分布式系統(tǒng)中的通信課件

分布式系統(tǒng)中的通信進(jìn)程間通信是一切分布式系統(tǒng)的核心。分布式系統(tǒng)中的通信都是基于底層網(wǎng)絡(luò)提供的低層消息傳遞機(jī)制的。通過(guò)消息傳遞來(lái)描述通信過(guò)程比使用基于共享存儲(chǔ)器的原語(yǔ)來(lái)描述要更困難。分布式系統(tǒng)中的通信進(jìn)程間通信是一切分布式系統(tǒng)的核心。14種廣泛使用的通信模型遠(yuǎn)程過(guò)程調(diào)用（remoteprocedurecall,RPC）遠(yuǎn)程方法調(diào)用（remotemethodinvocation,RMI）面向消息的中間件（message-orientedmiddleware,MOM）面向流的通信（stream-orientedcommunication）4種廣泛使用的通信模型2CommunicationProtocols

由于沒(méi)有共享存儲(chǔ)器，分布式系統(tǒng)中的所有通信都是基于（低層）消息交換的。Protocolsareagreements/rulesoncommunicationProtocolscouldbeconnection-orientedorconnectionlessApplicationPresentationSessionTransportNetworkDatalinkPhysicalApplicationprotocolPresentationprotocolSessionprotocolTransportprotocolNetworkprotocolDatalinkprotocolPhysicalprotocolNetwork7654321CommunicationProtocols由于沒(méi)3LayeredProtocolsAtypicalmessageasitappearsonthenetwork.2-2LayeredProtocolsAtypicalmes4DataLinkLayerDiscussionbetweenareceiverandasenderinthedatalinklayer.2-3DataLinkLayerDiscussionbetw5Client-ServerInteractionusingTCPNormaloperationofTCP(常規(guī)TCP的許多開(kāi)銷(xiāo)都耗費(fèi)在連接的管理上).TransactionalTCP(更經(jīng)濟(jì)的方式).2-4服務(wù)器執(zhí)行請(qǐng)求的操作告訴服務(wù)器關(guān)閉該連接客戶啟動(dòng)連接建立過(guò)程1,2,3要求客戶釋放連接將發(fā)送請(qǐng)求當(dāng)作連接的建立標(biāo)志(一個(gè)消息包含三條信息)把送回結(jié)果當(dāng)作連接的關(guān)閉標(biāo)志確認(rèn)連接的終止確認(rèn)已收到客戶的請(qǐng)求Client-ServerInteractionusin6MiddlewareProtocolsMiddleware:-Anapplicationthatlogicallylivesintheapplicationlayer-Containsmanygeneral-purposeprotocolsthatwarrant（代表）

theirownlayers-(會(huì)話層和表示層由一個(gè)單一的中間件層代替)ApplicationMiddlewareTransportNetworkDatalinkPhysicalApplicationprotocolMiddlewareprotocolTransportprotocolNetworkprotocolDatalinkprotocolPhysicalprotocolNetwork654321MiddlewareProtocolsMiddleware7Client-ServerCommunicationModelStructure:groupofserversofferingservicetoclients-Servers:offerservicestotheuserscalled“clients”-Clients:applicationsrequiringservicesfromservers-Example:WebServer/clients,Fileserver…Whyuseclient-servermodel-simplicity-low(er)overheadsclientkernelfileserverkernelprocessserverkernelterminalserverkernelClient-ServerCommunicationMo8Client-ServerCommunicationModelBasedonarequest/responseparadigm-Clientssendarequestaskingforservice(e.g.,afileblock)-Serverprocessesandreplieswithresult(orerror)Techniques:-Socket,RemoteProcedureCalls(RPC),RemoteMethodInvocation(RMI)

clientkernelfileserverkernelprocessserverkernelterminalserverkernelClient-ServerCommunicationMo9RemoteProcedureCall(RPC)Client-ServerprovidesamechanismforservicesindistributedsystemsBUT-requiresexplicitcommunication(send-receive)Q:Howdowemake“distributedcomputinglookliketraditional(centralized)computing”?Canweuseprocedurecalls?AcallsB->Asuspended,Bexecutes->Breturns,Aexecutes

?InformationfromA(caller)toB(callee)transferredusingparameters

?SomewhateasiersincebothcallerandcalleeexecuteinthesameaddressspaceRemoteProcedureCall(RPC)Cli10ConventionalProcedureCallParameterpassinginalocalprocedurecall:thestackbeforethecalltoread(fd,buf,nbytes)Thestackwhilethecalledprocedureisactive(執(zhí)行時(shí))調(diào)用方(主程序)把參數(shù)、返回地址反序壓入堆棧ConventionalProcedureCallPar11RPCInDistributedsystems:thecalleemaybeonadifferentsystem

-RemoteProcedurecall(RPC,允許程序調(diào)用位于其它機(jī)器上的進(jìn)程)-NOEXPLICITMESSAGEPASSING（對(duì)編程人員）

從而達(dá)到隱藏通信（send和receive），實(shí)現(xiàn)分布式系統(tǒng)中的訪問(wèn)透明性Goal:MakeRPClook(asmuchaspossible)likelocalprocedurecall

-allowremoteservicestobecalledasprocedures-callershouldnotbeawareofthefactthatthecalleeis(executing)onadifferentmachine(orviceversa)Althoughnomessagepassing(atuserlevel),parametersmuststillbepassed&resultsmuststillbereturned!RPCInDistributedsystems:the12ClientandServerStubsPrincipleofRPCbetweenaclientandserverprogram.

客戶通過(guò)執(zhí)行普通的（本地）過(guò)程調(diào)用來(lái)訪問(wèn)遠(yuǎn)程服務(wù)，它并不需要直接調(diào)用Send和Receive，消息傳遞的所有細(xì)節(jié)都隱藏在雙方的庫(kù)過(guò)程（Stub,Skeleton）中。ClientandServerStubs13StepsofaremoteprocedurecallClientprocedurecallsclientstubinnormalwayClientstubbuildsmessage,callslocalOSClient’sOSsendsmessagetoremoteOSRemoteOSgivesmessagetoserverstubServerstubunpacksparameters,callsserverServerdoeswork,returnsresulttothestubServerstubpacksitinmessage,callslocalOSServer’sOSsendsmessagetoclient’sOSClient’sOSgivesmessagetoclientstubStubunpacksresult,returnstoclientStepsofaremoteprocedureca14PassingValueParameters(1)StepsinvolvedindoingremotecomputationthroughRPC2-8PassingValueParameters(1)2-15PassingValueParameters(2)Problem:differentmachineshavedifferentdataformats-Intel:littleendian,SPARC:bigendianSolution:useastandardrepresentation-Example:externaldatarepresentation(XDR)PassingValueParameters(2)Pr16Marshalling:ReferenceParametersProblem:howdowepasspointers?

（指針只在它被使用的進(jìn)程的地址空間里面是有意義的）-Ifitpointstoawell-defineddatastructure,passacopytotheserverandtheserverstubpasses（使用）

apointertothelocalcopy

(來(lái)調(diào)用服務(wù)器程序）Whataboutdatastructurescontainingpointers?-Prohibit-Copy/restore,

machineindependentrepresentationCallbyCopy/restore:

首先由調(diào)用者將變量拷貝到Stack，調(diào)用完成之后，將堆棧中的變量復(fù)制回去覆蓋調(diào)用方該變量原先的值。參數(shù)從客戶發(fā)送到服務(wù)器，在那里修改，然后發(fā)送回客戶，覆蓋原來(lái)的值。

Marshalling（編組，即把參數(shù)打包進(jìn)消息中發(fā)送到服務(wù)器）:transformparameters/resultsintoabytestreamMarshalling:ReferenceParamet17BindingProblem:howdoesaclientlocateaserver?

（Client要發(fā)送消息到服務(wù)器，它需要知道服務(wù)器的地址）-UseBindingsServer-Exportserverinterfaceduringinitialization-Sendname,version#,uniqueidentifier,handle(address)tobinderClient(第一次調(diào)用遠(yuǎn)程過(guò)程，還沒(méi)有綁定到一個(gè)Server）-FirstRPC:sendmsgtobindertoimportserverinterface-Binder:checktoseeifsomeserverhasexportedtheinterface

?ReturnhandleanduniqueidentifiertoclientBinder充當(dāng)一個(gè)注冊(cè)服務(wù)器的角色。BindingProblem:howdoesacli18Binding:CommentsMethodflexible-canhandlemultipleserverswithsameinterface-bindercanpollaservertoseeifitisupandmayderegisteritifitisdownforfaulttolerance-canenforceauthentication(notgivinginterfacetousernotonlist)ExportingandimportingincursoverheadsBindercanbeabottleneck-UsemultiplebindersBindercandoloadbalancing

隨機(jī)地?cái)U(kuò)展（散播）Clients到這些Servers來(lái)平衡負(fù)載。Binding:CommentsMethodflexib19AsynchronousRPC(1)TheinterconnectionbetweenclientandserverinatraditionalRPCTheinteractionusingasynchronousRPC2-12AsynchronousRPC(1)Theinterc20AsynchronousRPC(2)AclientandserverinteractingthroughtwoasynchronousRPCs2-13AsynchronousRPC(2)2-1321WritingaClientandaServerThestepsinwritingaclientandaserverinDCERPC.2-14WritingaClientandaServerT22BindingaClienttoaServerClient-to-serverbindinginDCE.2-15BindingaClienttoaServer2-23RemoteMethodInvocation(RMI)RPCsappliedto(distributed)objects,i.e.,instancesofaclass-Class:object-orientedabstraction;modulewithdataandoperations-Separationbetweeninterfaceandimplementation-Interfaceresidesononemachine,implementationonanotherRMIssupportsystem-wideobjectreferences-ParameterscanbeobjectreferencesRemoteMethodInvocation(RMI)24DistributedObjectsWhenaclientbindstoadistributedobject,loadtheinterface(“proxy”)intoclientaddressspace-ProxyanalogoustostubsServerstubisreferredtoasaskeletonClientOSClientinvokesamethodClientServerOSSkeletonServerStateMethodInterfaceProxySkeletoninvokessamemethodatobjectSameinterfaceasobjectClientmachineServermachineObjectNetworkMarshalledinvocationispassedacrossnetworkDistributedObjectsClientOSCl25ProxiesandSkeletonsProxy:clientstub-MaintainsserverID,endpoint,objectID-Setsupandtearsdownconnectionwiththeserver-Doesserializationoflocalobjectparameters-Inpractice,canbedownloaded/constructedontheflySkeleton:serverstub-DoesdeserializationandpassesparameterstoserverandsendsresulttoproxyProxiesandSkeletonsProxy:cl26BindingaClienttoanObjectAnexamplewithimplicitbindingusingonlyglobalreferencesAnexamplewithexplicitbindingusingglobalandlocalreferencesDistr_object*obj_ref; //Declareasystemwideobjectreference

obj_ref=…; //Initializethereferencetoadistributedobject

obj_ref->do_something(); //Implicitlybindandinvokeamethod (a)Distr_objectobjPref; //Declareasystemwideobjectreference

Local_object*obj_ptr; //Declareapointertolocalobjects

obj_ref=…; //Initializethereferencetoadistributedobject

obj_ptr=bind(obj_ref); //Explicitlybindandobtainapointertothelocalproxy

obj_ptr->do_something(); //Invokeamethodonthelocalproxy (b)BindingaClienttoanObjectA27ParameterPassingThesituationwhenpassinganobjectbyreferenceorbyvalue.2-18ParameterPassingThesituation28TheDCEDistributed-ObjectModelDistributeddynamicobjectsinDCE.Distributednamedobjects2-19TheDCEDistributed-ObjectMod29面向消息的通信遠(yuǎn)程過(guò)程調(diào)用和遠(yuǎn)程對(duì)象調(diào)用都有助于隱藏分布式系統(tǒng)中的通信，也就是說(shuō)增強(qiáng)了訪問(wèn)透明性。不幸的是，這兩種機(jī)制并不總是適用的。特別是當(dāng)無(wú)法保證發(fā)出請(qǐng)求時(shí)接收端一定正在執(zhí)行的情況下，就必須有其他的通信服務(wù)。同時(shí)，RPC和RMI的同步特性也會(huì)造成客戶在發(fā)出的請(qǐng)求得到處理之前被阻塞，因而有時(shí)也需要采取其他辦法。這里所說(shuō)的“其他方法”就是消息傳遞機(jī)制。面向消息的通信遠(yuǎn)程過(guò)程調(diào)用和遠(yuǎn)程對(duì)象調(diào)用都有助于隱藏分布式系30Message-orientedcommunication-Persistenceandsynchronicity-Message-orientedtransientcommunication

?Berkeleysocket

?MPI-Message-orientedpersistentcommunicationmessage?MessagequeuingsystemsMessage-orientedcommunication31持久通信與暫時(shí)通信持久(persistent)通信需要傳輸?shù)南⒃谔峤恢笥赏ㄐ畔到y(tǒng)來(lái)存儲(chǔ)，直到將其交付給接收者為止。即在將消息成功交付給下一個(gè)通信服務(wù)器之前，消息一直存儲(chǔ)在通信服務(wù)器中。其典型例子是電子郵件系統(tǒng)。暫時(shí)(transient)通信通信系統(tǒng)只在發(fā)送和接收消息的應(yīng)用程序的運(yùn)行期間存儲(chǔ)消息。更準(zhǔn)確地說(shuō)，如果通信服務(wù)器無(wú)法將消息遞送到下一個(gè)服務(wù)器或者接收者，消息將會(huì)被簡(jiǎn)單地丟棄。持久通信與暫時(shí)通信持久(persistent)通信32異步通信與同步通信異步通信發(fā)送者把要傳輸?shù)南⑻峤恢罅⒓蠢^續(xù)執(zhí)行其其他程序，這意味著該消息存儲(chǔ)在位于發(fā)送端主機(jī)的本地緩沖區(qū)中，或者存儲(chǔ)在送達(dá)的第一個(gè)通信服務(wù)器上的緩沖區(qū)中。同步通信發(fā)送者在提交消息之后會(huì)被阻塞，直到消息已經(jīng)到達(dá)并存儲(chǔ)在接收主機(jī)的本地緩沖區(qū)中以后，也就是消息確實(shí)已經(jīng)傳送到接收者之后,才會(huì)繼續(xù)執(zhí)行其它程序。異步通信與同步通信異步通信33PersistenceandSynchronicity:CommentsTransientsynchronouscomm:response-based=>weakerforms,suchasdelivery-basedandreply-basedTransientasynchronouscomm:message-passingsystemsPersistentcomm:developingofmiddlewareforlarge-scaleinterconnectednetworks;failuremaskingandrecoveryPersistenceandSynchronicity:34Message-orientedTransientComm.Manydistributedsystemsbuiltontopofsimplemessage-orientedmodel-Example:Berkeleysockets-AsocketisanabstractrepresentationofacommunicationendpointPrimitiveMeaningSocketCreateanewcommunicationendpointBindAttachalocaladdresstoasocketListenAnnouncewillingnesstoacceptconnectionsAcceptBlockcalleruntilaconnectionrequestarrivesConnectActivelyattempttoestablishaconnectionSendSendsomedataovertheconnectionReceiveReceivesomedataovertheconnectionCloseReleasetheconnectionSocketprimitivesforTCP/IP.Message-orientedTransientCom35BerkeleySocketsConnection-orientedcommunicationpatternusingsockets.BerkeleySocketsConnection-or36TheMessage-Passing

Interface(MPI)Someofthemostintuitivemessage-passingprimitivesofMPI.PrimitiveMeaningMPI_bsendAppendoutgoingmessagetoalocalsendbufferMPI_sendSendamessageandwaituntilcopiedtolocalorremotebufferMPI_ssendSendamessageandwaituntilreceiptstartsMPI_sendrecvSendamessageandwaitforreplyMPI_isendPassreferencetooutgoingmessage,andcontinueMPI_issendPassreferencetooutgoingmessage,andwaituntilreceiptstartsMPI_recvReceiveamessage;blockiftherearenoneMPI_irecvCheckifthereisanincomingmessage,butdonotblockTheMessage-Passing

Interface37Message-orientedPersistentComm.MessagequeuingsystemsorMessage-OrientedMiddleware(MOM)-Supportasynchronouspersistentcommunication-Intermediatestorageformessagewhilesender/receiverareinactive-Exampleapplication:email-支持那些時(shí)間要求較為寬松的消息傳輸，比如那些要求幾分鐘完成的傳輸，而適用于那些必須在幾秒內(nèi)甚至幾微秒內(nèi)完成的傳輸。(不同于Socket及MPI)Message-orientedPersistentCo38Message-QueuingModel(1)Communicatebyinsertingmessagesinqueues消息由一系列通信服務(wù)器依次進(jìn)行轉(zhuǎn)發(fā)，最終送達(dá)目的地。即使在消息發(fā)送過(guò)程中接受者的機(jī)器未處于運(yùn)行狀態(tài)，消息也能送到。原則上，每一個(gè)應(yīng)用程序都擁有歸其私有的消息隊(duì)列，其他應(yīng)用程序可以發(fā)送消息到該隊(duì)列中。隊(duì)列只能由相應(yīng)的（也就是擁有該隊(duì)列的應(yīng)用程序）讀取，但是也可能有多個(gè)應(yīng)用程序共享單個(gè)隊(duì)列。Senderisonlyguaranteedthatmessagewillbeeventuallyinsertedinrecipient'squeue-NoguaranteesonwhenarrivingorifthemessagewillbereadMessage-QueuingModel(1)Commu39Message-QueuingModel(2)Fourcombinationsforloosely-coupledcommunicationsusingqueues.2-26Looselycoupledcommunication

消息進(jìn)入接收者的消息隊(duì)列時(shí)，接收者不必處于運(yùn)行狀態(tài)；即使發(fā)送者沒(méi)有運(yùn)行，接收者也可以讀出發(fā)送給它的消息。發(fā)送者和接收者可以彼此完全獨(dú)立地運(yùn)行。Message-QueuingModel(2)Four40Message-QueuingModel(3)Basicinterfacetoaqueueinamessage-queuingsystem.PrimitiveMeaningPutAppendamessagetoaspecifiedqueueGetBlockuntilthespecifiedqueueisnonempty,andremovethefirstmessagePollCheckaspecifiedqueueformessages,andremovethefirst.Neverblock.NotifyInstallahandlertobecalledwhenamessageisputintothespecifiedqueue.Message-QueuingModel(3)Basic41Example:CommunicationSystem主機(jī)通過(guò)網(wǎng)絡(luò)相連的通信系統(tǒng)的一般組織結(jié)構(gòu)Example:CommunicationSystem主42消息隊(duì)列系統(tǒng)消息由（發(fā)送者的）源隊(duì)列向（接收者的）目的隊(duì)列傳輸。全部隊(duì)列的集合是分布在多臺(tái)機(jī)器上的，因此，對(duì)于要傳輸消息的消息隊(duì)列系統(tǒng)來(lái)說(shuō)，他應(yīng)該維護(hù)一個(gè)由隊(duì)列到其所在網(wǎng)絡(luò)位置之間的映射關(guān)系，即維護(hù)一個(gè)存儲(chǔ)網(wǎng)絡(luò)位置（IP地址）所對(duì)應(yīng)的隊(duì)列名的數(shù)據(jù)庫(kù)。消息隊(duì)列系統(tǒng)消息由（發(fā)送者的）源隊(duì)列向（接收者的）目的隊(duì)列傳43GeneralArchitectureofaMessage-QueuingSystem(1)Therelationshipbetweenqueue-leveladdressingandnetwork-leveladdressing.GeneralArchitectureofaMess44隊(duì)列由隊(duì)列管理器來(lái)管理。一般來(lái)說(shuō)，隊(duì)列管理器與發(fā)送或者接收消息的應(yīng)用程序直接交互。然而，也有一些特殊的隊(duì)列管理器發(fā)揮了路由器或者中繼器的作用：他們將輸入的消息轉(zhuǎn)發(fā)給其他的隊(duì)列管理器。一種解決方案是，使用若干了解網(wǎng)絡(luò)拓?fù)涞穆酚善?。如果發(fā)送者A將目的地為B的消息放入其本地隊(duì)列中，該消息將會(huì)首先傳輸?shù)阶钹徑穆酚善?，該路由器知道如何將該消息往B的方向轉(zhuǎn)發(fā)。GeneralArchitectureofaMessage-QueuingSystem(2)隊(duì)列由隊(duì)列管理器來(lái)管理。一般來(lái)說(shuō)，隊(duì)列管理器與發(fā)送或者接收消45GeneralArchitectureofaMessage-QueuingSystem(3)Thegeneralorganizationofamessage-queuingsystemwithrouters.2-29GeneralArchitectureofaMess46Stream-orientedcommunication隨著多媒體分布式系統(tǒng)的出現(xiàn)，必須引入流（stream）的概念，以支持連續(xù)媒體的通信。-Datastream-Qualityofservices-StreamsynchronizationStream-orientedcommunication隨47DataStreamAdatastreamisasequenceofdataunitsDiscreteorcontinuous:-Discretestream:UNIXpipesorTCP/IPconnections-Continuousstream:audioorvideo（同步非常關(guān)鍵）Forcontinuousstream,threetransmissionmodes:-Asynchronoustransmissionmode:thedataitemsaretransmittedoneaftertheother,buttherearenofurthertimingconstraintsonwhentransmissionshouldtakeplace-Synchronoustransmissionmode:thereisamaximumend-to-enddelaydefinedforeachunitinadatastream-Isochronous（等時(shí)）

transmissionmode:dataunitsaretransferredontime,haveamaximumandminimumend-to-enddelayDataStreamAdatastreamisa48

（a）通過(guò)網(wǎng)絡(luò)在兩個(gè)進(jìn)程間建立流連接（b）在兩個(gè)設(shè)備間直接建立流連接分布式系統(tǒng)中的通信課件49ReadingMaterials

“DistributedSystems:PrinciplesandParadigms”byTanenbaumandVanSteen,PrenticeHall2002“DistributedSystems:ConceptsandDesign”byGeorgeCoulourisetc.,Addison-Wesley2001(Thirdedition)分布式系統(tǒng)設(shè)計(jì)，(美)JieWu著，高傳善等譯，機(jī)械工業(yè)出版社，2001分布式操作系統(tǒng)，AndrewS.Tanenbaum著，電子工業(yè)出版社，1999ReadingMaterials50分布式系統(tǒng)中的通信進(jìn)程間通信是一切分布式系統(tǒng)的核心。分布式系統(tǒng)中的通信都是基于底層網(wǎng)絡(luò)提供的低層消息傳遞機(jī)制的。通過(guò)消息傳遞來(lái)描述通信過(guò)程比使用基于共享存儲(chǔ)器的原語(yǔ)來(lái)描述要更困難。分布式系統(tǒng)中的通信進(jìn)程間通信是一切分布式系統(tǒng)的核心。514種廣泛使用的通信模型遠(yuǎn)程過(guò)程調(diào)用（remoteprocedurecall,RPC）遠(yuǎn)程方法調(diào)用（remotemethodinvocation,RMI）面向消息的中間件（message-orientedmiddleware,MOM）面向流的通信（stream-orientedcommunication）4種廣泛使用的通信模型52CommunicationProtocols

由于沒(méi)有共享存儲(chǔ)器，分布式系統(tǒng)中的所有通信都是基于（低層）消息交換的。Protocolsareagreements/rulesoncommunicationProtocolscouldbeconnection-orientedorconnectionlessApplicationPresentationSessionTransportNetworkDatalinkPhysicalApplicationprotocolPresentationprotocolSessionprotocolTransportprotocolNetworkprotocolDatalinkprotocolPhysicalprotocolNetwork7654321CommunicationProtocols由于沒(méi)53LayeredProtocolsAtypicalmessageasitappearsonthenetwork.2-2LayeredProtocolsAtypicalmes54DataLinkLayerDiscussionbetweenareceiverandasenderinthedatalinklayer.2-3DataLinkLayerDiscussionbetw55Client-ServerInteractionusingTCPNormaloperationofTCP(常規(guī)TCP的許多開(kāi)銷(xiāo)都耗費(fèi)在連接的管理上).TransactionalTCP(更經(jīng)濟(jì)的方式).2-4服務(wù)器執(zhí)行請(qǐng)求的操作告訴服務(wù)器關(guān)閉該連接客戶啟動(dòng)連接建立過(guò)程1,2,3要求客戶釋放連接將發(fā)送請(qǐng)求當(dāng)作連接的建立標(biāo)志(一個(gè)消息包含三條信息)把送回結(jié)果當(dāng)作連接的關(guān)閉標(biāo)志確認(rèn)連接的終止確認(rèn)已收到客戶的請(qǐng)求Client-ServerInteractionusin56MiddlewareProtocolsMiddleware:-Anapplicationthatlogicallylivesintheapplicationlayer-Containsmanygeneral-purposeprotocolsthatwarrant（代表）

theirownlayers-(會(huì)話層和表示層由一個(gè)單一的中間件層代替)ApplicationMiddlewareTransportNetworkDatalinkPhysicalApplicationprotocolMiddlewareprotocolTransportprotocolNetworkprotocolDatalinkprotocolPhysicalprotocolNetwork654321MiddlewareProtocolsMiddleware57Client-ServerCommunicationModelStructure:groupofserversofferingservicetoclients-Servers:offerservicestotheuserscalled“clients”-Clients:applicationsrequiringservicesfromservers-Example:WebServer/clients,Fileserver…Whyuseclient-servermodel-simplicity-low(er)overheadsclientkernelfileserverkernelprocessserverkernelterminalserverkernelClient-ServerCommunicationMo58Client-ServerCommunicationModelBasedonarequest/responseparadigm-Clientssendarequestaskingforservice(e.g.,afileblock)-Serverprocessesandreplieswithresult(orerror)Techniques:-Socket,RemoteProcedureCalls(RPC),RemoteMethodInvocation(RMI)

clientkernelfileserverkernelprocessserverkernelterminalserverkernelClient-ServerCommunicationMo59RemoteProcedureCall(RPC)Client-ServerprovidesamechanismforservicesindistributedsystemsBUT-requiresexplicitcommunication(send-receive)Q:Howdowemake“distributedcomputinglookliketraditional(centralized)computing”?Canweuseprocedurecalls?AcallsB->Asuspended,Bexecutes->Breturns,Aexecutes

?InformationfromA(caller)toB(callee)transferredusingparameters

?SomewhateasiersincebothcallerandcalleeexecuteinthesameaddressspaceRemoteProcedureCall(RPC)Cli60ConventionalProcedureCallParameterpassinginalocalprocedurecall:thestackbeforethecalltoread(fd,buf,nbytes)Thestackwhilethecalledprocedureisactive(執(zhí)行時(shí))調(diào)用方(主程序)把參數(shù)、返回地址反序壓入堆棧ConventionalProcedureCallPar61RPCInDistributedsystems:thecalleemaybeonadifferentsystem

-RemoteProcedurecall(RPC,允許程序調(diào)用位于其它機(jī)器上的進(jìn)程)-NOEXPLICITMESSAGEPASSING（對(duì)編程人員）

從而達(dá)到隱藏通信（send和receive），實(shí)現(xiàn)分布式系統(tǒng)中的訪問(wèn)透明性Goal:MakeRPClook(asmuchaspossible)likelocalprocedurecall

-allowremoteservicestobecalledasprocedures-callershouldnotbeawareofthefactthatthecalleeis(executing)onadifferentmachine(orviceversa)Althoughnomessagepassing(atuserlevel),parametersmuststillbepassed&resultsmuststillbereturned!RPCInDistributedsystems:the62ClientandServerStubsPrincipleofRPCbetweenaclientandserverprogram.

客戶通過(guò)執(zhí)行普通的（本地）過(guò)程調(diào)用來(lái)訪問(wèn)遠(yuǎn)程服務(wù)，它并不需要直接調(diào)用Send和Receive，消息傳遞的所有細(xì)節(jié)都隱藏在雙方的庫(kù)過(guò)程（Stub,Skeleton）中。ClientandServerStubs63StepsofaremoteprocedurecallClientprocedurecallsclientstubinnormalwayClientstubbuildsmessage,callslocalOSClient’sOSsendsmessagetoremoteOSRemoteOSgivesmessagetoserverstubServerstubunpacksparameters,callsserverServerdoeswork,returnsresulttothestubServerstubpacksitinmessage,callslocalOSServer’sOSsendsmessagetoclient’sOSClient’sOSgivesmessagetoclientstubStubunpacksresult,returnstoclientStepsofaremoteprocedureca64PassingValueParameters(1)StepsinvolvedindoingremotecomputationthroughRPC2-8PassingValueParameters(1)2-65PassingValueParameters(2)Problem:differentmachineshavedifferentdataformats-Intel:littleendian,SPARC:bigendianSolution:useastandardrepresentation-Example:externaldatarepresentation(XDR)PassingValueParameters(2)Pr66Marshalling:ReferenceParametersProblem:howdowepasspointers?

（指針只在它被使用的進(jìn)程的地址空間里面是有意義的）-Ifitpointstoawell-defineddatastructure,passacopytotheserverandtheserverstubpasses（使用）

apointertothelocalcopy

(來(lái)調(diào)用服務(wù)器程序）Whataboutdatastructurescontainingpointers?-Prohibit-Copy/restore,

machineindependentrepresentationCallbyCopy/restore:

首先由調(diào)用者將變量拷貝到Stack，調(diào)用完成之后，將堆棧中的變量復(fù)制回去覆蓋調(diào)用方該變量原先的值。參數(shù)從客戶發(fā)送到服務(wù)器，在那里修改，然后發(fā)送回客戶，覆蓋原來(lái)的值。

Marshalling（編組，即把參數(shù)打包進(jìn)消息中發(fā)送到服務(wù)器）:transformparameters/resultsintoabytestreamMarshalling:ReferenceParamet67BindingProblem:howdoesaclientlocateaserver?

（Client要發(fā)送消息到服務(wù)器，它需要知道服務(wù)器的地址）-UseBindingsServer-Exportserverinterfaceduringinitialization-Sendname,version#,uniqueidentifier,handle(address)tobinderClient(第一次調(diào)用遠(yuǎn)程過(guò)程，還沒(méi)有綁定到一個(gè)Server）-FirstRPC:sendmsgtobindertoimportserverinterface-Binder:checktoseeifsomeserverhasexportedtheinterface

?ReturnhandleanduniqueidentifiertoclientBinder充當(dāng)一個(gè)注冊(cè)服務(wù)器的角色。BindingProblem:howdoesacli68Binding:CommentsMethodflexible-canhandlemultipleserverswithsameinterface-bindercanpollaservertoseeifitisupandmayderegisteritifitisdownforfaulttolerance-canenforceauthentication(notgivinginterfacetousernotonlist)ExportingandimportingincursoverheadsBindercanbeabottleneck-UsemultiplebindersBindercandoloadbalancing

隨機(jī)地?cái)U(kuò)展（散播）Clients到這些Servers來(lái)平衡負(fù)載。Binding:CommentsMethodflexib69AsynchronousRPC(1)TheinterconnectionbetweenclientandserverinatraditionalRPCTheinteractionusingasynchronousRPC2-12AsynchronousRPC(1)Theinterc70AsynchronousRPC(2)AclientandserverinteractingthroughtwoasynchronousRPCs2-13AsynchronousRPC(2)2-1371WritingaClientandaServerThestepsinwritingaclientandaserverinDCERPC.2-14WritingaClientandaServerT72BindingaClienttoaServerClient-to-serverbindinginDCE.2-15BindingaClienttoaServer2-73RemoteMethodInvocation(RMI)RPCsappliedto(distributed)objects,i.e.,instancesofaclass-Class:object-orientedabstraction;modulewithdataandoperations-Separationbetweeninterfaceandimplementation-Interfaceresidesononemachine,implementationonanotherRMIssupportsystem-wideobjectreferences-ParameterscanbeobjectreferencesRemoteMethodInvocation(RMI)74DistributedObjectsWhenaclientbindstoadistributedobject,loadtheinterface(“proxy”)intoclientaddressspace-ProxyanalogoustostubsServerstubisreferredtoasaskeletonClientOSClientinvokesamethodClientServerOSSkeletonServerStateMethodInterfaceProxySkeletoninvokessamemethodatobjectSameinterfaceasobjectClientmachineServermachineObjectNetworkMarshalledinvocationispassedacrossnetworkDistributedObjectsClientOSCl75ProxiesandSkeletonsProxy:clientstub-MaintainsserverID,endpoint,objectID-Setsupandtearsdownconnectionwiththeserver-Doesserializationoflocalobjectparameters-Inpractice,canbedownloaded/constructedontheflySkeleton:serverstub-DoesdeserializationandpassesparameterstoserverandsendsresulttoproxyProxiesandSkeletonsProxy:cl76BindingaClienttoanObjectAnexamplewithimplicitbindingusingonlyglobalreferencesAnexamplewithexplicitbindingusingglobalandlocalreferencesDistr_object*obj_ref; //Declareasystemwideobjectreference

obj_ref=…; //Initializethereferencetoadistributedobject

obj_ref->do_something(); //Implicitlybindandinvokeamethod (a)Distr_objectobjPref; //Declareasystemwideobjectreference

Local_object*obj_ptr; //Declareapointertolocalobjects

obj_ref=…; //Initializethereferencetoadistributedobject

obj_ptr=bind(obj_ref); //Explicitlybindandobtainapointertothelocalproxy

obj_ptr->do_something(); //Invokeamethodonthelocalproxy (b)BindingaClienttoanObjectA77ParameterPassingThesituationwhenpassinganobjectbyreferenceorbyvalue.2-18ParameterPassingThesituation78TheDCEDistributed-ObjectModelDistributeddynamicobjectsinDCE.Distributednamedobjects2-19TheDCEDistributed-ObjectMod79面向消息的通信遠(yuǎn)程過(guò)程調(diào)用和遠(yuǎn)程對(duì)象調(diào)用都有助于隱藏分布式系統(tǒng)中的通信，也就是說(shuō)增強(qiáng)了訪問(wèn)透明性。不幸的是，這兩種機(jī)制并不總是適用的。特別是當(dāng)無(wú)法保證發(fā)出請(qǐng)求時(shí)接收端一定正在執(zhí)行的情況下，就必須有其他的通信服務(wù)。同時(shí)，RPC和RMI的同步特性也會(huì)造成客戶在發(fā)出的請(qǐng)求得到處理之前被阻塞，因而有時(shí)也需要采取其他辦法。這里所說(shuō)的“其他方法”就是消息傳遞機(jī)制。面向消息的通信遠(yuǎn)程過(guò)程調(diào)用和遠(yuǎn)程對(duì)象調(diào)用都有助于隱藏分布式系80Message-orientedcommunication-Persistenceandsynchronicity-Message-orientedtransientcommunication

?Berkeleysocket

?MPI-Message-orientedpersistentcommunicationmessage?MessagequeuingsystemsMessage-orientedcommunication81持久通信與暫時(shí)通信持久(persistent)通信需要傳輸?shù)南⒃谔峤恢笥赏ㄐ畔到y(tǒng)來(lái)存儲(chǔ)，直到將其交付給接收者為止。即在將消息成功交付給下一個(gè)通信服務(wù)器之前，消息一直存儲(chǔ)在通信服務(wù)器中。其典型例子是電子郵件系統(tǒng)。暫時(shí)(transient)通信通信系統(tǒng)只在發(fā)送和接收消息的應(yīng)用程序的運(yùn)行期間存儲(chǔ)消息。更準(zhǔn)確地說(shuō)，如果通信服務(wù)器無(wú)法將消息遞送到下一個(gè)服務(wù)器或者接收者，消息將會(huì)被簡(jiǎn)單地丟棄。持久通信與暫時(shí)通信持久(persistent)通信82異步通信與同步通信異步通信發(fā)送者把要傳輸?shù)南⑻峤恢罅⒓蠢^續(xù)執(zhí)行其其他程序，這意味著該消息存儲(chǔ)在位于發(fā)送端主機(jī)的本地緩沖區(qū)中，或者存儲(chǔ)在送達(dá)的第一個(gè)通信服務(wù)器上的緩沖區(qū)中。同步通信發(fā)送者在提交消息之后會(huì)被阻塞，直到消息已經(jīng)到達(dá)并存儲(chǔ)在接收主機(jī)的本地緩沖區(qū)中以后，也就是消息確實(shí)已經(jīng)傳送到接收者之后,才會(huì)繼續(xù)執(zhí)行其它程序。異步通信與同步通信異步通信83PersistenceandSynchronicity:CommentsTransientsynchronouscomm:response-based=>weakerforms,suchasdelivery-basedandreply-basedTransientasynchronouscomm:message-passingsystemsPersistentcomm:developingofmiddlewareforlarge-scaleinterconnectednetworks;failuremaskingandrecoveryPersistenceandSynchronicity:84Message-orientedTransientComm.Manydistributedsystemsbuiltontopofsimplemessage-orientedmodel-Example:Berkeleysockets-AsocketisanabstractrepresentationofacommunicationendpointPrimitiveMeaningSocketCreateanewcommunicationendpointBindAttachalocaladdresstoasocketListenAnnouncewillingnesstoacceptconnectionsAcceptBlockcalleruntilaconnectionrequestarrivesConnectAc