ParaView教程及算例Parallel_Visualization_第1页
ParaView教程及算例Parallel_Visualization_第2页
ParaView教程及算例Parallel_Visualization_第3页
ParaView教程及算例Parallel_Visualization_第4页
ParaView教程及算例Parallel_Visualization_第5页
已阅读5页,还剩38页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

ParallelVisualization,ParallelSystemsBackground,“Node”:analogoustoaworkstation,laptopetc:1ormoreCPUs(orcores)Physicalresources:memory,I/Odevices,graphicsetc.“Process”:Aninstanceofaprogram“Virtualized”resourcesnotably,memory“Parallelprocess”Shared-memoryor“Multi-core”1processso1memoryspace1“threadsofexecution”runningon1cores(OpenMP,pthreads)Requiresintra-processsynchronizationDistributed-memory1processCommunicatebymessage-passing(MPI)1process/nodeorprocessesmustsharephysicalresourcesofnode,Otherapproaches:,Co-ProcessorsNodehostsanacceleratorspecialprocessorprogrammedseparatelyfrommainprocessor(s)ProgrammanagesparallelexecutiononacceleratorGPGPU,CellProcessor,Cuda(Nvidia),OpenClHybridsDistributedmulti-threadedprocesseswithaccelerators?,ParallelVisualization,Why?PerformanceProcessingmaybetooslowononeCPUInteractivevisualizationrequiresreal-timeframeratesUselotsofCPUsShared-memory/multicoreordistributedDatamaybetoobigforavailablenodeVirtualmemoryworks,butpagingisslowUselotsofnodestoincreasephysicalmemorysizeBigshared-memory/multicorescalingiscostly($/CPU),TACCParallelVisualizationSystems,SpurorLonghorn,TACCParallelVisualizationSystems,LocalDisk,Memory,SpurorLonghorn,TACCParallelVisualizationSystems,Spur8nodes16corespernode(128total)4GPUspernode(32total)DDR(dualdatarate)InfinibandinterconnectLonghorn256nodes8corespernode(2048total)2GPUspernode(512total)QDR(quaddatarate)Infinibandinterconnect,VisualizationInParallel,ParallelVisualizationAppsAtTACC,EnsightGold,ParaViewandVisIt(IdontknowmuchaboutEnsight,wellfocusonParaViewandVisIt)PVandVisitareverysimilarinmanyways:BasedonVTKClient/ServerArchitecture,ClientApp,ServerProcess,Data,ParallelVisualizationAppsAtTACC,Bothutilizeadistributed,data-parallelserverarchitecture,ClientApp,MappingParallelPV,VisIttoTACCVisualizationSystems,1.Allocatesetofnodesonvisualizationsystem.ThiswillstartaVNCserveroneonenode,whichyouwillconnectto(asbefore),VNCviewer,MappingParallelPV,VisIttoTACCVisualizationSystems,2.Fromthatdesktop,launchthePVorVisItClientApp,VNCviewer,PV/VisItGUI,MappingParallelPV,VisIttoTACCVisualizationSystems,StartPVorVisItServerProcesses-PV,fromcommandline;VisItthroughGUI,VNCviewer,PV/VisItGUI,MappingParallelPV,VisIttoTACCVisualizationSystems,4.Multipleprocesses/nodetotakeadvantageofmultiplecores/node,VNCviewer,PV/VisItGUI,Way-ness:TheNumberofProcessesPerNode,Processmemoryislimitedto1/nofnodememoryforeachofnprocessesonthatnodeIfyouneedlargememoryperprocess,usefewerprocessespernodeWhywouldyouneedto?DataispartitionedinlargechunksVisualizationalgorithmscanexpandthedataWay-nessissetupatallocationtimeParalleljobslaunchedfromVNCdesktopwilladopttheway-nessspecifiedwhentheVNCserverjobislaunchedVNCsessionslaunchedthroughtheportalarefull-way,ParallelSessionsonLonghornviathePortal,VNCpasswordshouldalreadybesetBrowseto:SelectVNCasthesessiontypeThistime,specifynumberofnodes=k,togetknodesor8*kprocessesWay-nesswillbefull“#slots”willbethenumberofprocessesyougetHitSelectandwaitafewseconds.Gotothe“Rendering”tabandenteryourVNCpassword,*Ifyoureturntothe“Resources”pageandfindyourjob,youwillseeitrunningon8*kcores.,ParallelSessionsonLonghornviaqsub,VNCpasswordshouldalreadybesetsshtolonghorn:qsubaVNCsession:qsubqsuboptions/share/sge6.2/default/pe_scripts/job.vncvncoptionsRequiredqsuboptions:-Pproject_typeoneofvis,data,gpgpu,orhpc-qqueueoneofnormal,long,largemem,devel-Aallocationif1available,chooseone-pewaynessncoresjobsizeseveralmore,seethedocs,RunningParaviewinParallel,Startitasbefore:inthewhitexterm,moduleparaviewvglrunparaviewParaviewwillappear2.Inasecondxterm,starttheparallelservermoduleparaview(csh)envNO_HOSTSORT=1ibruntacc_xrunpvserver(bash)NO_HOSTSORT=1ibruntacc_xrunpvserverYouwillsee“Waitingforclient”,RunningParallelParaview(2),ConnectclienttoserverClickFile-ConnecttobringupChooseServerdialogIffirsttime:ClickAddServerSettheserverconfigurationnameto“manual”ClickConfigureFromStartupTypeselectManualandclickSaveSelectmanualfromthelistandclickConnectInclientxterm,youshouldsee“Waitingforserver”Inserverxterm,youshouldsee“Clientconnected.”,RunningVisItinParallel,Itsreallyeasy.moduleloadvisitvglrunvisit,ParaViewInParallel,Demonstrate:ClickSources-SphereMaxoutThetaResolutionandPhiResolutionClickApplyUpcomesawhitesphereClickFilters-Alphabetical-ProcessIdScalarsClickApplySegmentsarecoloredbywhichprocesshandlesthem,ParallelVisualizationAlgorithms#1:TaskParallelism,Dividetheoverallworkflowintotasksthatcanhappenindependentlyand,hence,concurrentlyLikelydoesnotscalewellRenderingmayitselfbeparallel,Time,Processes,Image,Readfile0,Readfile1,Readfile2,Streamline1,Cutplane2,Isosurface0,ParallelVisualizationAlgorithms#1:TaskParallelism,OftendoesnotscalewellToofewindependentpathsSupposeIsosurfacepredominates:,Time,Processes,Image,R.F.0,R.F.1,R.F.2,S.L.1,C.P.2,Isosurface0,ParallelVisualizationAlgorithms#2:DataParallelism,Datasetispartitionedamongtheprocessesandallprocessesexecutesameoperationsonthedata.Scaleswellaslongasthedataandoperationscanbedecomposed.,Image,Time,Readp0,Readp1,Readpn,p0,p1,pn,Dataset0,Similarfordataset1streamlines,Similarfordataset2cutplanes,ParallelVisualizationAlgorithms#2:DataParallelism,Generallyscaleswellif:Tasksareamenabletodata-parallelismManyare(isosurfacesetc.)Afewjustarent(streamlines!)WorkisevenlydistributedoverpartitionsAlternativeistohavepartitionsthanprocessorsforload-balancingInter-PartitionBoundariesInDistributedImplementationsSomeapplicationsrequireneighborhoodinformationtogetrightanswerRequiresaccesstodataassignedtoneighboringpartitions(calledghostzones)Overheadincreaseswithnumberofpartitions,PartitioningforData-Parallelism,PartitioningforData-Parallelism,Aside:WhyGhostZones?,Whathappenswhentheincidenttriangleslieindifferentpartitions?Adiscontinuityandtheeyeisdesignedtopickitup,PerformanceIssues,MemoryUtilizationParallelDataManagementSupportingparallelcomputeSupportingparallelI/ONon-dataparallelalgorithmsParallelrenderingoverhead,MemoryUtilization,Somevisualizationtechniquescausememoryusetoskyrocket!,MemoryUtilization,Regulargrids,Gridandmesharespecifiedby:(xyz)origin(dx,dy,dz)deltavectors(nx,ny,nz)countsRequiresdiddly-squatmemory,MemoryUtilization,Chopoffthecorner-UnstructuredGrids,Specifiedby:ExplicitlistofverticesExplicitlistoftriangles(2D)Memoryusecangoupmanytimes,MemoryUtilization,Badsolution:Clipvolume,extractbundary,MemoryUtilization,BetterSolution:extracttheboundaryandchop,addincap,InParaview,ParallelDataManagement,DatamustbedistributedacrossparallelprocessestotakeadvantageofresourcesHowdoesitgetthere?SomePVandVisItreadersimportinparallelExplicitparallelformatsuseseparatefilesforpartitions(.pvti,global.silo)Implicitparallelformatsparallelprocessesfigureoutwhattheyneed(.vti,brick-of-values)SomemayseemtoimportinparallelActually,importseriallyandthendistributeBadbadbad1processneedsenoughmemoryforentiredatasetplusadditionalspaceforpartitioningAndsomejustdontandleaveittoyou(D3filterinParaview)SeeBadbadbadabove,TACCParallelI/ONumbers,Numberofdisksperserver:-2LogicalUnits(LUNs)-6(+1redundant)disks/LUN,TACCParallelI/ONumbers,FilesystemServers(OSSnodes)-Spurhas50,SDRIBconnect-Longhorn12,QDRIBconnect,LustreFileSystem,Nofilesarelocalanyprocesscanaccessanyfile,oranypartofafileFilesareservedacrossnetworkBandwidthintonodeacrossinterconnectbandwidthfromafewlocaldisksMultiplepinch-pointsremainTotalnumberofdisks(30-40MB/secperdisk)Totalnumberoffileservers(IBconnectivity)LayoutoffilesbecomesimportantFilestripingacrossdisksineachfileserverFilestripingacrossfileservers,LusterOSSStriping,Stripeacross1OSSnodeforperformance(andcapacity)withdiminishing(negative)returnsSpurdefaultsto4,Longhornto2Stripingisunderusercontrollfsgetstripefileseelmm_stripe_countforcurrentstripinglfsgetstripedirstripe_countwillbeappliedtoallfileswrittentothatdirlfssetstripe

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论