ACHIEVING SCALABLE PARALLEL MOLECULAR DYNAMICS USING DYNAMIC

ABSTRACT. To achieve scalable parallel performance in Molecular Dynamics Simulations, we have modeled and implemented several dynamic spatial domain decomposition algorithms. The modeling is based upon the Bulk Synchronous Parallel architecture model (BSP)

ACHIEVINGSCALABLEPARALLELMOLECULARDYNAMICSUSING

DYNAMICSPATIALDOMAINDECOMPOSITIONTECHNIQUES

LARSNYLAND,JANPRINS,RUHUAIYUN,JANHERMANS,HYE-CHUNGKUM,ANDLEIWANG

ABSTRACT.ToachievescalableparallelperformanceinMolecularDynamicsSimulations,

wehavemodeledandimplementedseveraldynamicspatialdomaindecompositionalgo-

rithms.ThemodelingisbasedupontheBulkSynchronousParallelarchitecturemodel(BSP),

whichdescribessuperstepsofcomputation,communication,ing

thismodel,wehavedevelopedprototypesthatexplorethedifferingcostsofseveralspatial

decompositionalgorithms,andthenusethisdatatodriveimplementationofourMolecular

Dynamicssimulator,Sigma.

TheparallelimplementationisnotboundtothelimitationsoftheBSPmodel,allow-

ingustoextendthespatialdecompositionalgorithm.Foraninitialdecomposition,weuse

oneofthesuccessfuldecompositionstrategiesfromtheBSPstudy,andthensubsequently

useperformancedatatoadjustthedecomposition,dynamicallyimprovingtheloadbalance.

Themotivatingreasontousehistoricalperformancedataisthatthecomputationtopredict

abetterdecompositionincreasesincostwiththequalityofprediction,whilethemeasure-

mentofpastworkoftenhashardwaresupport,requiringonlyaslightamountofworkto

modifythedecompositionforfuturesimulationsteps.

Inthispaper,wepresentouradaptivespatialdecompositionalgorithms,theresultsof

modelingthemwiththeBSP,theenhancedspatialdecompositionalgorithm,anditsperfor-

manceresultsoncomputersavailablelocallyandatthenationalsupercomputercenters.

Keywords:Spatialdecomposition,Adaptiveloadbalancing,BSPcostmodeling,Molec-

ulardynamics,Parallelalgorithms

1.INTRODUCTION

AdrivinggoalofourresearchgroupistodevelopahighperformanceMolecularDynam-icssimulatortosupportbiochemistsintheirresearch.Ourgoalsaretostudylargetimescalebehaviorofmoleculesandtofacilitateinteractivesimulationswherethescientistcancon-trolthesimulation[9].Twomaincharacteristicsoftheproblemimpedeourgoal: rstisthelargenumberofinteractionsinrealisticsystems(solvatedbiomolecules),andsecondistheextremelysmallsimulatedtimestepthatisrequiredtoadequatelycapturethehighfre-quencymotions.Algorithmicimprovementsandapproximationtechniqueshavebeenusedtosuccessfullyimproveperformance,someofwhichwillbementionedhere.However,theprimaryfocusofthispaperistheparallelizationofMolecularDynamicssimulationsthatachievesscalableperformance.Todevelopparallelizationstrategiesthatmeetourgoals,wemodeltheproblematahigh-levelwherecompellingresultsaredeveloped,whicharethenusedasabasisforanimplementation.Intheimplementation,additionaloptimiza-tionshavebeenintroduced,requiringfurthermodi cationoftheparallelizationstrategies,resultinginascalableparallelimplementation.

1.1.MolecularDynamics(MD)Simulation.Moleculardynamicssimulatorsuseclassi-calmodelsofmotionandinteraction,ratherthanthemoremodernandcomplexmodelsof

ABSTRACT. To achieve scalable parallel performance in Molecular Dynamics Simulations, we have modeled and implemented several dynamic spatial domain decomposition algorithms. The modeling is based upon the Bulk Synchronous Parallel architecture model (BSP)

2NYLAND,PRINS,YUN,HERMANS,KUM,ANDWANG

quantummechanics,tocomputeandapplyforces.Problemssuchasdockingaligandinareceptor,performingstructurere nementorperformingsequenceanalysisareamongthemanyproblemsthatcanbeexploredwithMDsimulation.

Inthesimulation,acontinuousprocessisbrokendownintodiscretetimesteps,cho-sensmallenoughthatthediscretizationeffectsareminimized.Ateachstep,thesumofallforcesoneachatomiscalculated,andthenappliedwithregardtothedurationofthetimestep,updatingthepositionandvelocityvalues.Theforcesoriginatefrombondedandnon-bondedforcesbetweentheatoms.Thebondedforcesseektomaintainbondlengths,bondangles,anddihedralanglesonsinglebonds,two-bondchainsandthree-bondchains,respectively.Thenon-bondedforcesarecomprisedoftheelectrostaticforcesfromelectri-calchargesandtheLennard-Jonesforces.Thenon-bondedforcesaresymmetric(equalandopposite),occurbetweeneachpairofatoms,andvaryasaninversepowerofthedistancebetweentheatomsinvolved.

OneotheraspectofMDthatrequiresexplanationisthehandlingofboundaries.Therearetwochoices,openboundaryconditionsandperiodicboundaryconditions.Theopenboundaryconditionisasimulationofthemolecularsystemsasifitisinavacuum.Inperi-odicboundaryconditions(PBC),thesystemofatomsistypicallybox-shaped,andiscon-ceptuallyrepeatedto llspace.Thusanyatomsdriftingoutofthesimulationspacereenterthespaceontheoppositesidewiththesamevelocityanddirection.Forcesarealsocalcu-latedusingperiodicspace.Inthesimulationsdescribedhere,periodicboundaryconditionsareused.

Byfar,themosttime-consumingstepofthesimulationisthecomputationofthenon-bondedforces.Asingleatomhasbond-relatedforceswithalimitednumberofotheratoms,

interactions.Itbutthenon-bondedforcesexistbetweenallpairsofatoms,yielding

isinthispartoftheMDsimulationwhereapproximationsaresoughtandaccuracyisgivenup,allinanefforttoimproveperformance.Onesolutiontoreducethecalculationintro-ducesacutoffradius,wherenon-bondedforcesarecalculatedeachsimulationsteponlywhenthedistancebetweenapairofatomsislessthansomepresetradius,.Theremain-inglonger-rangeforcesareeithercalculatedbysomeothermethod(theFastMul-

Particle-MeshEwaldmethod[5]),calculatedlesstipoleAlgorithm[2],orthe

frequently,orcompletelyignored.

Ofcourse,thechoiceoftimestepdurationisanotherapproximation,wherelongertimestepsyieldfastercalculationswithlessaccurateresults.Thisisanintegrationprocessofanon-linearsystem,thusthetimestepsmustbesmallenoughtoaccountforthemostrapidchangesthatarelikelytooccur.Someworkhasbeendoneto ndalgorithmicsolutionsthatallowlongertimestepsw …… 此处隐藏：6044字，全部文档内容请下载后查看。喜欢就下载吧 ……