高级操作系统全chapter4支持2016ados

AdvancedOperatingDistributedSystemsConceptsanddesignQingZhu,DepartmentofComputerRenminChapterOperatingSystem Teacher:QingInformationSchool,2DistributedSystems–ConceptsandDesign,GeorgeCouloris,andTim

nEdition3,4,©PearsonEducation2001,2005ThelectureisbasedonthisChapterChapter44.74.84 5Virtualmachine NetworkoperatingsystemBothUNIXandWindowsareexamplesofnetworkoperatingsystems.Thedefiningcharacteristicisthatthenodesrunninganetworkoperatingsystemretainautonomy(自治)inmanagingtheirownprocessingresources.Inotherwords,therearemultiplesystemimages,onepernode.7DistributedoperatingsystemOnecouldenvisage(正视)anoperatingsysteminwhichusersareneverconcernedwithwheretheirprogramsrunorthelocationofanyresources.ThereisasinglesystemAnoperatingsystemthatproducesasinglesystemimagelikethisforalltheresourcesinadistributedsystemiscalledadistributedoperatingsystem.8Distributedoperatingsystem 9 9ChapterChapter44.74.8Figure6.1Systemlayers操作系统层libraries&

communication,...communication,...Computer&Computer& Kernelsandserverprocessesarerequiredatleast Concurrentprocessing并发处理：客户应该可以共享 Communication通信：资源管理器在网络上或计算机Figure CoreOSfunctionality操作系统内核功 CoreO/SProcessmanagementincludescreating,managinganddestroyingprocesses.EveryprocesshasanaddressspaceandoneormoreThreadmanagementincludescreating,synchronizingCommunicationsmanagementincludesallcommunicationsbetweenthreadsinthesamecomputerandmayincluderemote MemorymanagementincludescontrolofphysicalandTheO/SSupervisorTheSupervisor：dispatchesinterrupts,systemcalltrapsandexceptions,thecontrolofmemorymanagementunitandhardwarecaches,andmanipulationoftheprocessorandfloatingpointChapterChapter44.74.8ProtectionAllresourcesmustbeprotectedfromNotethatthethreattoasystem’sintegritydoesnotcomeonlyfrommaliciouslycontrivedcode.Benigncodethatcontainsabugorwhichhasunanticipatedbehaviormaycausepartoftherestofthesystemtobehaveincorrectly.对系统完整性的 Protectionmightalsopreventthebypassingofrequiredactivitiessuchasloginauthenticationandauthorization.保护也考 KernelsandMostprocessorshaveahardwaremoderegisterthatpermitsprivilegedinstructionstobestrictlycontrolled.Generally,thereisasupervisormodeforprivilegedinstructionsandausermodeforunprivilegedinstructions. Separateaddressspacesareallocated,andonlytheprivilegedkernelcanaccesstheprivilegedspaces.Usuallyasystemcalltrapisrequiredtoaccessprivilegedinstructionsfromuserspace.地址空间被分开，只有 ProtectionProtectioncomesataprice,includingtheprocessorcyclestoswitchbetweenaddressspaces,thesupervisionandprotectionofsystemcalltraps,andtheestablishment,authenticationandauthorizationofprivilegedusersand Sincealloverheadconsumesexpensiveresources,itisalwaysakeyconcernofITmanagers.ChapterChapter41用进程（任务）作为资源所有权单位(theunitofresourceownership)，量级进程）作为调度/执行/分派单位(theunitof共份资源

Linux、MacOSX、iOS 空间（用来局部变量和调用参数） 2内核级线程(Kernel-levelThreads> 实例：GNUPortablefrom用户级线程（UserLevel 线 实例：Windows、Linux、MacOS

WindowsNT、MacOSX、iOS3. 至强E5645CPU处理器、2 至强E5620CPU处理器、9个448核的NVIDIATeslaC2050GPU运算卡beexecutedsequentiallySpeedup=timetorunonsingleprocessor (1–f)+f/ofMultipleCoresFigure4.7 Figure4.7DatabaseWorkloadsonMultiple-ProcessorHardwareFigure4.8ScalingofDatabaseWorkloadsonMultipleProcessorChapterChapter44.74.8AtypicalprocessincludesanexecutionenvironmentandoneormoreAnexecutionenvironmentincludesanaddressthreadsynchronizationandcommunicationresourcessuchassemaphoresandcommunicationinterfacessuchassockets,andhigherlevelresourcessuchasopenfilesandSinceearlierprocessesmayhaveallowedonlyasinglethread,thetermmulti-threadedprocessisoftenusedforclarity.Figure6.3Addressspace地址空间0AddressAsshownonthepreviousdiagram,anaddressspaceisamanagementunitforthevirtualmemoryofaprocess.Itconsistsof regionsaccessiblebythethreadsoftheowning地址空间是进程虚存管理单元。是一个可以被进程的线程 Eachregionhasanextent(lowestvirtualaddressandsize),read/write/executepermissionsfortheprocessesthreadsandwhetheritgrowsupordown.每个区域包含如下性质：Itispageoriented,andgapsareleftbetweenregionstoallowforUnixAddressAddressspacesareageneralizationoftheUnixmodel,whichhadthreeAfixed,unmodifiabletextregioncontainingprogramAheap,extensibletowardhighervirtualAstack,extensibletowardlowervirtualAnindefinitenumberofadditionalregionshavesincebeenUNIXStackGenerallythereisaseparatestackforeachthread.Wheneverathreadorprocessisinterrupted,statusinformationisstoredonthestackthatwillpermittheprocessorthreadtocontinuefromthepointatwhichitwasinterrupted. 在栈中。容许进程或线程从这个断点继续执行。Usually,memoryallocatedtothestackisrecoveredwhentheprocessorthreadretrievestheinformationandresumes,asinterruptsoccurandresumeinalast-in,firstout(LIFO)manner.FileRegionsAfilestoredofflinecanbeloadedintoactivememory.Amappedfileisaccessedasanarrayofbytesinmemory.Suchafilecanreduceaccessoverheaddramatically,asitisordersofmagnitudefastertoaccessmemorythandiskfiles.映射文件：是一个在内存中可 SharedMemoryRegionsSometimesitisdesirabletosharememorybetweenprocesses,orbetweenaprocessandthekernel.Reasonsforsharingmemoryinclude:Librariesthatmightbelargeandwouldwasmemoryifeachprocessloadedacopy Kernelcallsthataccesssystemcallsand内核 DatasharingandcommunicationbetweenprocessesonsharedAnoperatingsystemcreatesprocessesatneed. adistributedenvironment,therearetwoindependentaspectsofthecreationprocess:ChoiceofatargetCreationofanexecutionenvironmentandaninitialthreadwithinitChoosingHostforProcessAtransfer decideswhethertosituatetheprocesslocallyor Alocation determineswhichnodeshouldhostthenewprocess.Locationpoliciesmaybestaticoradaptive.Staticpoliciesignorethecurrentstateofthesystemandaredesignedbasedontheexpectedlong-termcharacteristicsofthesystem.DynamicLocationPoliciesLoadsharingpoliciesusealoadmangertoallocateprocessestohosts.Sender-initiatedpoliciesrequirethenodecreatingtheprocesstospecifythehost.Migratorypoliciescanshiftprocessesbetweenhostsatanytime.Onceahostisselected,anewprocessrequiresanaddressspacewithinitializedcontentsanddefaultinformationsuchasfiles.Thenewaddressspacecanbedefinedstatically,orcopiedfromanexistingexecutionenvironment.一旦选择了主机，新进程需要一个包含地址空间和初始化信息Ifcopied,contentmaybesharedandnothingwrittentothenewenvironmentuntilsuchtimeasawriteinstructionoccursforeitherprocess.Thenthesharedcontentisdivided.Thistechniqueiscalledcopy-on-write(nextslide). Figure6.4 fromRA

a)Before b)AfterAsingleprocesscanhavemorethanoneactivitythreadsatthesametime.Forexample,itmaybeperforminganactivitywhileatthesametimeneedingtobeawareofbackgroundevents.Theremayalsobebackgroundactivitiessuchasloadinginformationintoabufferfromasocketorfile.Serversmayservicemanyrequestsfromdifferentusersatthesame一个进程可以同时有一个或多个线程。例如，同时执行一个背景活动，如，从socket或文件传送信息到器。在同时，服FigureandserverwithThread1

Thread2makesrequeststoserver

Receipt&

Input-NWorkerPoolArchitecture—apredeterminedfixednumberofthreadsisavailableforuseasneededandreturnedtothepoolafteruse.Thread-per-requestArchitecture—anewthreadisallocatedeachnewrequest,anddiscardedafterThread-per-connectionArchitecture—anewthreadisallocatedforeachnewconnection.Severalrequestscanusethatthreadsequentially.ThethreadisdiscardedwhentheconnectionisThread-per-objectArchitecture—anewthreadisallocatedforeachremoteobject.Allrequestsforthatobjectwaittousethatthread.Thethreadisdiscardedwhentheconnectiontotheobjectisdestroyed.对象线程体系结构：为每个 FigureAlternativeserverthreadingarchitectures(seealsoFigure

Whatisa ItisaprogramspathofMostprogramsrunasasinglethread,whichcouldcauseproblemsifaprogramneedsmultipleeventsoractionstooccuratthesametime.需要多或动作同时发生，就会产生问题。TheJavaVirtualMachineallowsanapplicationtohavemultiplethreadsrunningconcurrently.Whatisathread(cont.)Multi-threadingli llymeansmultiplelinesofasingleprogramcanbeexecutedatthesametime.However,itisdifferentfrommulti-processingbecauseallofthethreadssharethesameaddressspaceforbothcodeanddata,causingittobelessoverhead.So,bystartingathreadanefficientpathofexecutioniscreatedwhilestillsharingtheoriginaldataareafromtheFigureStateassociatedwithexecutionenvironmentsandExecution AddressspacetablesCommunicationinterfaces,openfilesSemaphores,othersynchronization

SavedprocessorPriorityandexecutionstate(suchSoftwareinterrupthandlingListofthread Executionenvironment Pagesofaddressspaceresidentinmemory;hardwarecache ThreadstatesRunningState:Athreadissaidtobeintherunningstatewhenitisbeingexecuted.ReadyState(NotRunnable):Athreadinthisstateisreadyforexecution,butisnotcurrentlyexecuted.OnceathreadgetsaccesstotheCPU,itgetsmovedtotheRunningstate. ThreadstatesDeadState:Athreadreachesthe“dead”statewhentherunmethodhasfinishedexecution.WaitingState(yielding):Inthisstatethethreadiswaitingforsomeactiontohappen.Oncethatactionhappens,thethreadgetsintothereadystate.Threadsinthewaitingstatecouldbe ,suspended,blocked,orwaitingfora等待态：在这个状态线程等待的发生。一旦时间发生，线程转成就绪态。等待态的线程可以睡眠、挂起、阻塞或等UsagesofThreadsThreadsareusedforallsortsofapplications,fromgeneral ctivedrawingapplicationstogames.Forinstanceaprogramisnotcapableofdrawingpictureswhilereadingkeystrokes.Sotheprogrameitherhastogivefullattentiontolisteningtokeystrokesordrawingpictures,otherwiseonethreadcanlistentothekeyboardwhiletheotherdrawsthepictures.UsagesofThreadsAnothergoodusageofthreadsinonasystemwithmultipleCPUsorcores.InthiscaseeachthreadwouldberunonaseparateCPU,resultingistrueparallelisminsteadoftimesharing.FigureJavathreadconstructorandmanagement CreatesanewthreadintheSUSPENDEDstate,whichwillbelongtogroupandbeidentifiedasname;thethreadwillexecutetherun()methodoftarget.Setandreturnthethread’spriority.Athreadexecutestherun()methodofitstargetobject,ifithasone,andotherwiseitsrun()method(ThreadimplementsJavathreadconstructorandmanagementmethods ChangethestateofthethreadfromSUSPENDEDto sleep(int CausethethreadtoentertheSUSPENDEDstateforthespecified EntertheREADYstateandinvokethe DestroytheFigureJavathreadsynchronization BlocksthecallingthreadforuptothespecifiedtimeuntilthreadhasInterruptsthread:causesittoreturnfromablockingmethodcallsuchas object.wait(longmillisecs,intBlocksthecallingthreaduntilacallmadetonotify()ornotifyAll()onobjectwakesthethread,orthethreadisinterrupted,orthespecifiedtimehaselapsed.Wakes,respectively,oneorallofanythreadsthathavecalledwait()onCreatingThreadsinTherearetwowaysinwhichJavaallowsforthecreationofathread.ExtendingaImplementinganCreatingThreadsThefirstistosimplyextendtheThreadclassandoverloadtherun()method.importpublicclassExampleextends{publicvoid{}}CreatingThreadsThesecondistoimplementtheRunnableUsingthiswecancreatethesameimportpublicclassExampleimplements{publicvoid{}}Therearetwowaystosynchronizethreadsbothofwhicharebasedonthesynchronizedkeyword,blocksandAsynchronizedmethodisonewhichspecifiessynchronizedaspartofits synchronizedvoidf(){/*body*/Blocksynchronizationtakesanargumentofwhichobjecttolock.Thisallowsanymethodtolockanyobject.synchronized(this){/*body*/Onethingtonoteisthatthesynchronizedkeywordisnotpartofamethod’ssignature,soitcan’tbeinheritedwhensubclassed.SuspendingThreadsLikethestopmethod,suspend()wasalsodeprecatedduetoallowingdeadlocktohappeninsomecases. Similarlytothestop()methodweneedtoaddsomethingtotheclassandtherunmethodtoallowthethreadtosuspenditselfwhenasked,andresumewhenSuspendingThreads(cont.)TherearetwowaysinwhichtosuspendaThethreadcallswait()onThethreadcalls线程调用：wait(，Thentoresumethethreadnotify()ornotifyAll()mustbecalled.然后，恢复线程调用notify(orprivatevolatilebooleansusThread=false;publicvoidsuspendThread(booleanval){susThread=val;}Suspendingpublicvoidrun()ThreadmyThread==Thread.currentThread();while(myThread==exampleThread){trysynchronized(this){}}catch(InterruptedException}}SuspendingThreadsIntheexampleitisimportanttorealizethatthecheckshouldbedoneafterthesleep()inrun()sothatwhenthethreadisresumeditdoesn’t ygobacktothewaitstate.该设置检查，以线程被恢复时，它不能很快地恢FastThreadsisahierarchic,event-basedthreadschedulingsystem.Itmanagesakernelonacomputerwithoneormoreprocessorsandasetofapplicationprocesses.Eachprocesshasauserlevelscheduler,whilethekernelallocatesvirtualprocessorstoprocesses.Partaofthenextslideshowsakernelallocationprocessesonathree-processormachine.Figure6.10Scheduleractivations

P

Ptoprocesses

KernelFigure6.10bthattheprocessnotifiesthekernelwhenavirtualprocessorisidleandnolongerneededorwhenanextravirtualprocessoristhatthekernelnotifiesaprocesswhenascheduleractivationnotifiesthatprocess’schedulerofanevent.Therearefourtypesofscheduleractivationevents,shownonthenextslide. User-LevelSchedulerAnewVirtualProcessorSchedulerActivationSchedulerActivationSchedulerActivationChapterChapter44.74.8InvocationInvocationperformanceisacriticalfactorindistributedsystemdesign.Themoredesignersseparatefunctionalitybetweenaddressspaces,themoreremoteinvocationsarerequired.必须越多的使用调用。Theinvocationofalocalprocesstakes cewithinlocalmemoryandmayinvolveafewtensofinstructioncycles.Invocationofaremoteprocessinvolvesnetworkactivitiesandpossiblyaccesstofiles,andmayrequirebillionsofinstructioncyclesinprocessorsrunningatspeedsmeasuredingigahertz. TheseinvocationactivitiesareexternaltothedesiredprocessingactivitiesandincreasecostswithoutaddingLatencyNullInvocationcostsarethedelaysrequiredtosetupcommunicationsandthenon-goalperformingoverheadofinvocations.Thesefixedoverheadcostsmeasurethelatencyoftheconnection.Substantialeffortsgointominimizingandreducinglatencycostsindistributedapplications.Slide6.11showssomeofthedelaysfordifferenttypesofinvocations.FigureInvocationsbetweenaddress

trapinstruction

Controltransfervia

User

UserRPC/RMI(between

InvocationCostsasapercentageofAsmoregoalworkisplishedforafixedamountofoverhead,theoverheadislessofaconcernasapercentageoftotalWithverysmalldatasizes,mostofthesystemtimebespentinoverheadWithlargedatatransfers,theoverheadcostsmaynegligibleasapercentageoftotalSlide6.12illustratesthisasagraphofRPCagainstpacketsizeinRPCSlide6.12 FigureRPCdelayagainstparameterRPC

size(bytes)RPCdelayagainstparametersizeThedelayisroughlyproportionaltothesizeuntilthesizereachesathresholdataboutnetworkpacketsize.Beyondthatthreshold,atleastoneextrapackethastobesent,tocarrytheextradata.Dependingontheprotocol,afurtherpacketmightbeusedtoacknowledgethisextrapacket.JumpinthegraphoccureachtimethenumberofpacketincreasesLightweightOnewaytoreducetheoverheadofaremoteinvocationistosharesomeofthecosts.Whileprocessinvocationisexpensive,someactivitiescanbedoneonceandthenreused,whileothersmustbedoneforeachcommunication.减少 LightweightRPCattemptstominimizeoverheadbysharingprocessactivitieswithinparentprocessesthatarepre-established.Overheadcostscanbereducedbyasmuchastwo-FigureAlightweightremoteprocedure

AA

andcopyresults

Trapto

ConcurrentInvocationAnotherapproachtoreducingcommunicationoverheadistoreducethenumberofmessagessenttoestablishtheconnectionandtocontinueprocessingwhilecommunicatinginsteadofsuspendingthreadsuntilamessageisreceived.SerializedandconcurrentinvocationsarecomparedonthenextFigureTimesforserializedandconcurrent

Thread-basedSingle isnotgood intsockfd;FILE*fp;{pthread_ttid;fp=fopen(…);/*Starttheusualway*/sockfd=Socket(…);/*Createathreadtosenddata*/Pthread_create(&tid,NULL,write_func,NULL);/*readdatafromsockfd*//*waitforchildthread*/Pthread_join(tid,NULL);}void*write_func(void{charsendline[MAXLINE];while(moredatainfp)ReadfromfpintoWritesendline[]intoShutdown(sockfd,SHUT_WR);}void{charwhile(moredatafromsockfd)Readfromsockfdintorecvline[];Writefromrecvline[]tostdout;}LockingLockinginMutexMutex–formutualintcounter=void*thread_func(void{int/*unprotectedcode*/val=counter;counter=val+return}intcounter=ptread_mutex_tmutex=void*thread_func(void{int/*protectedbymutex*/Pthread_mutex_lock(&mutex);val=counter;counter=val+1;Pthread_mutex_unlock(&mutex);return}ConditionVariable–forThinkofConditionVariable–forProducersandconsumersruninseparateProducerproducesdataandconsumerconsumesProducerhastoinformtheconsumerwhendataisConsumerhastoinformproducerwhenbufferspaceisWithoutWithoutCondition/*Globalsintdata_avail=pthread_mutex_tdata_mutex=void*producer(void{ProducedataInsertdatainto}{/*donothing–keeploo Extractdatafromif(queueisdata_avail=0;}WithWithConditionintdata_avail=pthread_cont_tdata_cond=PTHREAD_COND_INITIALIZER;{ProducedataInsertdataintodata_avail=1;}void*consumer(void{while(!data_avail)/*sleeponconditionvariable*/}/*wokenupExtractdatafromqueue;if(queueisempty)data_avail=}ChapterChapter44.74.8KernelModeandUserMostCPUssupporttwolevelsofInkernelmode,allinstructionsarepermittedandallmemoryandregistersareaccessible. Inusermode,memoryandregisteraccessisGenerally,almostalloperatingsystemscodeisexecutedinkernelmode,whileuserprocessesarelimitedtousermode.在用户模式，内存和寄存器被限制。一般地，几乎所有的KernelArchitectureMonolithicKernelsthatserveallthefunctionsoftheO/Smaynotbeideallysuitedfordistributedapplications.Theyaremassive,undifferentiatedandintractable.AnalternativeistohavethekernelperformthemostbasicionsandallowmicrokernelsthatcanbeadaptedforspecializedfunctionstomanagesystemFigure

FigureTheroleofthemicrokernel ThemicrokernelsupportsmiddlewareviaBasicServicesofanTaskTaskLocal/RemoteInterProcessCommunicationsFileSystemSecurityIntertaskHighPerformanceMulti-userMemoryDynamicRun-TimeExtendedExtendedservicesareacategoryofmiddlewarethatexploitthepotentialofnetworkstodistributeDBMS,TPMonitors,andNetworkOperatingServiceUbiquitousAccesstonetworkfileandprintBinarylargeobjectsGlobaldirectoriesandNetworkYellowAuthenticationandauthorizationSystemNetworkDatabaseandTransactionInternetServerPCAsymmetricMultiprocessingSymmetricMultiprocessingAllprocessorsareMultiserverGraphicalUserInterfaceStructureconsistsofanicon,aprimarywindowwith bar,andoneormoresecondarywindowsSupportingtasksmayhavesecondarywindowsand Anapplicationrepresentsatask,andusersmustfollowthestructureofataskChapterChapter4 虚拟是指通过某种技术把一个物理实体变成若干个逻辑上例如，在单CPU多道分时系统中，通过多道程序技术和分VirtualMachinesandVirtualization enablesasinglePCorservertosimultaneouslyrunmultipleoperatingsystemsormultiplesessionsofasingleOS（单机或amachinecanhostnumerousapplicationsincludingthosethatrunondifferentoperatingsystems,onasingle hostoperatingsystemcansupportaeachhasthecharacteristicsofaparticularOSand,insomeversionsofvirtualization,thecharacteristicsofaparticularhardware

Figure2.13showsatypicalarrangement.Thevirtualmachinemonitor(VMM)

VirtualMemoryVirtualMachineArchitectureInstructionsetarchitectureISA)指令系统体系结ApplicationbinaryinterfaceABI)应用程序二进制Applicationprogramminginterface(API)应用程序HowDoesItWork?Eachvirtualmachineisanemulationofacompletehardwaremachine+guestOS+applications,buttheVMexistsinuserTheVMMand(whenitispresent)thehostoperatingsystemruninkernelspace/mode.HowcanaguestOSrunninginuserspaceexecuteprivileged(特权)instructions?Itcan’t.Whenittries,theexecutionistrappedtotheVMM,whichthentakeschargeofexecutingtheinstructioninquestion.ProcessVirtualProcessVirtualMachine（进程虚拟机Isavirtual tformforexecutingasingleDoesnotlooklikeanyreal TheprocessVMinterprets codethathasbeencompiledtorunonthevirtualmachineExamples:JVM, ’sCLIJITSystemVirtualSystemVirtualMachine（系统虚拟机EachvirtualmachineisanisolatedenvironmentthatsharessystemresourceswithotherVMsappearstotheguestOSandtheappsitrunsasanexactcopyoftherealhardwareTheguestOSman