高级操作系统 第7章 分布式系统.ppt

LearningachosenvalueTherearesomeoptions:Eachacceptor,wheneveritacceptsaproposal,informsallthelearners.Acceptorsinformsadistinguishedlearner(usuallytheproposer)andletthedistinguishedlearnerbroadcasttheresult.ClassicPaxos情形1:PaxoswithoutProposerCompetitionPreparesetabc=0abc=X(1)Response(1)PromiseResponseClassicPaxosAcceptabc=XAccept(1,setabc=0)Promise(1,setabc=0)Response(1)PromiseClassicPaxosAcceptabc=0Acceptedsetabc=0accpetedAccept(1,setabc=0)PromiseClassicPaxos情形2:PaxoswithProposerCompetitionPrepareabc=Xsetabc=0setabc=1(1)(2)ClassicPaxosabc=XResponse(2)PromiseResponse(1)PromiseResponseResponsePrepareClassicPaxosPrepare/Acceptabc=XAccept(1,setabc=0)Promise(1,setabc=0)ResponseResponse(2)PromiseResponse(1,setabc=0)Response(1)PromiseClassicPaxosAcceptabc=XAccept(1,setabc=0)Promise(1,setabc=0)Accept(2,setabc=0)Promise(2,setabc=0)abc=0Response(2)PromiseTunableknobs(可调旋钮)Acceptorshavemanyoptionstoresponse:Preparerequest:No/YesAcceptrequest:No/Yesifitdidn’tpromisenottodosoBackofftimeafterabandonaproposal:exponentialback-off/pre-assignedvaluesShouldwewaitfornodestoonlineineachphase?ApplicationsChubbylockservice.Petal:Distributedvirtualdisks.Frangipani:Ascalabledistributedfilesystem.Raftalgorithm：InSearchofanUnderstandableConsensusAlgorithm(ExtendedVersion)，DiegoOngaroandJohnOusterhout，StanfordUniversity，May20,2014.Raftisaconsensusalgorithmformanagingareplicatedlog.Itproducesaresultequivalentto(multi-)Paxos,anditisasefficientasPaxos,butitsstructureisdifferentfromPaxos；Raft中有三类角色：LeaderFollowerCandidate常用算法：Bully算法和环算法。RaftalgorithmRaft协议强依赖Leader节点的可用性来确保集群数据的一致性。类似于事务的两阶段提交协议完成数据的一致性。Raftalgorithm接下来只要解决协议中的故障问题：1.数据到达Leader节点前；2.数据到达Leader节点，但未复制到Fol