocalrecordedprogramplayback(statePl).Second,sincetime-shiftedprogramsarestoredonlyonthepoolofnetworkdisks,weaddastateX,whichindicatesthatthePVRistransmittingatleastoneliveprogramtothepoolofnetworkdisksinreal-time.IncaseusersjustwanttostoreTVprogramsontheirPVRsforlaterviewing,theprogramscanbecopiedintheinternalnetworkinnon-real-time.Thus,weneedanadditionalfourstates:P2PcopyingofprogramstoanotherPVR(oCp),P2PcopyingofprogramsfromanotherPVR(iCp),copyingofprogramstothepoolofnetworkdisks(oCd),whichisreplication,andcopyingofprogramsfromthepoolofnetworkdisks(iCd). ItisclearthatPVRsequippedwithmultipletunersanddecodersbecomemuchmoresophisticatedifwecombineavarietyofprimitivestatessuchasL,T,R,Pl,Pp,Pd,X,oCp,iCp,oCd,andiCd.SupposethataPVRisnetworkedwithntunersandmdecoders.Sincesystemresourcesarelimited,thePVRshouldhavetheabilitytodecidewhetherornotitwillacceptanewrequest.ThePVRacceptsonlyrequeststhatsatisfythefollowingdesignconstraintsatthenextstate:

PAGE 90

Figure5{3. Representationofaspecicprogramposition. Sincetheaboveconstraintscanbecheckedatanytimewithoutdeterminingallpossiblenextstatesinadvance,thiscanbeanecientadmissioncontrolforincomingrequests. Beforedescribingthoseschemes,wesuggesthowtoidentifyaspecicpositionofabroadcastprogramamongPVRs.Inabroadcastingenvironment,allthePVRsshouldreceivethesamecontentforaspecictimeperiodaslongastheyaretunedtothesamechannel.However,PVRsmaystartstoringthesameprogramat

PAGE 91

dierenttimesbecausethecontentisbroadcastwithoutaprogramstartindicationanditisveryhardforallPVRstosynchronizeaglobaltime[79].Asaresult,itisalsodiculttoidentifythespecicpositionofthesameprogramamongPVRssincethepositionistypicallyrepresentedbyabyteosetfromthebeginningoftheprogramle.Moreover,ifaprogrambeginstobestoredafterithasalreadystartedbroadcasting,itbecomesmorediculttondtheosetfromthebeginningoftheprogram. Sincebasicstand-alonePVRsplayonlyprogramsstoredontheirowndisks,eachPVRworksindependentlyofwhichpositionsothersarecurrentlystoringorplaying.However,inourarchitecture,aPVRmaytakeovertransmissionresponsi-bilityfromanotherPVRtomatchacommunicationpair,switchplaybackrequestsbetweenPVRs,andkeepstoringthesameprogramfortime-shiftingbetweenPVRs(theseschemesaredescribedlater).ThismeansthatitisessentialtodetermineuptowhichpositionofaprogrameachPVRhasplayedortransmittedtoavoidtheplayingorstoringdiscontinuitycausedbyincorrectpositioning. Toaddressthisprogramstoragediscontinuity,wehavedevelopedanewprogrampositionrepresentationthatcanidentifythespecicpositionofthesameprogramamongPVRs.ThebroadcastingtimeissynchronizedbetweenabroadcastingserverandSTBswithtimetablessuchasSystemTimeTable(STT)andTimeDateTable(TDT)inATSCandDVBstandards,respectively.Thetimetablesaresupposedtobetransmittedeveryonesecond,multiplexedinbroadcastMPEG-2streams.Sincethetablesarebroadcastatxedpositionsinthestream,theyaresuitableforreferencetimepositions.Wethereforerepresentprogrampositionswithatimecontainedinatableandarelativebyteosetfromthetimeinsteadofusingonlybyteosetfromthebeginningofaprogramle.Figure 5{3 illustratesaspecicprogrampositionina(time,relativeleoset)representation.

PAGE 92

Table5{1. Denitionofsymbolsusedforourschedulingschemes Symbol Denition pSi cRDi cRIi pLi 5{1 liststhesymbolsusedforschedulingschemesinthissection.Theprexescandpusedinsomesymbolsindicateasetofsystemcomponents,suchasPVRsandnetworkdisks,andasetofTVprograms,respectively. AftertheCMSreceivesaprogramtransmissionrequestfromasystemcompo-nent,ittriestondanothercomponenttobestmatchwithitasacommunicationpairbyinvestigatingwhatprogramseachcomponentisstoringandtransmitting.Eachsystemcomponentcanalsoissuethematchingrequestwhentheexisting

PAGE 93

Figure5{4. Anexampleofmatchingacommunicationpair. full-duplextransmissionshouldbechangedintohalf-duplextransmissionduetotheterminationofoneprogramhavingbeingplayedbackbytheothercomponentofthecommunicationpair.Figure 5{4 describeshowtheschemeworks. SupposethatCirequestspa.WerstcheckwhetherornotcRDiTcSaisempty.TheresultsetindicatesasetofcomponentsthatarecurrentlystoringpaandreceivingatleastoneprogramofpSidirectlyfromCi.Ifitisnotempty,itmeansthatthereexistsatleastonecomponenttocommunicatewithCiinfull-duplexmode.Wechoosethecomponentwiththelongestoverlappingtimewithpaamongthem.InFigure 5{4 (a),therearetwocomponentsbelongingtotheresultset,CjandCk.Sincepbhaslongeroverlappingtimewithpathanpc,wechooseCj,whichistransmittingpbtoCi.Then,CiandCjstarttransmittingandreceivingpbandpabetweeneachothersimultaneously. However,iftheabovesetisempty,itisnotpossibletodirectlymatchacom-municationpairsinceCiisnotcurrentlytransmittinganyprogramtothesystemcomponentsbelongingtocSa.Thus,weneedtondanothersetofcomponents

PAGE 94

thataretransmittingprogramsbelongingtopSi,sothatCicantakeoverthetrans-missionresponsibilityofoneoftheprograms.Inotherwords,thematchismadebyindirectlysearchingacommunicationpair.WeobtaincRIibydenitionasfollows: thenwecheckifcRIiTcSaisempty.Ifnot,itmeansthatatleastonecomponentisreceivingatleastoneprogramofpSifromcomponentsotherthanCi.Wechooseonecomponentbasedonthelengthofoverlappingtimebetweenprogramsaswell.Afterchoosingonecomponent,Citakesoverthetransmissionresponsibilityoftheprogramtothechosencomponentinplaceoftheothercomponent,sinceCiisstoringtheprogram.InFigure 5{4 (b),therearealsotwocomponentsbelongingtotheset,CmandCn.WechooseCmsincepdtransmittedtoCmfromCohaslongeroverlappingtimewithpathanpf.OnceCogivesuptransmittingpd,CiandCmstartcommunicatingpdandpasimultaneously. Ifbothoftheaboveintersectionsareempty,wearenotabletomakeanymatchwithCi.Thus,wechooseonecomponentfromcSasothatthecomponentwillhavethemostavailablediskbandwidth.Eventhoughinthiscasewearenotabletofurtherutilizenetworkbandwidth,wecanachievetheeectofdiskloadbalancingamongcomponents.InFigure 5{4 (c),therearethreecomponentsbelongingtocSa,Cp,Cq,andCr.WechooseCpbecauseithasthemostavailablediskbandwidth.

PAGE 95

Figure5{5. AnexampleofswitchingtransmissionresponsibilitybetweentwoPVRs. withstripingortoreplicatepopularprograms.Asmentionedabove,however,thoseload-distributiontechniquescannotbeapplieddirectlytoourarchitecturewhereeachPVRincludesonlyaverysmallnumberofdisksanditisundereachuser'scontroltodecidewhichprogramstostore.Thus,itisnotpossiblefortheCMStoecientlyplacedataoverallthePVRswithoutpermission.Thisimpliesthatweneedadynamicload-distributionschemethatissuitableforourarchitecture. Toachieveload-balancingamongthesystemcomponents,wersttrytoswitchplaybackrequestsandthenreplicatepopularprogramstothepoolofnetworkdisks.Beforethenumberofrequestsforaspecicprogramreachesacertainpointrequiringload-balancingschemes,theCMSchoosesapropersystemcomponenttoprocesstherequestsbasedontheavailablediskbandwidthofeachcomponentortheaboveschemetomatchacommunicationpair. Fortheloadsonaspecicprogramtobebalanced,however,werstperformtheswitchingschemetominimizethenumberofreplications,whichrequireadditionaldiskandnetworkbandwidth.Figure 5{5 illustrateshowtheswitchingis

PAGE 96

performed.SupposethatCsrequestspa.SincethesystemcomponentsbelongingtocSadonothaveavailablediskbandwidthbeforereplicatingpatothepoolofnetworkdisks,wetrytomakeroomforthenewrequestforpabyndinganothersystemcomponenttotakeoveroneoftheexistingrequestsbeingtransmittedbyoneofcSa.Todoso,weobtainpUTa,asetofprogramsthatcSaaretransmitting,asfollows: InFigure 5{5 ,cSa=fCi;Cj;CkgarecurrentlytransmittingpUTa=fpb;pc;pd;pe;pf;pgg,whichbecomecandidatestobeswitchedwithpa.WethenndcUSathatarestor-ingpUTabydenitionasfollows: InFigure 5{5 ,theprogramsbelongingtopUTaarestoredoncUSa=fCm;Cn;Cog.WechooseonecomponentamongcUSawhosediskbandwidthismostavailable.Thus,oneofthesystemcomponentsbelongingtocSacanacceptthethenewrequestforpabyhandingoneoftheexistingrequestsovertothechosencompo-nent.Inotherwords,theexistingrequestisswitchedtothechosencomponenttomakeroomforthenewrequestforpa.InFigure 5{5 ,Cmischosenbasedonthecriterion.AnexistingrequestforpbisswitchedtoCmfromCisincepbisstoredonCmaswellasbeingtransmittedbyCi.TheincreaseddiskbandwidththroughtheswitchingofpbeventuallyenablesCitoacceptthenewrequestforpa. NotethatfortheswitchingtheleastloadedcomponentisalwayschosenamongcUSasothattheloadofeachcomponentcanbemaintainedatasimilardegree.Furthermore,whenthereplicationworkstogetherwiththeswitching,italsoenhancestheloaddistributionofthesystem.Onceaprogramisreplicated,therearemorechancesthatnetworkdiskscanhandleallfuturerequestsforthe

PAGE 97

replicatedprogramsincetheyusuallyhavemorediskbandwidththanPVRs.Forthesamereason,networkdisksalsotendtotakeovertheexistingrequestsforthereplicatedprogramthroughtheswitchingschemesothatPVRscanaccepttherequestsforotherprograms.Accordingly,throughswitchingandreplicationtheseload-balancingcapabilitiescanimprovethesystemperformancebysignicantlyreducingthenumberofreplications. Ifwecannotacceptthenewrequestforpabyswitching,however,weshouldeventuallyreplicatepa.WeemployadynamicreplicationpolicytoreectsuchP2Psharingcharacteristics[39].Weevaluatethefollowingconditiontriggeringswitchingandreplication:thesumofthediskbandwidthnecessaryforbothonereplicationofpaandexpectedrequestsforpaduringthereplicationexceedsthetotalavailablediskbandwidthofcSa. Toobtainanequationrepresentingthiscondition,werstcalculatethetotalavailablediskbandwidthofcSa.LetDiebandDiudenotetheeectiveanduseddiskbandwidthofCi,respectively.Givenadiskblocksize(Bi),theseektime(Dist),rotationallatency(Dirl),anddatatransferrate(Ditr)ofCi,Diebcanbecomputedasfollows: whereDist,Dirl,andDitrrepresentaveragevaluestoreectcommonsituations.Also,wecanobtainDiusimplybyaddingdiskbandwidthusedforR,Pl,Pp,Pd,oCp,iCp,oCd,andiCd.Thus,thetotalavailablediskbandwidthofcSaiscomputedasPCi2cSa(DiebDiu). Toadmitthesubsequentrequestsforpaassoonaspossible,allthePVRsbelongingtocSatransmitpainparallelatafasterrate.Thus,thetimenecessarytoreplicatepaisda=(n(cSa)),whereda,,andn(cSa)denoteadurationofpawhendisplayedatanormalrate,theactualreplicatingratecomparedtoa

PAGE 98

normalrate,andthenumberofPVRsbelongingtocSa,respectively.Giventheexpectedrequestrateofpa=a,theexpectednumberofrequestsforpauntilthereplicationiscompletedis(da=(n(cSa)))a.ThediskbandwidthrequiredfortheexpectedrequestscanbeobtainedbymultiplyingRabytheexpectednumber.Inaddition,sincepaisreplicatedtimesasfastasthenormalplaybackrate(Ra),thediskbandwidthnecessaryforthereplicationisRa.Therefore,thefollowinginequalityrepresentstheswitchingandreplicationtriggeringcondition: Onceaprogramischosen,werstcheckiftherearePVRsthataretunedtothechannelfortheprogram.Ifso,weconsiderthosePVRstunedtotheprogramasasetofcandidatePVRstotransmittheprogramfortime-shiftingtothepoolofnetworkdisks;otherwiseallthePVRsbecomecandidates.Onceasetof

PAGE 99

candidatePVRsisxed,wechooseonePVRwhosetunersarelikelyonaveragetobemostavailableduringitsbroadcastingtime.ForeachcandidatePVR,wedividethetotalremainingdurationoftheprogram(R)intontimeperiodsthatareseparatedbysomeeventsthataecttuneravailability.Forexample,assoonasaPVRnishesrecordingaprogramortransmittingfortime-shifting,onemoretunerwillbecomeavailable.Ontheotherhand,thenumberofavailabletunerswilldecreasebyoneafterthereservedrecordingstarts.Thus,wecanpredictthetuneravailabilityofeachPVRbasedontheinformationonscheduledeventsbythetimetheprogramstartstobebroadcast.However,thepredictionaccuracymightdependonadditionaluserinteractionsthataecttuneravailabilityduringitsbroadcastingtime.Foreachtimeperiodtifori=1;2;:::;n,wecanthenobtainbothtuneravailabilityTA(ti)bysubtractingthenumberofoccupiedtunersfromthetotalnumberoftunersandtime-weightingTW(ti)bydividingRbyti.Weeventuallyobtaintime-weightedTAbyaccumulating(TA(ti)TW(ti))foreachperiodandchoosethePVRwiththelargesttime-weightedTA. NotethateachPVRcanrecordnumerousprogramsfromdierentchannelsatthesametimewithonlysmalldelaysfromtheirbroadcastingtimesbysharingthemthroughthepoolofnetworkdisks.ThisbringsintheeectthateachPVRvirtuallyextendssignicantlyitsnumberoftunersforconcurrentrecordingsofmultiplechannels.

PAGE 100

Figure5{6. Maximumdelayinswitchingtransmissionresponsibilityofatime-shiftedprogrambetweentwoPVRs. forthechannelisbeingswitchedtoanotherPVR.Inthesecondcase,however,wemaymisssomeprogramdatabecauseitisnotpossibletopredictwhenusersissuesomeinteractionsthatuseuptunersforthetransmissionoftime-shiftedprograms.Therearetwopossibilitiesforthiscase.OnepossibilityisthatthereisnootherPVRtunedtotheprogram.SincesomedelayincludingFC-ALarbitrationtimeanddatatransmissiontimewilloccurwhilethetransmissionresponsibilityisbeingswitchedbetweentwoPVRs,itisinevitablethatsomedataoftheprogramwillbemissedduringtheswitching.TheotherpossibilityisthatthereisatleastonePVRtunedtotheprogram.IfoneofthosePVRstakesoverthetransmissionresponsibilityoftheprogram,itispossibletoguaranteethestoragecontinuityoftheprogramaslongasthePVRisbueringthedatabroadcastduringtheswitching. Wethusanalyzethebuerrequirementforcontinuousstorageoftime-shiftingprogramsforthiscase.Figure 5{6 showsthemaximumdelayfortransmissionswitchingofaprogrambetweentwoPVRs.Toensurethestoragecontinuityofthe

PAGE 101

programforanycase,weassumethefollowingworstcasescenarioincomputingthedelay:whenthenumberofsystemcomponentsattachedtotheFC-ALisN,thesourceandtargetPVRsfortheswitchinghavethehighestFC-ALaddresses,N1andN,respectively(deviceswithloweraddresseshavehigherprioritiesintheFC-AL).SupposethatCN1iscurrentlytransmittingpatothepoolofnetworkdisks(denotedbyNDsinFigure 5{6 ),butitshouldpassthetransmissionresponsibilitytoanotherPVRattimet0becauseitrunsoutoftunersduetoanadditionalrequestfortheuseofatuner.Sincethisisnotpredictable,CN1shouldimmediatelynotifytheCMSofthiseventwiththeinformationonlatesttransmittedprogrampositionsothatthetransmissionresponsibilitycanbetakenoverbyanotherPVRasquicklyaspossible.TotransferthisinformationdatatotheCMS,CN1participatesinarbitrationtogetFC-ALaccess.SinceCN1hasverylowpriority,itshouldwaituntilalmostallthesystemcomponentsacquireloopaccessandnishtransmittingdata.Afterobtainingtheloopaccess,CN1sendstheCMStheprogrampositionlatesttransmittedtothepoolofnetworkdisks.TheseoperationsoccurduringPeriod(1). TheCMSthendecideswhichPVRwilltakeoverthetransmissionresponsibil-itybasedontuneravailabilityofeachPVR.TheCMSchoosesCN,assumingthatitistunedtopaandthenumberofavailabletunersisgreatestonaverageduringthebroadcastingtimeofpa.Afterthat,theCMSsendsCNarequesttoresumetransmittingpafromthelateststoredposition.SincetheCMSdoesnotusethefairnessarbitrationalgorithmtotransfercommanddataimmediately,itcanacquiretheloopaccessandtransmittherequestpacketveryquickly.ThistransferoccursduringPeriod(2). Afterreceivingtherequest,CNeventuallyparticipatesinarbitration,acquirestheloopaccess,andstartstransmittingpatothepoolofnetworkdisksagain.

PAGE 102

However,itmustalsowaitfortheloopaccessuntilallthePVRsnishtransmit-tingtheirdatasinceithasthelowestpriority,whichoccursduringPeriod(3).Inaddition,CNneedstotransmitpaataslightlyfasterrateuntilitcatchesupwiththeliveprogramposition(Catch-upPeriod).Lateron,itcanreachthelivebroadcastingpositionofpa(LivePeriod).Toavoidmissingsomedatabroadcastduringtheswitchingoftransmissionresponsibility,CNthereforeshouldkeepstor-ingbroadcastdataofpaintothebueruntilitstartstransmittingthedatatothepoolofnetworkdisks(i.e.,duringPeriods(1),(2),and(3)). Letusevaluatethemaximumdelayfortheswitchingoftransmissionresponsi-bilityforpa.Eachloopaccesstimeconsistsofarbitrationtime,openingconnectiontime,closingconnectiontime,anddatatransfertime[6].WhentransferringdatabetweentwosystemcomponentsovertheFC-AL,thetransmissiondelayisdeter-minedbytheloopoverhead,suchasper-nodedelayandpropagationdelay,anddatatransfertime.Theper-nodedelay(Lnd)isaddedforeachhopbetweenthesendingandreceivingdevicesandthepropagationdelay(Lpd)isdeterminedbythephysicaldistance.Inaddition,datatransfertimeisaectedbytheFC-ALdatatransferrate(Ltr),theamountofdatatobetransferred,andthenumberofreceiverbuersforowcontrol. Threemessageswitchings,thearbitration,theopeningconnection,andtheclosingconnection,requiretheentirelooptriptime,whichiscomputedasshowninEq. 5{6 .Lar,Lopn,andLclsdenotethearbitrationtime,theopeningconnectiontimeandtheclosingconnectiontime,respectively.NisthetotalnumberofsystemcomponentsattachedtotheFC-AL. TotransmitoneblockovertheFC-AL,theblockshouldbedividedintoframessincetheFC-ALtransmitsdatainunitsofframe.Thus,thenumberofframes

PAGE 103

belongingtooneblock(B=F)dependsontheblocksize(B)becausetheframesize(F)isxed.Inaddition,denotesaconstantofoverheaddegreecausedbytheowcontroldependingonthenumberofreceivebuers.Nhdenotesthenumberofhopsbetweentwosystemcomponents.Thus,theFC-ALoverheadtimenecessarytotransmitoneblock(Lbo)is F((NhLnd)+Lpd):(5{7) Ontheotherhand,sinceacommandsize(C)isusuallymuchshorterthantheframesize,theoverheadtimeforonecommand(Lco)is Thedatatransmissiontimeiscalculateddependingondatasizetobetrans-mitted.Thus,itshouldbeB=LtrandC=Ltrforoneblockandonecommand,respectively. Assumingthatallthesystemcomponentsparticipateinarbitration,thelengthofeachperiodisdeterminedbythesumofarbitrationtimeanddatatransmissiontimeofeachsystemcomponent.Thus,weobtainthefollowingmaximumdelayfortheswitchingofthetransmissionresponsibilitysincethereareN1blocksandonecommandtobetransmittedduringPeriod(1),onecommandduringPeriod(2),andNblocksduringPeriod(3): Delay=(2N+1)(Lar+Lopn+Lcls)+(2N1)Lbo+B Ltr+2Lco+C Ltr:(5{9) Therefore,aslongaseachPVRhasaslargeabuerasissucienttosustaintheabovedelay,wecanguaranteestoragecontinuityofeachprogramatanytime.Wenallyobtainthebuersizenecessaryforthetransmissionswitchingofan

PAGE 104

Table5{2. SummaryofsimulationparametersforTVcontentsharing Category Parameter Value FC-AL datatransferrate 400MB/s propagationdelay 4ns/meter pernodedelay 6words(240ns) distance 200meters Disk capacity 120GB datatransferrate 22.9{47.5MB/s seektime 1.1{16ms rotationallatency 4.17ms HD-qualityTVprogramasfollows: Buffer Size=Max DelayRhd;(5{10) whereRhddenotestheplaybackrateofanHD-qualityprogram. ThearrivalrateofprogramplaybackrequestsfollowsaPoissondistributionreectingvaryingaveragevalues.Eachrequestcanbeacceptedaslongasthecorrespondingresourcesareavailable;otherwiseitwillberejected.Theobtainedsimulationresultsarebasedontheassumptionthatthetolerablerejectionratiois5%.Itisalsoassumedthat100PVRsareconnectedtoaninternalFC-ALnetworkandhalfofthemonaverageareavailableataspecictime.Thenumberofchannelsprovidedis50. Table 5{2 showsthesimulationparametersusedfortheFC-ALanddisks.TheFC-ALhasadatatransferrateof400MB/s,whichatthistimeisavailable.The

PAGE 105

Figure5{7. Eectofmatchingacommunicationpair. per-nodedelayoftheinterfacetoforwardaframeis6words(i.e.,240nanosec-onds).Thepropagationdelayandmaximumdistancebetweentwodevicesare4nanoseconds/meterand200meters,respectively.ThediskmodelisbasedontheIBMDeskStar120GXPmodelwhosecapacityis120Gbytes.Itsaveragedatatransferrate,seektime,androtationallatencyare35.2MB/s,8.8ms,and4.17ms,respectively. WeassumethatbroadcastTVprogramsconsistof50%HD-quality(19.4Mbps)and50%SD-quality(5Mbps)programs.Thus,theaverageprogramplaybackratewillbeapproximately12.2Mbps.Eachprogramisonaverage50minuteslong.Basedontheseaveragevalues,eachPVRthereforecanstoreupto22programs.Zipfdistributionisusedtodeterminetherequestfrequencyforeachprogramwithaparameterof0.271whichmodelshighlyskewedpopularitydistribution(how-ever,duetolackofanactuallarge-scaletrial,therealdistributionisatpresentunknown).Weconductedthesimulationsfor24hourswhilevaryingseveralpara-meters,suchasthenumberofdisks,thearrivalrateofrequests,andthediskblocksize,toseetheimpactofeachparameteronsystemperformance.

PAGE 106

Figure5{8. Thedataportiontransmittedinfull-duplexmodeusingourmatchingscheme. Figure 5{7 showstheFC-ALutilizationineachcaseasthenumberofprogramtransmissionrequestsovertheFC-ALincreases.Forconvenienceofcomparison,wedenetheFC-ALutilizationastheutilizationofacommunicationchannelthatiscurrentlyaperformancebottleneck(i.e.,beingusedmore).Inotherwords,onlythecriticalchannelisconsideredwhenevaluatingtheutilizationeventhoughbothchannelsarebeingutilized.WecanseethataftertheFC-ALchannelutilizationreachesapproximately86%,itstartstobesaturatedineachcaseduetotheowcontrolmechanism[19].Mostimportantly,thematchingschemecansupportmany

PAGE 107

Figure5{9. PercentageofmatchingsformedindividuallyfromRDiandRIi. morerequestswhentheFC-ALutilizationisequal.Forexample,inacasewhereFC-ALutilizationis86%,233requestsaresupportedwhenmatchingisnotusedwhilethematchingschemeextendsthenumberofsupportedrequeststo323,whichresultsinabouta39%performanceimprovement.ThisisbecausethebandwidthofbothchannelsoftheFC-ALcanbefurtherutilizedbymatchingasmanycommunicationpairsaspossible.Figure 5{8 illustratesthisresult,showingthedataportiontransmittedinfull-duplexmodeovertheFC-AL.Thedataportionatfull-duplextransmissionreachesasmuchas63%when323requestsaresupported.Datacannotbetransmittedinfull-duplexmodeunlessthematchingschemeisused.Itcanalsobeseenthatthedataportiontransmittedinfull-duplexmodecontinuestoincreaseasthenumberoftransmissionrequestsincreases.Whenthenumberofrequestsis20,thedataportionis14.5%,butitgoesupto62.1%for275requests.Thereasonforthisisthatwithincreasedtransmissionrequests,theprobabilityofndingcommunicationpairsincreasesaccordingly. Inaddition,itisworthnotingthatmorematchingpairswereformedfromcRDithanfromcRIi.Figure 5{9 illustrateseachpercentageofthematchingsformedfromcRDiorcRIi.TheaveragepercentageofmatchingsfromcRIiisabout74%,whilethepercentagefromcRDiis26%.TheratioforbothcRDiandcRIiis

PAGE 108

Figure5{10. Eectofswitchingplaybackrequests. similarinallthenumbersofrequests.ThisindicatesthatmatchingcommunicationpairsthroughindirectsearchingismuchmoreeectivethansimpledirectmatchingsinceindirectsearchingcansignicantlyextendthenumberofmatchingcandidatesandisnotlimitedtothesystemcomponentsthatarereceivingsomeprogramsdirectlyfromCi. Figure 5{10 showshowmanydiskaccessrequestseachPVRcanissueonaverageforanhourwhilevaryingthenumberofnetworkdisks.ItcanbeseenthateachPVRcanissueabout18%morediskaccessrequestswithbothswitchingand

PAGE 109

Figure5{11. Improvementinnumberofreplicationsbyswitching. Figure5{12. AverageFC-ALbandwidthsavedbyswitching. replicationschemesthanwithonlyareplicationscheme.Forexample,whensixnetworkdisksareused,thesupporteddiskaccessrequestrateperPVRis10.1fortherstcaseand8.56forthesecondcase.Sincethereare50PVRsinthenetwork,thetotalnumberofrequestssupportedincreasesby77.Thisdemonstratesthattheeectofswitchingplaybackrequestsisconsiderable.Thefollowingguresshowthereasons. Figure 5{11 illustratestheaveragenumbersofreplicationsforanhourwithtwocases.Forafaircomparisonoftwocases,thediskaccessrequestratesperPVRaresetequaltotheonesusedinthesecondcaseofFigure 5{10 (i.e.,the

PAGE 110

Figure5{13. Numberofswitchingsvs.replications. replicationonlycase)dependingonthenumberofnetworkdisks.Theaveragenumberofreplicationsperhourwaslessthan0.1forbothswitchingandreplicationschemes,whileitwas9.2foronlyreplicationscheme.ThismeansthatforthisrequestratewecanachieveloadbalancingofplaybackrequestsamongPVRswithonlyaswitchingschemesincethereplicationsarerarelyperformed.ThisisbecausetheswitchingschemebyitselfisabletodistributetheloadsamongPVRsquiteevenlybychoosingtheleastloadedPVRasatargetforswitching.Accordingly,theFC-ALbandwidththatcanbesavedisproportionaltothetotalplaybackrateforthereducednumberofreplicationsshowninFigure 5{12 ,andthenumberofnetworkdisksrequiredtostorethereplicatedprogramsalsodecreasestoasimilardegree. Inaddition,toseetheeectofswitchingwhencombinedwithreplication,wemeasuredthenumberofswitchingsandreplications,asshowninFigure 5{13 .ThediskaccessrequestratesperPVRissettotheonesusedintherstcaseofFigure 5{10 (i.e.,whenbothswitchingandreplicationareused).Ourexperimentalresultsshowedthatthenumberofswitchingswasonaverageabout12timesgreaterthanthatofreplications.Thereasonisthatswitchingandreplicationworktogetherinthefollowingway:toreduceasmanyreplicationsaspossiblebyachieving

PAGE 111

Figure5{14. TheaccumulatedaccessfrequencyaccordingtoZipfdistribution. load-balancingamongPVRs,theswitchingsareperformedwheneverthetotalnumberofrequestsforeachprogramexceedsitsthreshold.Ontheotherhand,mostofreplicationsarenotstarteduntilthetotaldiskaccessrequestsapproachthecapacityoftheentiresystemsincetheswitchingshavekeptthemdistributedquiteevenlyamongPVRs.Onceaprogramisreplicated,networkdiskshandlefuturerequestsforthereplicatedprogramsincetheiravailablediskbandwidthtendstobegreatestcomparedtotheindividualPVRs.Asaresult,thePVRsstoringthereplicatedprogramcanacceptotherprogramswiththeavailablebandwidththatshouldotherwisebeusedfortherequestsforthereplicatedprogram.Therefore,wecanseethatthereplicationperformsmuchbetterwhencombinedwiththeswitchingthanwhenitworksaloneeventhoughthesamenumberofreplicationsareperformed,asshowninFigure 5{13 .

PAGE 112

Figure5{15. Time-shiftinghoursvs.hitratio. 5{14 showsthepercentageofaccumulatedaccessfrequencyof50programsaseachprogramisaddedonebyonefromthemostpopularprogramaccordingtotheZipfdistribution.Theaccessfrequencyofthemostpopularprogramisgreaterthan13%,whilethatoftheleastpopularoneislessthan0.8%.Thus,itcanbeseenthatweneedtostoreonly20%ofthemostpopularprogramssothatthepercentageofaccumulatedaccessfrequencycanreach50%. Basedontheskewedaccessfrequencydistribution,wehavedeterminedhowmanyhoursoftime-shiftingthepoolofnetworkdiskscanprovidedependingonthedesiredhitratiowhilevaryingthenumberofdisksfrom1to30,asshowninFigure 5{15 .Fromtheexperiments,wehaverstfoundoutthat55,65,75,85,and95%hitratiocanbeobtainedbystoringabout23,35,49,66,87%,respectively.Inaddition,asthenumberofdisksincreasedineachhitratiocase,thetime-shiftinghourswereextendedasexpected.Forexample,thepoolofnetworkdiskscouldholdthemostpopular23%oftheprogramsforabout59hourswith30disks(i.e.,

PAGE 113

Figure5{16. DelayinswitchingtransmissionresponsibilityofaprogrambyFC-ALtransmissionsize. inthe55%hitratiocase),whiletheycouldholdthemforonlyabout2hourswithonedisk.Itcanalsobeseenthatthetime-shiftinghourssupportedinthe55%hitratiocasearefourtimeslongerthanthoseinthe95%hitratiocaseeventhoughthedierenceofhitratioisonly40%.Thetime-shiftinghoursincreasemuchmorerapidly(byapproximately10times)comparedtothedierenceofhitratio.Thisisduetotheeectofthesignicantskewedprogrampopularitydistribution.Ineachcase,somesmalluctuationsofcurvesoccurredbecausethecompositionratiooftime-shiftedprogramswithdierentqualitiescanvaryineachsimulationrun.Sinceuserstendtorecordandwatchrecentlybroadcastprograms,itthereforeisnecessarytondtheproperhitratioandnumberofdisks,consideringsuchTVviewers'accesspatterns. 5{16 and 5{17 showmaximumdelays

PAGE 114

Figure5{17. Requiredbuersizeforstoragecontinuityofatime-shiftedprogram. andrequiredbuersizesforprogramstoragecontinuity,respectively.EachPVRtransmitsonediskblockwithoneFC-ALarbitration.Withthesmalltransmissionsize,althoughwecanreducetheresponsetimewithonlyasmallbuer,thediskperformancecanbedegradedsignicantlyduetotheseektimeandrotationallatency,whichiscriticalforoverallsystemperformance.Ontheotherhand,whenalargetransmissionsizeisused,wecanimprovethediskperformancewhilealargerbuerforthelongerdelayisrequired.Intheworstcase,aslongasalltheprogramsthateachPVRiscurrentlytunedtoarealwaysbueredforatmost0.7second(i.e.,1.7MBbuerpertunershouldbereservedforthispurpose),wecanavoidthedisplaydisruptionforthetime-shiftingprogramscausedbydatamissingduringtheswitching.

PAGE 115

experimentalresultssupportedtheireectiveness.Byplacingnetworkdisksinthenetwork,itcanbeseenthatPVRscansharetime-shiftedprogramsbetweeneachother,extendingtime-shiftinghourssignicantly.CombinedwithnetworkedPVRsandhigh-speednetworks,ourarchitectureisexpectedtoprovideanenvironmentwherepeoplecanwatchandsharewhattheywantwhentheywantthroughmanydierentchannels.

PAGE 116

Thisdissertationmadesomecontributionsfordistributedmultimediaservicesbasedonnetworkedstoragesystemsbyprovidingbueringschemesandahigh-speedstreamingarchitecturetoutilizesystemresourcesinecientways.WeadoptedtheFC-ALtechnologyasahigh-speedstorageandstreamingnetworktoconnectthesystemcomponentsinvolvedinsuchenhancedservices.Intherstpartofthisdissertation,weproposedanecientprefetchingschemetoexploitthediskcacheusingthefairnessarbitrationalgorithm.Inthesecondpart,wefurtherextendedthebueringschemebyintegratingtheintervalcachingtechniqueintotheprefetchingtechnique.Weshowedthattheproposedschemecouldsimultaneouslytakebenetofbothprefetchingandcachingbasedondynamicthresholdvaluestodeterminewhethertoprefetchortocachestreams.Sincemoderndiskscontinuetohavelargercachememory,theproposedschemesarethereforeexpectedtoimprovetheperformanceevenmoreinthefuture.Inthethirdpart,weproposedanovelintegratedstorage/networkarchitecturetosharehigh-qualitybroadcastTVcontentsbetweenconsumerdevices.Wealsoproposedschedulingschemesofmatchingcommunicationpairsandswitching/replicationforthesystemload-balancingtofurtherutilizesystemresources. Ibelievethattherearestillmoreissuesthatweneedtoaddresstofurtherimprovetheperformanceofdistributedmultimediasystems.Therefore,Iwillcontinuetoworkonnetworkedstoragesystems,includinganothernetworkedstoragesystem,Network-AttachedStorages(NASs),inmyfutureresearch.Amongmanypotentialtopics,thefollowingtwoissueswillbegivenhigherpriorities:therstissueishowtosupportuserinteractiveVCRfunctionssuchasfast 104

PAGE 117

forward/rewindinmultimediaserversbasedonnetworkedstoragesystems.TherearevariousschemesforsupportingVCRfunctionsinVODsystems,butweneedtoaddresstheissuesoccurringwhentheyareincorporatedwithourbueringscheme.OnceVCRfunctionsareused,theaccesspatternstovideolesarenolongersequential,whichmayalsocausetheload-unbalancingproblemwithoutproperdataplacementandschedulingschemes.WealsoneedsomeschemestoaddresstheresponsetimeofVCRfunctionstoreducethedelayasmuchaspossible.Thesecondissueishowtoexploitsystemresourcesofmultimediaserversbasedonnetworkedstoragesystemswhenheterogeneousdisksareemployed.Sinceinsuchaenvironmenteachdiskhasdierentdiskcapacityanddatatransferrate,weshouldmodifytheproposedschedulingpolicybyadjustingdiskblocksizeandnetworktransmissionblocksizeindividually.Withourcurrentschedulingpolicy,itisverydiculttopredictthenetworkacquisitiontimeofeachdiskduetosignicantvariationofeachdisk'saccesstime.

PAGE 118

[1] G.A.GibsonandR.V.Meter.Networkattachedstoragearchitecture.CommunicationsoftheACM,43(11):37{45,2000. [2] ANSI(AmericanNationalStandardsInstitute).FiberChannelArbitratedLoop(FC-AL-2).WashingtonD.C.,2001. [3] A.F.Benner.FibreChannelforSANs.McGraw-Hill,NewYork,USA,2001. [4] T.Clark.DesigningStorageAreaNetworks.AddisionWesley,Boston,USA,2003. [5] E.KimandJ.C.L.Liu.Anintegratednetwork/storagearchitectureforsharinghigh-qualitybroadcastTVcontents.IEEECommunicationsMagazine,43(9):86{93,2005. [6] E.KimandJ.C.L.Liu.Time-awareprefetchingforon-demandvideoservicesinaresidentialservicegateway.MultimediaToolsandApplicationsJournal,DOI:10.1007/s11042-006-0017-3,2006. [7] D.H.C.Du,J.C.L.Liu,J.Hsieh,andR.J.Vetter.Buildingvideo-on-demandservers.SpecialIssueonMultimediaforJournalofTelecommunicationSystems,9(3-4):255{286,1998. [8] D.H.C.Du,T.Chang,J.Hsieh,S.Shim,andY.Wang.Twoemergingserialstorageinterfacesforsupportingdigitallibraries:SerialStorageArchitecture(SSA)andFiberChannel-ArbitratedLoop(FC-AL).MultimediaToolsandApplication,10(2):179{203,2000. [9] P.TriantallouandS.Harizopoulos.Prefetchingintosmartdiskcachesforhighperformancevideoservers.InProc.IEEEMultimediaComputingandSystems,pages800{805,Florence,Italy,1999. [10] J.M.Almeida,D.L.Eager,andM.K.Vernon.Ahybridcachingstrategyforstreamingmediales.InProc.SPIE/ACMConf.onMultimediaComputingandNetworking,pages200{212,SanJose,USA,2001. [11] A.Dan,D.M.Dias,R.Mukherjee,D.Sitaram,andR.Tewari.Bueringandcachinginlarge-scalevideoservers.InProc.IEEEInternationalComputerConf.,pages217{225,SanFrancisco,USA,1995. 106

PAGE 119

[12] A.DanandD.Sitaram.Ageneralizedintervalcachingpolicyformixedinteractiveandlongvideoworkloads.InProc.MultimediaComputingandNetworkingConf.,pages344{351,SanJose,USA,1996. [13] N.J.SarhanandC.R.Das.CachingandschedulinginNAD-basedmultimediaservers.IEEETransactionsonParallelandDistributedSystems,15(10):921{933,2004. [14] W.ShiandS.Ghandeharizadeh.Controlledbuersharingincontinuousmediaservers.MultimediaToolsandApplications,23(2):131{159,2004. [15] C.AmbuhlandB.Weber.Parallelprefetchingandcachingishard.InProc.TheoreticalAspectsofComputerScience,LNCS,volume2996,pages211{221,Montpelier,France,2004. [16] P.Cao,E.W.Felten,A.R.Karlin,andK.Li.Astudyofintegratedprefetchingandcachingstrategies.InProc.ACMSIGMETRICS,pages188{197,Ottawa,Canada,1995. [17] T.KimbrelandA.R.Karlin.Near-optimalparallelprefetchingandcaching.SIAMJournalonComputing,29(4):1051{1082,2000. [18] S.ChenandM.Thapar.AFibreChannel-basedarchitectureforinternetmultimediaserverclusters.InProc.InternationalAlgorithmsandArchitecturesforParallelProcessing,pages437{450,Melbourne,Australia,1997. [19] J.R.HeathandP.J.Yakutis.High-speedstorageareanetworksusingaFiberChannelArbitratedLoopinterconnect.IEEENetwork,14(2):51{56,2000. [20] G.Ma,A.Khaleel,andN.Reddy.Performanceevaluationofstoragesys-temsbasedonnetwork-attacheddisks.IEEETransactionsonParallelandDistributedSystems,11(9):956{968,2000. [21] E.KimandJ.C.L.Liu.EcientschedulingschemesforHD-qualitycontentdistribution.InProc.IEEEConsumerCommunicationsandNetworkingConf.,pages1103{1108,LasVegas,USA,2006. [22] E.KimandJ.C.L.Liu.DesignofHD-qualitystreamingnetworksforreal-timecontentdistribution.IEEETransactionsonConsumerElectronics,52(2):392{401,2006. [23] D.J.Gemmell,H.M.Vin,D.D.Kandlur,P.V.Rangan,andL.A.Rowe.Multimediastorageservers:Atutorial.IEEEComputer,28(5):40{49,1995. [24] T.Kim,H.Bahn,andK.Koh.Popularity-awareintervalcachingformultime-diastreamingservers.IEEElectronicsLetters,39(21):1555{1557,2003. [25] J.Y.B.Lee.Parallelvideoservers-atutorial.IEEEMultimedia,5(2):20{28,1998.

PAGE 120

[26] M.Y.Y.Leung,J.C.SLui,andL.Golubchik.Useofanalyticalperformancemodelsforsystemsizingandresourceallocationininteractivevideo-on-demandsystemsemployingdatasharingtechniques.IEEETransactionsonKnowledgeandDataEngineering,14(3):615{637,2002. [27] D.Wu,Y.T.Hou,W.Zhu,Y.Q.Zhang,andJ.MPeha.Streamingvideoovertheinternet:approachesanddirections.IEEETransactionsonCircuitsandSystemsforVideoTechnology,11(3):282{300,2001. [28] R.R.YehiaandI.O.Mahgoub.Storagesystemandmultimedia:Classicationandextensions.DistributedandParallelDatabases,7(4):429{442,1999. [29] D.H.C.Du,J.Hsieh,T.Chang,Y.Wang,andS.Shim.Interfacecomparision:SSAversusFC-AL.IEEEConcurrency,6(2):55{70,1998. [30] T.M.Ruwart.PerformancecharacterizationoflargeandlongFiberChannelarbitratedloops.InProc.IEEEMassStorageSystemsSymposium,pages11{21,SanDiego,USA,1999. [31] E.KimandJ.C.L.Liu.Exploitationofdiskcacheforvodservicesinnetwork-attachedstoragesystems.InIEEEConf.onInformationTechnology:ResearchandEducation,pages120{124,Hsinchu,Taiwan,2005. [32] E.KimandJ.C.L.Liu.Optimizingprefetchinaresidentialgatewaywithnetworkedstoragesystems.InProc.IEEEConf.onMultimedia&Expo,pages1050{1053,Amsterdam,Netherlands,2005. [33] ANSI(AmericanNationalStandardsInstitute).FiberChannelPhysicalandSignalingInterface(FC-PH).WashingtonD.C.,1994. [34] L.Qiang,X.Huagang,andZ.Qishan.CreditdeterminationforowcontrolinFC-AL.InProc.DigitalAvionicsSystemsConf.,pages24{28,SaltLakeCity,USA,2004. [35] S.Berson,S.Ghandeharizadeh,R.R.Muntz,andX.Ju.Staggeredstripinginmultimediainformationsystems.InProc.ACMSIGMODConf.,pages79{90,Minneapolis,USA,1994. [36] M.S.Chen,D.Kandlur,andP.Yu.Storageandretrievalmethodstosupportfullyinteractiveplayoutinadisk-arraybasedvideoservers.MultimediaSystems,3(3):126{135,1995. [37] P.ShenoyandV.Harric.Ecientstripingtechniquesforvariablebitratecontinuousmedialeservers.PerformanceEvaluation,38(2):175{199,1999. [38] A.DanandD.Sitaram.Anonlinevideoplacementpolicybasedonbandwidthtospaceratio(BSR).InProc.ACMSIGMODConf.,pages376{385,SanJose,USA,1995.

PAGE 121

[39] P.W.K.Lie,J.C.S.Lui,andL.Golubchik.Threshold-baseddynamicreplica-tioninlarge-scalevideo-on-demandsystems.MultimediaToolsandApplica-tions,11(1):35{62,2000. [40] D.N.Serpanos,L.Georgiadis,andT.Bouloutas.MMPacking:aloadandstoragebalancingalgorithmfordistributedmultimediaservers.IEEETransactionsonCircuitsandSystemsforVideoTechnology,8(1):13{17,1998. [41] J.L.Wolf,P.S.Yu,andH.Shachnai.DASDdancing:Adiskloadbalanc-ingoptimizationschemeforvideo-on-demandcomputer.InProc.ACMSIGMETRICSConf.,pages157{166,Ottawa,Canada,1995. [42] C.C.Aggarwal,J.L.Wolf,andP.S.Yu.Onoptimalbatchingpoliciesforvideo-on-demandstorageservers.InProc.IEEEMultimediaComputingandSystems,pages253{258,Hiroshima,Japan,1996. [43] A.Dan,D.Sitaram,andP.Shahabuddin.Dynamicbatchingpoliciesforanon-demandvideoserver.MultimediaSystems,4(3):112{121,1996. [44] S.Sheu,K.A.Hua,andW.Tavanapong.Chaining:ageneralizedbatch-ingtechniquesforvideo-on-demandsystems.InProc.IEEEMultimediaComputingandSystems,pages110{117,Ottawa,Canada,1997. [45] C.C.Aggarwal,J.L.Wolf,andP.S.Yu.Onoptimalpiggybackmergingpoliciesforvideo-on-demandsystems.InProc.ACMSigmetrics,pages200{209,Philadelphia,USA,1996. [46] L.Golubchik,J.C.S.Lui,andR.R.Muntz.Adaptivepiggybacking:anoveltechniquefordatasharinginvideo-on-demandstorageservers.MultimediaSystems,4(3):140{155,1996. [47] A.Hondroulis,C.Harizakis,andP.Triantallou.Optimalcachememoryexploitationforcontinuousmedia:Tocacheortoprefetch?MultimediaToolsandApplications,23(3):203{220,2004. [48] V.Srinivasan,E.S.Davidson,andG.S.Tyson.Aprefetchtaxonomy.IEEETransactionsonComputers,53(2):126{140,2004. [49] J.TseandA.J.Smith.CPUcacheprefetching:Timingevaluationofhardwareimplementations.IEEETransactionsonComputers,47(5):509{526,1998. [50] R.Karedla,J.Love,andB.Wherry.Cachingstrategiestoimprovedisksystemperformance.IEEEComputer,27(3):38{46,1994. [51] V.Soloviev.Prefetchinginsegmenteddiskcacheformulti-disksystems.InProc.ACM/IEEEWorkshoponInput/OuputsinParallelandDistributedSystems,pages69{82,Philadelphia,USA,1996.

PAGE 122

[52] D.A.PattersonK.KeetonandJ.Hellerstein.Acaseforintelligentdisks(IDISKS).SIGMODRecord,27(3):42{52,1998. [53] E.Riedel,G.Gibson,andC.Faloutsos.Activestorageforlarge-scaledataminingandmultimediaapplications.InProc.InternationalConf.onVLDB,pages62{73,NewYork,USA,1998. [54] S.Jung,E.Kim,andD.Lee.DesignandimplementationofanenhancedPersonalVideoRecorderfordtv.IEEETransactionsonConsumerElectronics,47(4):915{920,2001. [55] E.Kim,H.Son,andB.Kang.Designandimplementationofanenhancedper-sonalvideorecorderforhdtv.InProc.IEEEConf.onConsumerElectronics,pages316{317,LosAngeles,USA,2001. [56] B.Marusic,T.Finkst,A.Jere,M.Meza,S.Dobravec,andJ.Tasic.Shareit!-contenttransferinhome-to-homenetworks.InProc.IEEEMediterraneanElectrotechnicalConf.,volume2,pages669{672,Dubrovnik,Croatia,2004. [57] A.J.CahillandC.J.Sreenan.AnecientCDNplacementalgorithmforthedeliveryofhigh-qualityTVcontent.InProc.IASTEDConf.onInternetandMultimediaSystemsandApplications,pages364{369,Grindelwald,Switzerland,2005. [58] C.D.Cranor,M.Green,C.Kalmanek,D.Shur,S.Sibal,J.E.VanderMerwe,andC.J.Sreenan.Enhancedstreamingservicesinacontentdistributionnetwork.IEEEInternetComputing,5(4):66{75,2001. [59] M.Chen,D.Kandlur,andP.Yu.Supportforfullyinteractiveplayoutinadisk-array-basedvideoserver.InProc.ACMMultimediaConf.,pages391{398,SanFrancisco,USA,1994. [60] J.W.DingandY.M.Huang.Resource-basedstriping:Anecientstripingstrategyforvideoserversusingheterogeneousdisk-subsystems.MultimediaToolsandApplications,19(1):29{51,2003. [61] E.Kim,T.G.Kwon,Y.Nah,andSukhoLee.Supportingdynamicrelocationofvideodataindisk-array-basedvideoservers.InProc.Conf.onDatabaseSystemsforAdvancedApplications,pages501{510,Melbourne,Austrailia,1997. [62] T.G.Kwon,Y.Choi,andS.Lee.Diskplacementforarbitrary-rateplaybackinaninteractivevideoserver.ACM/SpringerMultimediasystems,5(4):271{281,1997. [63] M.H.Lee,C.H.Wen,C.Y.Cheng,andY.J.Oyang.Designingafullyinterac-tivevideo-on-demandserverwithanoveldataplacementandretrievalscheme.IEEETransactionsonConsumerElectronics,41(3):851{858,1995.

PAGE 123

[64] Y.WangandD.H.C.Du.Weightedstripinginmultimediaservers.InProc.IEEEConf.onMultimedia.ComputingandSystems,pages102{109,Ottawa,Canada,1997. [65] A.Dan,M.G.Kienzle,andD.Sitaram.Dynamicsegmentreplicationpolicyforloadbalancinginvideo-on-demandservers.MultimediaSystems,3(3):93{103,1995. [66] N.VenkatasubramanianandS.Ramanathan.Loadmanagementindistributedvideoservers.InProc.Conf.onDistributedComputingSystems,pages31{39,Baltimore,USA,1997. [67] J.C.L.Liu,D.H.C.Du,andJ.A.Schnepf.Supportingrandomaccessonreal-timeretrievalofdigitalcontinuousmedia.ComputerCommunications,18(3):145{159,1995. [68] Y.H.Chu,J.Chuang,andH.Zhang.Acasefortaxationinpeer-topeerstreamingbroadcast.InProc.ACMSIGCOMMWorkshoponPracticeandTheoryofIncentivesinNetworkedSystems,pages205{212,Portland,USA,2004. [69] M.Hefeeda,A.Habib,B.Botev,D.Xu,andB.Bhargave.PROMISE:peer-to-peermediastreamingusingCollectCast.InProc.ACMMultimediaConf.,pages45{54,Berkeley,USA,2003. [70] K.Kalapriya,S.K.Nandy,andK.VenkateshBabu.Canstreamingofstoredplaybackvideobesupportedonpeertopeerinfrastructure?InProc.Conf.onAdvancedInformationNetworkingandApplication,pages200{203,Fukuoka,Japan,2004. [71] R.K.RajendranandD.Rubenstein.OptimizingthequalityofscalablevideostreamsonP2Pnetworks.InProc.SIGMETRICS,pages396{397,NewYork,USA,2004. [72] S.Sen,J.Rexford,andD.Towsley.Proxyprexcachingformultimediastreams.InProc.INFOCOMconf.,pages1310{1319,NewYork,USA,1999. [73] B.Wang,S.Sen,M.Adler,andD.Towsley.Optimalproxycacheallocationforecientstreamingmediadistribution.InProc.INFOCOM,pages1726{1735,NewYork,USA,2002. [74] K.Wu,P.S.Yu,andJ.Wolf.Segment-basedproxycachingofmultimediastreams.InProc.WWWConf.,pages36{44,HongKong,2001. [75] Z.L.Zhang,Y.Wang,D.H.C.Du,andD.Su.Videostaging:aproxy-server-basedapproachtoend-to-endvideodeliveryoverwide-areanetworks.IEEE/ACMTransactionsonNetworking,8(4):429{442,2000.

PAGE 124

[76] R.Bhagwan,S.Savage,andG.M.Voelker.Understandingavailability.InProc.InternationalWorkshoponPeer-to-PeerSystems,pages256{267,Berkeley,USA,2003. [77] K.LabonteJ.ChuandB.Levine.Availabilityandlocalitymeasurementsofpeer-to-peerlesystems.InProc.ofSPIE,volume4868,pages310{321,Boston,USA,2002. [78] S.Saroiu,P.K.Gummadi,andS.D.Gribble.Ameasurementstudyofpeer-to-peerlesharingsystems.InProc.MultimediaComputingandNetworking,pages156{170,SanJose,USA,2002. [79] J.Yoon,H.Kim,S.Chun,J.Kim,andS.Sull.Real-timevideoindexingsystemforlivedigitalbroadcastTVprograms.InProc.Conf.onImageandVideoRetrieval,pages261{269,Dublin,Ireland,2004. [80] C.K.Chang,C.C.Shih,T.Nguyen,andP.Mongkolwat.Apopularity-baseddataallocationschemeforacluster-basedvodserver.InProc.COMPSAC,pages62{67,Seoul,Korea,1996. [81] C.Griwodz,M.Bar,andLCWolf.Long-termmoviepopularitymodelsinvideo-on-demandsystemsorthelifeofanon-demandmovie.InProc.ACMMultimediaConf.,pages349{357,Seattle,USA,1997. [82] T.D.C.LittleandD.Venkatesh:.Popularity-basedassignmentofmoviestostoragedevicesinavideo-on-demandsystem.MultimediaSystems,2(6):280{287,1995.

PAGE 125

EunsamKimwasborninBusan,SouthKorea.HereceivedhisB.S.andM.S.degreesincomputerengineeringfromSeoulNationalUniversity,Korea,in1994and1996,respectively.HeworkedfortheDigitalTVResearchLabofLGElectronicsinKoreafrom1996to2002.HeisexpectedtoreceiveaPh.D.degreeintheDepartmentofComputerandInformationScienceandEngineering,UniversityofFlorida.Hisresearchinterestsincludedistributedmultimediasystems,networkedstoragesystems,andmultimediaconsumernetworkingandplatforms.Hepublishedfourjournalpapersandveconferencepapersintheaboveresearchareas. 113