Introduction to MPI

《Introduction to MPI》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、1.MessagePassingFundamentalsMessagePassingFundamentalsAsaprogrammer,youmayfindthatyouneedtosolveeverlarger,morememoryintensiveproblems,orsimplysolveproblemswithgreaterspeedthanispossibleonaserialcomputer.Youcanturntoparallelprogrammingandparallelcomputerstosatisfyt

2、heseneeds.UsingparallelprogrammingmethodsonparallelcomputersgivesyouaccesstogreatermemoryandCentralProcessingUnit(CPU)resourcesnotavailableonserialcomputers.Hence,youareabletosolvelargeproblemsthatmaynothavebeenpossibleotherwise,aswellassolveproblemsmorequickly.One

3、ofthebasicmethodsofprogrammingforparallelcomputingistheuseofmessagepassinglibraries.Theselibrariesmanagetransferofdatabetweeninstancesofaparallelprogramrunning(usually)onmultipleprocessorsinaparallelcomputingarchitecture.Thetopicstobediscussedinthischapterare!Theba

4、sicsofparallelcomputerarchitectures.!Thedifferencebetweendomainandfunctionaldecomposition.!Thedifferencebetweendataparallelandmessagepassingmodels.!Abriefsurveyofimportantparallelprogrammingissues.1.1.ParallelArchitecturesParallelArchitecturesParallelcomputershavet

5、wobasicarchitectures:distributedmemoryandsharedmemory.Distributedmemoryparallelcomputersareessentiallyacollectionofserialcomputers(nodes)workingtogethertosolveaproblem.Eachnodehasrapidaccesstoitsownlocalmemoryandaccesstothememoryofothernodesviasomesortofcommunicati

6、onsnetwork,usuallyaproprietaryhigh-speedcommunicationsnetwork.Dataareexchangedbetweennodesasmessagesoverthenetwork.Inasharedmemorycomputer,multipleprocessorunitsshareaccesstoaglobalmemoryspaceviaahigh-speedmemorybus.Thisglobalmemoryspaceallowstheprocessorstoefficie

7、ntlyexchangeorshareaccesstodata.Typically,thenumberofprocessorsusedinsharedmemoryarchitecturesislimitedtoonlyahandful(2-16)ofprocessors.Thisisbecausetheamountofdatathatcanbeprocessedislimitedbythebandwidthofthememorybusconnectingtheprocessors.Thelatestgenerationofp

8、arallelcomputersnowusesamixedshared/distributedmemoryarchitecture.Eachnodeconsistsofagroupof2to16processorsconnected1IntroductiontoMPI–createdbyt