Large-Scale Process Simulation and Optimization in

Large-ScaleProcessSimulationandOptimizationinaHighPerformanceComputingEnvironmentMarkA.StadtherrDepartmentofChemicalEngineeringUniversityofNotreDameNotreDame,IN46556USAE-mail:[e-mail]97,Boston,MA,October12-16,1997AbstractHighperformancecomputing(HPC)technology,includingparalleland/orvectorprocessing,providesopportunitiestosolveprocessoptimizationandsimulationproblemsfasterandmorereliablythaneverbefore,thusenablingthesolutionofincreasinglylargescaleproblems,eveninarealtimeenvironment.ThispresentationwillfocusonrecentadvancesinHPCtechnologyandmethodsforexploitingitinprocessoptimizationandsimulation.Ofparticularinterestaremethodsforthelarge,sparselinearequationsystemsthatoftenariseinlarge-scaleprocessengineeringproblems,andthatoftenrepresentacomputationalbottleneck.Alsoofinterestisanapproachforguaranteeingthereliablesolutionofprocessengineeringproblems.1IntroductionThefuturesuccessofthechemicalprocessindustriesdependsontheabilitytodesignandoperatecomplex,highlyinterconnectedplantsthatareprotableandthatmeetquality,safety,environmentalandotherstandards.Towardsthisgoal,processsimulationandoptimizationtoolsareincreasinglybeingusedindustriallyineverystepofthedesignprocessandinsubsequentplantoperations.Toperformrealisticprocesssimulationforverylargescaleindustrialprocesses,however,requiresadequatecomputationalresources.Today,highperformancecomputing(HPC)technology,includingparalleland/orvectorcomputing,providesthecomputationalpowertoreal-isticallymodel,simulate,design,andoptimizecomplexchemicalmanufacturingprocesses,steady-andunsteady-state.Tobetterusethisleadingedgetechnologyinprocesssimulationrequirestheuseoftechniquesthatecientlyexploitvectorandparallelprocessing.Sincemostcurrentlyusedtechniquesforsolvingsuchproblemsweredevelopedforuseonconventionalserialmachines,itisoftennecessarytorethinkproblemsolvingstrategiesinordertotakefulladvantageofHPCpower.Highperformancecomputingtypicallyinvolvessomeformofparallelprocessing,whichinrecentyearshasbeenrapidlyenteringthemainstreamofcomputertechnology.Thoughsingleprocessorperformancewillcontinuetoimprove,themostimmediatewaytoimproveasystem’sperformanceisthroughtheuseofmultipleprocessors,asopposedtowaitingforthenextgenerationofsingleprocessors.Furthermore,physicallimitsonsingleprocessorspeedswilleventuallybereached.Since1inprinciplethereisnoupperlimitonthespeedofaparallelprocessingmachine,thisrepresentstheinevitablefutureofcomputing,notonlyforthefasteststate-of-the-artsupercomputers,butalsofordesk-basedmachinesandservers.Infactitistheboomingnetworkanddatabaseservermarketthathasbeendrivingparallelprocessingintothemainstreamtoday.Parallelprocessingtakesonmanyforms.TheprocessorsusedmaybeCISC(ComplexIn-structionSetComputing)processors,suchastheIntelPentiumII,RISC(ReducedInstructionSetComputing)processors,suchastheMIPSR10000orSunUltraSPARC-II,orvectorprocessors,suchasusedintheCRAYT90.Whilehistoricallyvectorprocessorshavebeenveryexpensiveandproducedinfairlylowvolume,therearesomeindicationstodaythatvectorprocessingtechnologywillbeultimatelybeincorporatedwithsuper-scalarRISCtechnologyinprocessorsproducedasahighervolumecommodity.Whateverprocessorsareusedinthesystem,theycanbeconnectedinavarietyofways,andcanaccessmemoryinanumberofdierentways.Mostsystemscanbeclassiedaseithershared-memoryordistributed-memory,orsomehybridthereof.Bothhigh-endparallel/vectorsupercomputers,suchastheCRAYT90andlower-endsymmetricmultiprocessors(SMPs),suchasafourprocessorCompaqProLiant6000,featureuniformaccesstoasharedmem-ory.Thisarrangementmaynotscalewelltolargernumbersofprocessors,sootherdesigns,suchastheCray/SGIOrigin2000,featureanonuniformaccesstomemorythatthoughphysicallydis-tributedmaybeconsideredlogicallyshared.Distributed-memorysystemsincludebothmassivelyparallelmachines,suchastheCRAYT3E,andnetwork-basedsystems.Whilefortheformer,ashared-memoryprogrammingmodelmaystillbeusefulinsomecases,forthelatteronemustusu-allyrelyonmessagepassing,usingpopularprotocolssuchasPVM(ParallelVirtualMachine)orMPI(MessagePassingInterface),tomovedatafromprocessortoprocessorandtomemory.Withtherapidadvancementofnetworkingtechnology,network-basedparallelsystemsarebecomingcommon.Essentiallyanycollectionofmachinesonanetworkcanbeusedasaparallelcomputingsystem.Themachinesinthenetwork-basedsystemmayrangefromsimpleworkstationstoSMPstoparallel/vectorsupercomputers.Suchaheterogeneousnetworkofcomputationalresourcesissometimesreferredtoasametacomputer.Theconceptofmetacomputinginthecontextofchem-icalprocessengineeringwasdiscussedoriginallybyStadtherretal.(1993).AnexcellentrecentexampleoftheimplementationofmetacomputingforprocesssimulationandoptimizationistheSimulatorManageratBayerAG(Brull,1997).Thisallowsdierentunitsinalargeproblemtobeconsideredinparallelonthemostappropriatemachinewiththemostappropriatesoftware,withtheoverallproblemconvergedbytheSimulationManagerusingasimultaneous-modularapproach(e.g.,ChenandStadtherr,1985;ChimowitzandBielinis,1987).InSections2-4below,thefocusisonusingvectorandparallelcomputin