并行程序設計 中文課件 08 基于消息傳遞的程序設計技術(shù)

ParallelProgrammingInstructor:ZhangWeizhe(張偉哲)ComputerNetworkandInformationSecurityTechniqueResearchCenter,SchoolofComputerScienceandTechnology,HarbinInstituteofTechnologyProgrammingUsingtheMessage-PassingParadigm3IntroductionProgramingwithMPIProgrammingwithPVMComparisionwithMPIandPVMOutline4AParallelMachineModelInterconnect…TheclusterAnodecancommunicatewithothernodesbysendingandreceivingmessagesoveraninterconnectionnetwork可以通過在互連網(wǎng)絡上發(fā)送和接收消息來與其他節(jié)點進行通信的節(jié)點ThevonNeumanncomputer5Principlesof

Message-PassingProgrammingEachprocessorinamessagepassingprogramrunsaseparateprocess(sub-program,task)Thelogicalviewofamachinesupportingthemessage-passingparadigmconsistsofpprocesses,eachwithitsownexclusiveaddressspace.AllvariablesareprivateEachdataelementmustbelongtooneofthepartitionsofthespace;hence,datamustbeexplicitlypartitionedandplaced.CommunicateviaspecialsubroutinecallsAllinteractions(read-onlyorread/write)requirecooperationoftwoprocesses-theprocessthathasthedataandtheprocessthatwantstoaccessthedata.6Principlesof

Message-PassingProgramming消息傳遞程序中的每個處理器都運行一個單獨的進程（子程序，任務）支持消息傳遞范例的機器的邏輯視圖由p個進程組成，每個進程都有自己的獨占地址空間。所有變量都是私有的每個數(shù)據(jù)元素必須屬于該空間的一個分區(qū);因此，必須明確分區(qū)和放置數(shù)據(jù)。通過特殊的子程序通話所有交互（只讀或讀/寫）都需要兩個進程的協(xié)作，這兩個進程是具有數(shù)據(jù)訪問數(shù)據(jù)的進程。7Principlesof

Message-PassingProgrammingSPMDSingleProgramMultipleDataSameprogramrunseverywhere EachprocessonlyknowsandoperatesonasmallpartofdataMPMDMultipleProgramMultipleData Eachprocessperformadifferentfunction(input,problemsetup,solution,output,display)8MessagesMessagesarepacketsofdatamovingbetweenprocessesThemessagepassingsystemhastobetoldthefollowinginformation:SendingprocessSourcelocationDatatypeDatalengthReceivingprocess(es)DestinationlocationDestinationsize9MessagePassingMessage-passingprogramsareoftenwrittenusingtheasynchronous異步orlooselysynchronous松散同步paradigms.Asynchronouscommunicationdoesnotcompleteuntilthemessagehasbeenreceived.Anasynchronouscommunicationcompletesassoonasthemessageisonitsway

10IntroductionProgramingwithMPIProgrammingwithPVMComparisionwithMPIandPVMOutline11WhatisMPI?ThedevelopmentofMPIstartedinApril1992.MPIwasdesignedbytheMPIForum(adiversecollectionofimplementors,librarywriters,andendusers)quiteindependentlyofanyspecificimplementation

MPI由MPI論壇（實施者，圖書館作家和最終用戶的多樣化集合）設計，完全獨立于任何具體的實現(xiàn)Website

//mpi/12WhatisMPI?MPIdefinesastandardlibraryformessage-passingthatcanbeusedtodevelopportablemessage-passingprogramsusingeitherCorFortran.Afixedsetofprocessesiscreatedatprograminitialization,oneprocessiscreatedperprocessor

mpirun–np5programEachprocessknowsitspersonalnumber(rank)EachprocessknowsnumberofallprocessesEachprocesscancommunicatewithotherprocessesProcesscan’tcreatenewprocesses(inMPI-1)13WhatisMPI?MPI定義了消息傳遞的標準庫，可用于使用C或Fortran開發(fā)便攜式消息傳遞程序。在程序初始化時創(chuàng)建一組固定的進程，每個處理器創(chuàng)建一個進程每個進程知道其個人號碼（等級）每個進程知道所有進程的數(shù)量每個進程都可以與其他進程進行通信進程無法創(chuàng)建新進程（在MPI-1中）14MPI:theMessagePassingInterfaceTheminimalsetofMPIroutines.MPI_InitInitializesMPI.MPI_FinalizeTerminatesMPI.MPI_Comm_sizeDeterminesthenumberofprocesses.MPI_Comm_rankDeterminesthelabelofcallingprocess.MPI_SendSendsamessage.MPI_RecvReceivesamessage.15StartingandTerminatingtheMPILibrary

MPI_InitiscalledpriortoanycallstootherMPIroutines.ItspurposeistoinitializetheMPIenvironment.MPI_Finalizeiscalledattheendofthecomputation,anditperformsvariousclean-uptaskstoterminatetheMPIenvironment.Theprototypesofthesetwofunctionsare:

intMPI_Init(int*argc,char***argv) intMPI_Finalize()

MPI_InitalsostripsoffanyMPIrelatedcommand-linearguments.MPI_Init也剝離任何與MPI相關(guān)的命令行參數(shù)。AllMPIroutines,data-types,andconstantsareprefixedby“MPI_”.ThereturncodeforsuccessfulcompletionisMPI_SUCCESS.所有MPI例程，數(shù)據(jù)類型和常量都以“MPI_”作為前綴。成功完成的返回碼為MPI_SUCCESS。16CommunicatorsAcommunicatordefinesacommunicationdomain-asetofprocessesthatareallowedtocommunicatewitheachother.InformationaboutcommunicationdomainsisstoredinvariablesoftypeMPI_Comm.CommunicatorsareusedasargumentstoallmessagetransferMPIroutines.Aprocesscanbelongtomanydifferent(possiblyoverlapping)communicationdomains.MPIdefinesadefaultcommunicatorcalledMPI_COMM_WORLDwhichincludesalltheprocesses.17Communicators通信者定義通信域-允許彼此通信的一組進程。有關(guān)通信域的信息存儲在MPI_Comm類型的變量中。通信器用作所有消息傳輸MPI例程的參數(shù)。進程可以屬于許多不同（可能重疊）的通信域。MPI定義了一個名為MPI_COMM_WORLD的默認通訊器，包括所有進程。18QueryingInformationThe

MPI_Comm_size

and

MPI_Comm_rank

functionsareusedtodeterminethenumberofprocessesandthelabelofthecallingprocess,respectively.Thecallingsequencesoftheseroutinesareasfollows:

intMPI_Comm_size(MPI_Commcomm,int*size) intMPI_Comm_rank(MPI_Commcomm,int*rank)

Therankofaprocessisanintegerthatrangesfromzerouptothesizeofthecommunicatorminusone.進程的等級是從零到通信者的大小減一的整數(shù)。19OurFirstMPIProgram#include<mpi.h>main(intargc,char*argv[]){ intnpes,myrank;

MPI_Init(&argc,&argv); MPI_Comm_size(MPI_COMM_WORLD,&npes); MPI_Comm_rank(MPI_COMM_WORLD,&myrank); printf("Fromprocess%doutof%d,HelloWorld!\n", myrank,npes);

MPI_Finalize();}20ParallelProgrammingWithMPICommunication通訊Basicsend/receive(blocking)基本發(fā)送/接收（阻塞）Non-blocking非阻塞Collective集體Synchronization同步Implicitinpoint-to-pointcommunication隱含的點對點通信Globalsynchronizationviacollectivecommunication通過集體交流進行全球同步ParallelI/O(MPI2)21Basic

SendingandReceivingMessagesThebasicfunctionsforsendingandreceivingmessagesinMPIaretheMPI_SendandMPI_Recv,respectively.Thecallingsequencesoftheseroutinesareasfollows:

intMPI_Send(void*buf,intcount,MPI_Datatype datatype,intdest,inttag,MPI_Commcomm) intMPI_Recv(void*buf,intcount,MPI_Datatype datatype,intsource,inttag, MPI_Commcomm,MPI_Status*status)

MPI_Send22MPI_SendThemessagetobesentisdeterminedbypointingtothememoryblock(buffer),whichcontainsthemessage.Thetriad,whichisusedtopointtothebuffer(buf,count,type),isincludedintotheparametersofpracticallyalldatapassingfunctions,Theprocesses,amongwhichdataispassed,shouldbelongtothecommunicator,specifiedinthefunctionMPI_Send,Theparametertagisusedonlywhenitisnecessarytodifferentiateamongthemessagesbeingpassed.Otherwise,anarbitraryintegernumbercanbeusedastheparametervalue23MPI_Send要發(fā)送的消息通過指向包含該消息的存儲器塊（緩沖器）來確定。用于指向緩沖區(qū)（buf，count，type）的三元組被包含在幾乎所有數(shù)據(jù)傳遞函數(shù)的參數(shù)中，數(shù)據(jù)傳遞的過程應屬于MPI_Send函數(shù)中指定的通信器，僅當需要區(qū)分正在傳遞的消息之間時才使用參數(shù)標簽。否則，可以使用任意整數(shù)作為參數(shù)值24MPI_Recv2526SendingandReceivingMessages

MPIallowsspecificationofwildcardargumentsforbothsourceandtag.IfsourceissettoMPI_ANY_SOURCE,thenanyprocessofthecommunicationdomaincanbethesourceofthemessage.IftagissettoMPI_ANY_TAG,thenmessageswithanytagareaccepted.Onthereceiveside,themessagemustbeoflengthequaltoorlessthanthelengthfieldspecified.27SendingandReceivingMessages

MPI允許為源和標簽指定通配符參數(shù)。如果source設置為MPI_ANY_SOURCE，則通信域的任何進程都可以是消息的源。如果標簽設置為MPI_ANY_TAG，則會接受帶有任何標簽的消息。在接收端，該消息的長度必須等于或小于指定的長度字段。28MPIDatatypes

MPIDatatypeCDatatypeMPI_CHARsignedcharMPI_SHORTsignedshortintMPI_INTsignedintMPI_LONGsignedlongintMPI_UNSIGNED_CHARunsignedcharMPI_UNSIGNED_SHORTunsignedshortintMPI_UNSIGNEDunsignedintMPI_UNSIGNED_LONGunsignedlongintMPI_FLOATfloatMPI_DOUBLEdoubleMPI_LONG_DOUBLElongdoubleMPI_BYTEMPI_PACKED29Point-to-pointExampleProcess0 Process1#defineTAG999floata[10];intdest=1;MPI_Send(a,10,MPI_FLOAT,dest,TAG,MPI_COMM_WORLD);#defineTAG999MPI_Statusstatus;intcount;floatb[20];intsender=0;MPI_Recv(b,20,MPI_FLOAT,sender,TAG,MPI_COMM_WORLD,&status);MPI_Get_count(&status,MPI_FLOAT,&count);30Non-blockingCommunicationInordertooverlapcommunicationwithcomputation,MPIprovidesapairoffunctionsforperformingnon-blocking非阻塞sendandreceiveoperations.intMPI_Isend(void*buf,intcount,MPI_Datatypedatatype, intdest,inttag,MPI_Commcomm, MPI_Request*request)intMPI_Irecv(void*buf,intcount,MPI_Datatypedatatype, intsource,inttag,MPI_Commcomm, MPI_Request*request)Theseoperationsreturnbeforetheoperationshavebeencompleted.FunctionMPI_Testtestswhetherornotthenon-blockingsendorreceiveoperationidentifiedbyitsrequesthasfinished.intMPI_Test(MPI_Request*request,int*flag, MPI_Status*status)

Non-blockingCommunicationThefollowingschemeofcombiningthecomputationsandtheexecutionofthenonblockingcommunicationoperationispossible:組合計算和非阻塞通信操作的執(zhí)行的以下方案是可能的：31

ProgramsEvaluatingofMPIProgramExecutionTime

Theexecutiontimeneedstoknowforestimatingtheobtainedspeedupofparallelcomputation執(zhí)行時間需要知道估計獲得的并行計算加速Obtainingthetimeofthecurrentmomentoftheprogramexecutionisprovidedbymeansofthefollowingfunction:Theaccuracyoftimemeasurementcandependontheenvironmentoftheparallelprogramexecution.Thefollowingfunctioncanbeusedinordertodeterminethecurrentvalueoftimemeasurementaccuracy:時間測量的準確度可以取決于并行程序執(zhí)行的環(huán)境?？梢允褂靡韵鹿δ軄泶_定時間測量精度的當前值：3233MPICollectiveCommunicationRoutinesthatsendmessage(s)toagroupofprocessesorreceivemessage(s)fromagroupofprocesses將消息發(fā)送到一組進程或從一組進程接收消息Potentiallymoreefficientthanpoint-to-pointcommunication潛在地比點對點通信更有效率ExamplesBroadcastReductionBarrierScatterGatherAll-to-allCollectiveCommunication-BroadcastThereistheneedfortransmittingthevaluesofthevectorXtoalltheparallelprocesses,AnevidentwayistousetheabovediscussedMPIcommunicationfunctionstoprovideallrequireddatatransmissions:Therepetitionofthedatatransmissionsleadstosummingupthelatenciesofthecommunicationoperations,Therequireddatatransmissionscanbeexecutedwiththesmallernumberofiterations34CollectiveCommunication-Broadcast需要將矢量X的值發(fā)送到所有并行進程，一個明顯的方法是使用上述討論的MPI通信功能來提供所有必需的數(shù)據(jù)傳輸：數(shù)據(jù)傳輸?shù)闹貜蛯е驴偨Y(jié)通信操作的延遲，所需的數(shù)據(jù)傳輸可以用較少的迭代次數(shù)執(zhí)行3536CollectiveCommunication-BroadcastTheone-to-allbroadcastoperationis:

intMPI_Bcast(void*buf,intcount,MPI_Datatypedatatype,intsource,MPI_Commcomm)SendsdatafromroottoallothersinagroupCollectiveCommunication-Broadcast3738CollectiveCommunication–ReduceTheall-to-onereductionoperationis:

intMPI_Reduce(void*sendbuf,void*recvbuf,intcount, MPI_Datatypedatatype,MPI_Opop,inttarget, MPI_Commcomm)Combinesdatafromallprocessesingroup-Performs(associative)reductionoperation(SUM,MAX)-ReturnsthedatatooneprocessCollectiveCommunication–Reduce39CollectiveCommunication–ReduceTheBasicMPIOperationTypesforDataReductionFunctions…4041CollectiveCommunication–synchronizationThebarriersynchronizationoperationisperformedinMPIusing:

intMPI_Barrier(MPI_Commcomm)

Abarrieroperationsynchronisesanumberofprocessors.屏障操作同步多個處理器。42CollectiveCommunication–ScatterThecorrespondingscatteroperationis:

intMPI_Scatter(void*sendbuf,intsendcount, MPI_Datatypesenddatatype,void*recvbuf, intrecvcount,MPI_Datatyperecvdatatype, intsource,MPI_Commcomm)

SendseachelementofarrayinroottoseparateprocessCollectiveCommunication–GatherGatheringDatafromAlltheProcessestoaProcess…Gatheringdatafromalltheprocessestoaprocessisreversetodatascattering.ThefollowingMPIfunctionprovidestheexecutionofthisoperation:4344CollectiveCommunication–GatherThegatheroperationisperformedinMPIusing:

intMPI_Gather(void*sendbuf,intsendcount, MPI_Datatypesenddatatype,void*recvbuf, intrecvcount,MPI_Datatyperecvdatatype, inttarget,MPI_Commcomm)

Collectsdatafromsetofprocesses45OtherCollectiveCommunicationMPIalsoprovidestheMPI_Allgatherfunctioninwhichthedataaregatheredatalltheprocesses.

intMPI_Allgather(void*sendbuf,intsendcount, MPI_Datatypesenddatatype,void*recvbuf,intrecvcount, MPI_Datatyperecvdatatype,MPI_Commcomm)Iftheresultofthereductionoperationisneededbyallprocesses,MPIprovides:

intMPI_Allreduce(void*sendbuf,void*recvbuf,intcount,MPI_Datatypedatatype,MPI_Opop,MPI_Commcomm)Tocomputeprefix-sums,MPIprovides:

intMPI_Scan(void*sendbuf,void*recvbuf,intcount,MPI_Datatypedatatype,MPI_Opop,MPI_Commcomm)

46OtherCollectiveCommunicationTheall-to-allpersonalizedcommunicationoperationisperformedby:

intMPI_Alltoall(void*sendbuf,intsendcount, MPI_Datatypesenddatatype,void*recvbuf, intrecvcount,MPI_Datatyperecvdatatype, MPI_Commcomm)Usingthiscoresetofcollectiveoperations,anumberofprogramscanbegreatlysimplified.使用這套核心集體操作，可以大大簡化一些程序。4748TopologiesandEmbeddings

MPIallowsaprogrammertoorganizeprocessorsintologicalk-dmeshes.TheprocessoridsinMPI_COMM_WORLDcanbemappedtoothercommunicators(correspondingtohigher-dimensionalmeshes)inmanyways.Thegoodnessofanysuchmappingisdeterminedbytheinteractionpatternoftheunderlyingprogramandthetopologyofthemachine.MPIdoesnotprovidetheprogrammeranycontroloverthesemappings.

49TopologiesandEmbeddings

MPI允許程序員將處理器組織成邏輯k-d網(wǎng)格。MPI_COMM_WORLD中的處理器ID可以通過多種方式映射到其他通訊器（對應于高維網(wǎng)格）。任何此類映射的優(yōu)點由底層程序的交互模式和機器的拓撲結(jié)構(gòu)決定。MPI不提供程序員對這些映射的任何控制。50TopologiesandEmbeddingsDifferentwaystomapasetofprocessestoatwo-dimensionalgrid.and(b)showarow-andcolumn-wisemappingoftheseprocesses,(c)showsamappingthatfollowsaspace-llingcurve(dottedline)(d)showsamappinginwhichneighboringprocessesaredirectlyconnectedinahypercube.51CreatingandUsing

CartesianTopologies

Wecancreatecartesiantopologiesusingthefunction:我們可以使用以下功能創(chuàng)建笛卡爾拓撲：

intMPI_Cart_create(MPI_Commcomm_old,intndims, int*dims,int*periods,intreorder,MPI_Comm*comm_cart) Thisfunctiontakestheprocessesintheoldcommunicatorandcreatesanewcommunicatorwithdimsdimensions.該功能采用舊通信器中的進程，并創(chuàng)建一個具有dims維度的新通信器。Eachprocessorcannowbeidentifiedinthisnewcartesiantopologybyavectorofdimensiondims.每個處理器現(xiàn)在可以在這個新的笛卡爾拓撲由尺寸DIMS的向量標識。Example:CalculatingπThevalueofconstantπcanbecomputedbymeansoftheintegralTocomputethisintegralthemethodofrectanglescanbeusedfornumericalintegration52Example:CalculatingπCyclicschemecanbeusedtodistributethecalculationsamongtheprocessorsPartialsums,thatwerecalculatedondifferentprocessors,havetobesummed5354Calculatingπ–SequentialProgram

intnum_steps=1000;doublewidth;voidmain() { inti; doublex,pi,sum=0.0; width=1.0/(double)num_steps;

for(i=1;i<=num_steps;i++){ x=(i-0.5)*width; sum=sum+4.0/(1.0+x*x); } pi=sum*width;}55MPIExample–Calculatingπ56MPIExample–Calculatingπ57CollectiveCommunication–Calculatingπ58CollectiveCommunication–Calculatingπ59IntroductionProgramingwithMPIProgrammingwithPVMComparisionwithMPIandPVMOutline60WhatisPVM?

ThedevelopmentofPVMstartedinsummer1989atOakRidgeNationalLaboratory(ORNL).Isasoftwarepackagethatallowsaheterogeneouscollectionofworkstations(hostpool)tofunctionasasinglehighperformanceparallelmachine(virtual)PVM,throughits

virtualmachine

providesasimpleyetusefuldistributedoperatingsystemIthas

daemon

runningonallcomputersmakingupthe

virtualmachine61WhatisPVM?PVM的發(fā)展始于1989年夏天在橡樹嶺國家實驗室（ORNL）。是一個軟件包，允許異構(gòu)收集的工作站（主機池）作為單個高性能并行機器（虛擬）PVM通過其虛擬機提供了一個簡單而有用的分布式操作系統(tǒng)它具有在組成虛擬機的所有計算機上運行的守護程序62PVMResourcesWebsite

/pvm/pvm_home.html

Book PVM:ParallelVirtualMachine

AUsers'GuideandTutorialforNetworkedParallelComputing

AlGeist,AdamBeguelin,JackDongarra,WeichengJiang,RobertManchek,VaidySunderam

/pvm3/book/pvm-book.html

63HowPVMisDesigned64BasicPVMFunctionsEnrollsthecallingprocessintoPVMandgeneratesauniquetaskidentifierifthisprocessisnotalreadyenrolledinPVM.IfthecallingprocessisalreadyenrolledinPVM,thisroutinesimplyreturnstheprocess'stid.將調(diào)用進程注冊到PVM中，并生成唯一的任務標識符，如果此進程尚未注冊在PVM中。如果調(diào)用進程已經(jīng)注冊在PVM中，這個例程只需返回進程的tid。 tid=pvm_mytid();StartsnewPVMprocesses.Theprogrammercanspecifythemachinearchitectureandmachinenamewhereprocessesaretobespawned.開始新的PVM流程。程序員可以指定要生成進程的機器體系結(jié)構(gòu)和機器名稱。numt=pvm_spawn("worker",0,PvmTaskDefault,"",1,&tids[i]);65BasicPVMFunctionsTellsthelocalpvmdthatthisprocessisleavingPVM.ThisroutineshouldbecalledbyallPVMprocessesbeforetheyexit.Addhoststothevirtualmachine.Thenamesshouldhavethesamesyntaxaslinesofapvmdhostfile.pvm_addhosts(hostarray,4,infoarray);Deleteshostsfromthevirtualmachine.

pvm_delhosts(hostarray,4);66BasicPVMFunctionsImmediatelysendsthedatainthemessagebuffertothespecifieddestinationtask.Thisisablocking,sendoperation.Returns0ifsuccessful,<0otherwise.

pvm_send(tids[1],MSGTAG);Multicasts組播amessagestoredintheactivesendbuffertotasksspecifiedinthetids[].Themessageisnotsenttothecallereveniflistedinthearrayoftids.

pvm_mcast(tids,ntask,msgtag);67BasicPVMFunctionsBlocksthereceivingprocessuntilamessagewiththespecifiedtaghasarrivedfromthespecifiedtid.Themessageisthenplacedinanewactivereceivebuffer,whichalsoclearsthecurrentreceivebuffer.

pvm_recv(tid,msgtag);Sameaspvm_recv,exceptanon-blocking非阻塞receiveoperationisperformed.Ifthespecifiedmessagehasarrived,thisroutinereturnsthebufferidofthenewreceivebuffer.Ifthemessagehasnotarrived,itreturns0.Ifanerroroccurs,thenaninteger<0isreturned.

pvm_nrecv(tid,msgtag);68BasicPVMFunctionspvm_barrier("worker",5);pvm_bcast("worker",msgtag);pvm_gather(&getmatrix,&myrow,10,PVM_INT,msgtag,"workers",root);pvm_scatter(&getmyrow,&matrix,10,PVM_INT,

msgtag,"workers",root);pvm_reduce(PvmMax,&myvals,10,PVM_INT,msgtag,"workers",root);69PVMExample:HelloWorld!#include<stdio.h>#include"pvm3.h"main(){intcc,tid;charbuf[100];printf("i'mt%x\n",pvm_mytid());cc=pvm_spawn("hello_other",0,0,"",1,&tid);if(cc==1){ cc=pvm_recv(-1,-1); pvm_bufinfo(cc,0,0,&tid); pvm_upkstr(buf); printf("fromt%x:%s\n",tid,buf);}else printf("can'tstarthello_other\n");pvm_exit();exit(0);}#include"pvm3.h"main(){ intptid; charbuf[100]; ptid=pvm_parent(); strcpy(buf,"hello,worldfrom"); gethostname(buf+strlen(buf),64); pvm_initsend(PvmDataDefault); pvm_pkstr(buf); pvm_send(ptid,1); pvm_exit(); exit(0);}70SetuptoUsePVMSetPVM_ROOTandPVM_ARCHinyour.cshrcfileBuildPVMforeacharchitecturetypeCreatea.rhostsfileoneachhostlistingallthehostsyouwishtouseCreatea$HOME/.xpvm_hostsfilelistingallthehostsyouwishtouseprependedbyan``&''71StartingPVMBefore

wegooverthestepstocompileandrunparallelPVMprograms,youshouldbesureyoucanstartupPVMandconfigureavirtualmachine.OnanyhostonwhichPVMhasbeeninstalledyoucantype%pvmandyoushouldgetbackaPVMconsolepromptsignifyingthatPVMisnowrunningonthishost.Youcanaddhoststoyourvirtualmachinebytypingattheconsolepromptpvm>addhostnameAndyoucandeletehosts(excepttheoneyouareon)fromyourvirtualmachinebytypingpvm>deletehostnameIfyougetthemessage``Can'tStartpvmd,''thencheckthecommonstartupproblemssectionandtryagain.72StartingPVM在完成編譯并運行并行PVM程序的步驟之前，您應該確保可以啟動PVM并配置虛擬機。在任何安裝了PVM的主機上，您可以鍵入%pvm并且您應該返回PVM控制臺提示符，表示PVM正在此主機上運行。您可以通過在控制臺提示符下鍵入來將主機添加到虛擬機pvm>addhostname并且您可以通過鍵入從虛擬機中刪除主機（您所在的除外）pvm>deletehostname如果您收到消息“無法啟動pvmd”，請檢查常見的啟動問題部分，然后重試。73StartingPVMToseewhatthepresentvirtualmachinelookslike,youcantypepvm>confToseewhatPVMtasksarerunningonthevirtualmachine,youtypepvm>ps–aOfcourseyoudon'thaveanytasksrunningyet;that'sinthenextsection.Ifyoutype``quit"attheconsoleprompt,theconsolewillquitbutyourvirtualmachineandtaskswillcontinuetorun.AtanyUnixpromptonanyhostinthevirtualmachine,youcantype%pvmandyouwillgetthemessage``pvmalreadyrunning"andtheconsoleprompt.Whenyouarefinishedwiththevirtualmachine,youshouldtypepvm>haltThiscommandkillsanyPVMtasks,shutsdownthevirtualmachine,andexitstheconsole.ThisistherecommendedmethodtostopPVMbecauseitmakessurethatthevirtualmachineshutsdowncleanly.YoushouldpracticestartingandstoppingandaddinghoststoPVMuntilyouarecomfortablewiththePVMconsole.AfulldescriptionofthePVMconsoleanditsmanycommandoptionsisgivenattheendofthischapter.74StartingPVM要查看當前虛擬機的外觀，可以鍵入pvm>conf要查看虛擬機上運行的PVM任務，請鍵入pvm>ps–a當然你還沒有任何運行的任務;這在下一節(jié)。如果您在控制臺提示符下鍵入“quit”，則控制臺將退出，但您的虛擬機和任務將繼續(xù)運行。在虛擬機中的任何主機上的任何Unix提示符下，可以鍵入%pvm您將收到消息“pvmalreadyrunning”和控制臺提示。完成虛擬機后，您應該鍵入pvm>halt該命令可以殺死任何PVM任務，關(guān)閉虛擬機，并退出控制臺。這是阻止PVM的推薦方法，因為它確保虛擬機完全關(guān)閉。您應該練習啟動和停止并將主機添加到PVM，直到您對PVM控制臺感到滿意為止。在本章末尾給出了PVM控制臺及其許多命令選項的完整說明。75IntroductionProgramingwithMPIProgrammingwithPVMComparisionwithMPIandPVMOutline76PVMandMPIGoalsPVMAdistributedoperatingsystemPortabilityHeterogeneityHandlingcommunicationfailuresMPIAlibraryforwritingapplicationprogram,notadistributedoperatingsystemportabilityHighPerformanceHeterogeneityWell-definedbehavior77PVMandMPIGoalsPVM分布式操作系統(tǒng)可移植性異質(zhì)性處理通信故障MPI用于編寫應用程序的庫，而不是分布式操作系統(tǒng)可移植性高性能異質(zhì)性明確的行為78WhatisNotDifferent?Portability

–sourcecodewrittenforonearchitecturecanbecopiedtoasecondarchitecture,compiledandexecutedwithoutmodification(tosomeextent)Support

MPMD

programsaswellas

SPMDInteroperability–theabilityofdifferentimplementationsofthesamespecificationtoexchangemessagesHeterogeneity(tosomeextent)

PVM&MPIaresystemsdesignedtoprovideuserswithlibrariesforwritingportable,heterogeneous,MPMDprograms79WhatisNotDifferent?可移植性-為一個架構(gòu)編寫的源代碼可以復制到第二個架構(gòu)，編譯并執(zhí)行，無需修改（在某種程度上）支持MPMD程序以及SPMD互操作性-相同規(guī)范的不同實現(xiàn)交換消息的能力異質(zhì)性（某種程度上）PVM和MPI是為用戶提供寫入便攜式，異構(gòu)的MPMD程序的庫的系統(tǒng)80ProcesscontrolAbilitytostartandstoptasks,tofindoutwhichtasksarerunning,andpossiblywheretheyarerunning.PVM

containsallofthesecapabilities–itcanspawn/killtasksdynamicallyMPI-1

hasnodefinedmethodtostartnewtask.

MPI-2

containfunctionstostartagroupoftasksandtosendakillsignaltoagroupoftasks81Processcontrol能夠啟動和停止任務，查找正在運行的任務，以及可能運行的位置。PVM包含所有這些功能-

它可以動態(tài)地生成/刪除任務MPI-1沒有定義的方法來啟動新任務。MPI-2包含啟動一組任務并向一組任務發(fā)送殺死信號的功能82ResourceControlPVM

isinherently

dynamic

innature,andithasarichsetofresourcecontrolfunctions.HostscanbeaddedordeletedloadbalancingtaskmigrationfaulttoleranceefficiencyMPI

isspecificallydesignedtobe

static

innaturetoimproveperformance83ResourceControlPVM本質(zhì)上是動態(tài)的，它具有豐富的資源控制功能。主機可以添加或刪除負載均衡任務遷移容錯效率MPI是專門設計為靜態(tài)的，以提高性能84Virtualtopology

-onlyfor

MPIConvenientprocessnamingNamingschemetofitthecommunicationpatternSimplifieswritingofcodeCanal