LMS-Virtual.Lab声学计算加速

LMSVirtual.LabREV8AAcousticTrainingLMSVirtual.LabREV8AHowtospeedupyourcomputation?2copyrightLMSInternational-2008ƒSYSNOISERev5.6usesadynamicmemoryallocationƒDuringaresolution(schematic!)ƒMEMORYisaSYSNOISEEnvironmentVariable,tosetthemaximummemorywhichcanbeused,youcanmodifyit:Tools→Options→AcousticsandinthetabLMSSysnoisespecifythemaximummemory.NodesPropsB.C.MatricesResultsMEMORYMemoryAllocation(1)3copyrightLMSInternational-2008Memoryisreservedintotalvirtualmemory,i.e.thesumofphysicalRAMmemoryandswapspace:ƒaccesstophysicalRAMmemoryisextremelyfastƒswapspaceisadedicatedmemoryareaontheharddisk,withmuchsloweraccessƒonUNIXsystems,aprocesscanneverallocatemorememorythantheswapspacesizeƒrule-of-thumb:swapspacesize=2*PhysicalRAMsizeMemoryAllocation(2)4copyrightLMSInternational-2008ƒIN-COREsolverƒworkspaceissufficientlylargetostoresystemmatricesduringassemblyanddecompositionstagesoftheanalysisƒtwoin-coresituationsexistƒworkspacecompletelyinphysicalRAMmemoryƒworkspaceinphysicalRAMmemoryandswapareaƒOUT-OF-COREsolverƒworkspaceisnotsufficientlylargetostoresystemmatricesduringassemblyanddecompositionstagesoftheanalysisƒsystemmatricesareassembledinblockswhicharewrittentoascratchfile(*.TMP)ondisk;decompositionwillalsoproceedblockbyblockandreadblocksfromthescratchfileƒworkspacemustbesufficientlylargetocontaintheassemblyanddecompositionblocks(seelater)In-CoreversusOut-of-CoreSolvers(1)5copyrightLMSInternational-2008ƒSolutionefficiencyƒin-coresolverusingphysicalRAMmemoryonlyƒin-coresolverusingphysicalRAMmemoryandswapspace:muchtimecanbelostin“swapping”fromRAMmemorytotheharddisk(largeelapsedorwall-clocktimecomparedtoCPUStatsinEchoWindow)=inefficient!ƒout-of-coresolveralsoinvolvesdiskaccess,butwritingorreadingofeachblockrequiresasingleoperation:efficientuse!ƒRecommendationsƒletSYSNOISEdecideuponin-coreorout-of-coresolverƒsetMEMORYtoavaluethatdoesnotexceedphysicalRAMmemoryIn-CoreversusOut-of-CoreSolvers(2)6copyrightLMSInternational-2008In-CoreversusOut-of-CoreSolvers(3)ƒIN-COREsolverspeedƒinRev5.6thiswillalwaysbethefastestmethodƒIN-COREsolverlimitsƒForthe32-bitversionofSYSNOISE,absolutelimitforin-corememory:2GBytesRAMƒForthe64-bitversionofSYSNOISE,nopracticallimitforin-corememory7copyrightLMSInternational-2008ƒOUT-OF-COREsolverlimitsƒpracticallimittoproblemsizeimposedbyavailableharddiskspaceforblockstorageƒRev5.3A•single*.TMPscratchfilecontainingtheblocks•absolutelimitformanyUNIXsystems:2GBytessizeƒRev5.4–5.6•I-BEMsolverallowsupto8different*.TMPfiles•environmentvariablesTMPDIR1,TMPDIR2,…•effectivelimit:16GBytesIn-CoreversusOut-of-CoreSolvers(4)8copyrightLMSInternational-2008ƒOUT-OF-COREmemoryallocationtipsƒMemoryallocationmustbesufficientlylargeforsimultaneousstorageofmultipleassemblyordecompositionblocksƒBEMDIRECT:1assemblyblockor2decompositionblocksƒBEMINDIRECT:2assemblyblocksor2decompositionblocksƒInsufficientmemoryallocationforblockstoragewillleadtoanabortedanalysisIn-CoreversusOut-of-CoreSolvers(5)9copyrightLMSInternational-2008ƒSelectionofsolverlibraryisgovernedbytheenvironmentvariableSOLVERƒSOLVER=1⇒LINPACKsolverƒSOLVER=2⇒userdefinedsolverƒSOLVER=3⇒reservedforfutureiterativesolversƒSOLVER=4⇒LAPACKsolverƒSOLVER=7⇒LMSsolverƒSOLVER=8⇒Matrix-leveldistributedsolverƒSettingofSOLVERcanbemodifiedpriortostartingtheresponseanalysisƒDefaultsettingin‘SYSNOISE.STP’file:SOLVER=7SolverLibraries(1)10copyrightLMSInternational-2008ƒDistributedcomputingprinciplesƒnetworked¶llelcomputingƒfrequencylevel&matrixleveldistributionƒNetSolverlicensestructureƒNetSolvertechnicalprinciplesƒMPImechanismƒprocessorscheduling&loadbalancingƒRunningSYSNOISENetSolverƒcommandstructureƒMPIhostfilestructureƒSYSNOISEParallelonIBMSP-2DistributedComputing-NETSOLVER11copyrightLMSInternational-2008ƒLargeSYSNOISEjobscanbeprocessedinparallelonmultipleprocessors:distributedcomputingsplitsthelargejobinsub-jobswhichrunsimultaneouslyoneachprocessorƒRev5.6supportstwotypesofparallelcomputing:ƒon”real”multi-processorparallelmachines(e.g.IBMSP-2;SGIOrigin;HP.-class;DECAlphaServer)ƒonasetofnetworkedsingle-ormulti-processorUNIXworkstations,actingasavirtualparallelmachine(SGIR4K,R8K&R10K;HPB-class,C-class&J-class;IBMRS/6000;DECALPHA;SUNUltraSparc)DistributedComputingPrinciples(1)12copyrightLMSInternational-2008ƒSYSNOISERev5.6:ƒdistributedcomputingisalsosupportedonNTplatforms(networkedsingle-ormulti-processorPC’s)ƒHybridparallelconfigurations(UNIX+NT)arenotsupportedDistributedComputingPrinciples(2)13copyrightLMSInternational-2008ƒFrequencyleveldistributionƒeachprocessorrunsadifferentsubsetofallrequestedanalysisfrequenciesƒavailableforbothnetworkandparallelconfigurationsƒMatrixleveldistributionƒtosolververylargeproblems,thesystemmatrixissplitintoblocks,andeachprocessorassemblesandsolvesadifferentblockƒinRev5.6availableforgeneralUNIXorNTnetworkconfigurationsthroughtheMPImechanismDistributedComputingPrinciples(3)14copyrightLMSInternational-2008NetSolverLicense(1)ƒDistributedcomputingrequiresNetSolverLicense:ƒNumberofKernellicensesƒaSYSNOISEsessionalwaysuses1Kernellicenseƒpre-andpost