google云计算系列课程第一讲介绍.ppt

google云计算系列课程第一讲介绍.ppt

《google云计算系列课程第一讲介绍.ppt》由会员分享，可在线阅读，更多相关《google云计算系列课程第一讲介绍.ppt（37页珍藏版）》请在冰豆网上搜索。

DistributedComputingSeminar,Lecture1:

IntroductiontoDistributedComputing&SystemsBackground,ChristopheBisciglia,AaronKimball,&SierraMichels-SlettvetSummer2007,Exceptwhereotherwisenoted,thecontentsofthispresentationareCopyright2007UniversityofWashingtonandarelicensedundertheCreativeCommonsAttribution2.5License.,CourseOverview,5lectures1Introduction2TechnicalSide:

MapReduce&GFS2Theoretical:

AlgorithmsfordistributedcomputingReadings+QuestionsnightlyReadings:

http:

/http:

/,Outline,IntroductiontoDistributedComputingParallelvs.DistributedComputingHistoryofDistributedComputingParallelizationandSynchronizationNetworkingBasics,ComputerSpeedup,MooresLaw:

“Thedensityoftransistorsonachipdoublesevery18months,forthesamecost”（1965）,Image:

TomsHardwareandnotsubjecttotheCreativeCommonslicenseapplicabletotherestofthiswork.,Image:

TomsHardware,Scopeofproblems,Whatcanyoudowith1computer?

Whatcanyoudowith100computers?

Whatcanyoudowithanentiredatacenter?

Distributedproblems,Renderingmultipleframesofhigh-qualityanimation,Image:

DreamWorksAnimationandnotsubjecttotheCreativeCommonslicenseapplicabletotherestofthiswork.,Distributedproblems,Simulatingseveralhundredorthousandcharacters,HappyFeetKingdomFeatureProductions;LordoftheRingsNewLineCinema,neitherimageissubjecttotheCreativeCommonslicenseapplicabletotherestofthework.,Distributedproblems,Indexingtheweb（Google）SimulatinganInternet-sizednetworkfornetworkingexperiments（PlanetLab）Speedingupcontentdelivery（Akamai）,Whatisthekeyattributethatalltheseexampleshaveincommon?

Parallelvs.Distributed,Parallelcomputingcanmean:

VectorprocessingofdataMultipleCPUsinasinglecomputerDistributedcomputingismultipleCPUsacrossmanycomputersoverthenetwork,ABriefHistory1975-85,ParallelcomputingwasfavoredintheearlyyearsPrimarilyvector-basedatfirstGraduallymorethread-basedparallelismwasintroduced,Image:

ComputerPicturesDatabaseandCrayResearchCorpandisnotsubjecttotheCreativeCommonslicenseapplicabletotherestofthiswork.,“Massivelyparallelarchitectures”startrisinginprominenceMessagePassingInterface（MPI）andotherlibrariesdevelopedBandwidthwasabigproblem,ABriefHistory1985-95,ABriefHistory1995-Today,Cluster/gridarchitectureincreasinglydominantSpecialnodemachineseschewedinfavorofCOTStechnologiesWeb-wideclustersoftwareCompanieslikeGoogletakethistotheextreme,Parallelization&Synchronization,ParallelizationIdea,Parallelizationis“easy”ifprocessingcanbecleanlysplitintonunits:

ParallelizationIdea

（2）,Inaparallelcomputation,wewouldliketohaveasmanythreadsaswehaveprocessors.e.g.,afour-processorcomputerwouldbeabletorunfourthreadsatthesametime.,ParallelizationIdea（3）,ParallelizationIdea（4）,ParallelizationPitfalls,Butthismodelistoosimple!

Howdoweassignworkunitstoworkerthreads?

Whatifwehavemoreworkunitsthanthreads?

Howdoweaggregatetheresultsattheend?

Howdoweknowalltheworkershavefinished?

Whatiftheworkcannotbedividedintocompletelyseparatetasks?

Whatisthecommonthemeofalloftheseproblems?

ParallelizationPitfalls

（2）,Eachoftheseproblemsrepresentsapointatwhichmultiplethreadsmustcommunicatewithoneanother,oraccessasharedresource.Goldenrule:

Anymemorythatcanbeusedbymultiplethreadsmusthaveanassociatedsynchronizationsystem!

WhatisWrongWithThis?

Thread1:

voidfoo（）x+;y=x;,Thread2:

voidbar（）y+;x+=3;,Iftheinitialstateisy=0,x=6,whathappensafterthesethreadsfinishrunning?

Multithreaded=Unpredictability,Whenwerunamultithreadedprogram,wedontknowwhatorderthreadsrunin,nordoweknowwhentheywillinterruptoneanother.,Thread1:

voidfoo（）eax=memx;inceax;memx=eax;ebx=memx;memy=ebx;,Thread2:

voidbar（）eax=memy;inceax;memy=eax;eax=memx;addeax,3;memx=eax;,Manythingsthatlooklike“onestep”operationsactuallytakeseveralstepsunderthehood:

Multithreaded=Unpredictability,Thisappliestomorethanjustintegers:

PullingworkunitsfromaqueueReportingworkbacktomasterunitTellinganotherthreadthatitcanbeginthe“nextphase”ofprocessingAllrequiresynchronization!

SynchronizationPrimitives,Asynchronizationprimitiveisaspecialsharedvariablethatguaranteesthatitcanonlybeaccessedatomically.Hardwaresupportguaranteesthatoperationsonsynchronizationprimitivesonlyevertakeonestep,Semaphores,AsemaphoreisaflagthatcanberaisedorloweredinonestepSemaphoreswereflagsthatrailroadengineerswouldusewhenenteringasharedtrack,Onlyonesideofthesemaphorecaneverbered!

（Canbothbegreen?

）,Semaphores,set（）andreset（）canbethoughtofaslock（）andunlock（）Callstolock（）whenthesemaphoreisalreadylockedcausethethreadtoblock.Pitfalls:

Must“bind”semaphorestoparticularobjects;mustrememberto