+ All Categories
Home > Documents > wwww www - University of Richmondggilfoyl/research/justin2013DNP.pdfwwww www wwww...

wwww www - University of Richmondggilfoyl/research/justin2013DNP.pdfwwww www wwww...

Date post: 09-Jul-2020
Category:
Upload: others
View: 53 times
Download: 0 times
Share this document with a friend
1
Applicability of Parallel Computing to Partial Wave Analysis (PWA) J. Ruger - Christopher Newport University G.P. Gilfoyle - University of Richmond D.P. Weygand - Jefferson Lab Introduction Jefferson Laboratory zJLR9H is a US national laboratory built to probe the atomic nucleus and to unravel the quarkVgluon nature of matterT Our best theory of the strong interaction that binds nuclei together is Quantum jhromodynamics zQj+H which is not well understood in the nuclear environmentT We are developing new technologies to analyze the large data sets that will be produced by one of the large detectors under construction in Hall 9 as part of the 0FGeV UpgradeT jLRS0F The primary goal of experiments using the jLRS0F detector at energies up to 0F GeV is the study of the internal nucleon dynamicsT Towards this endG the detector has been tuned for studies of reactions in a wide kinematic range such as inclusive processesT The large acceptance of jLRS0F leads to high data rates since most of the debris from the collision of the electron beam with the target is detectedT The computational requirements to acquireG reconstructG analyzeG and simulate these data are highT We expect to collect about 0,G,,, events per second where each event contains 0, k9ytes of dataT For typical runningG this rate means we will gather about MG,,, G9ytes per day and have to store about one million G9ytes each yearT It is the goal of our group to have the software in place on the day of first beams in jLRS0F to calibrateG reconstructG and analyze these new data and to keep pace with the high ratesT Motivation To understand the strong interactionG or Quantum jhromodynamics zQj+H it is necessary to understand the spectrum of bound states of the interactionG that is mesons and baryonsT One of the best tools for the analysis of these reactions is Partial Wave Rnalysis zPWRHT Using PWR we can extract the properties of individual particles from overlapping mass spectra as seen in FigT 5 and FigT IT WeG when analyzing jLRS0F dataG will have to analyze millions of events that may require these computationally intensive calculationsT PWR requires fitting angular momentum states to large sets of dataT Rt 3,, ms per event per wave it will take roughly nine months to do one fit on the data scale that jLRS0F acquiresT This timing analysis takes into account the number of wavesG number of events and likelihood calculationsT The motivation for pursuing research with Xeon PhiWs and its use in Partial Wave Rnalysis is to minimize the overall time required to perform calculationsT 9y exploiting the vector and multiVthreading capabilities of the MIj processor available from IntelG the likelihood function zFigT IH is multiVthreaded to enhance performance of PWR calculationsT Intel Xeon Phi 1 Jefferson Lab has a 0PVnode Intel Xeon Phi zMIjH cluster 1 Each node contains I Intel Xeon Phi M00,P joVprocessors 1 Intel Xeon Phi M00,P 1 P, cores for a total of FI, threads at 0T,M5 GHz 1 Supports 0P single precision or 4 double precision calcT 1 Jefferson Lab offers unique capability of running applications either in native mode on the jPU or through offloading from the hostT 1 jreating multiVthreaded software packages is easy with the PhiWs using OPENVMP and OPENVMPIT 1 jode created for the Xeon Phi can be ran on single threaded platforms or GPU increasing valueT jomputation Results R representation of the processing power gained from the inclusion of a Xeon Phi is a comparison of the run time for PWR logVlikelihood calculation versus the number of calculations deployed on the Xeon Phi and Host machineT Figure M shows that using the Xeon Phi decreased the processing time by a factor of 0PT This will reduce the time need to do the 7 month PWR fit talked about in motivation to about two weeksT To show the ability of the Xeon Phi to multiVthread logVlikelihood calculationsG Figure P represents the parallel scaling over available threadsT The data showed that maximum efficiency was found using half the available threadsT This will limit the amount of time spent in thread management having 0F, threads always availableT The results show promise in using the Intel Xeon Phi for computationally intensive calculationsT The ability to decrease run time by a factor of 0P allows data processing to happen quicker and minimizes the amount of time spent per fitT Rcknowledgments I would like to thank +rT Jerry Gilfoyle and the University of Richmond for the funding and opportunity to research PWRT SecondG I would like to thank +rT +ennis Weygand for the computational guidance received during the projectT LastG I would like to thank +rT 9alint Joo for his knowledge and guidance when working with the Xeon Phi and Jefferson LabWs Xeon Phi jlusterT π 2 π 2 π V mass spectrum a0z0FP,H aFz05F,H piFz0P3,H FigT 5 FigT I FigT 0 FigT F Phase +ifference The Model isobar raw mass spectrum Our goal is to find a mathematical parametrization that explains the experimental observation that is differential cross sections zFigT 0HT R major portion of this parametrization is the calculation of the Likelihood function zLHG which is defined as a product of probabilitiesT L is calculated using the quotient of the Intensity +istribution zIzτHH and normalization integralT In LG the term outside the product in FigT F is the Poisson probability of observing n events and reflects the fact that we use the extended maximum likelihood methodT The integral in the denominator of the summed term contains the acceptance ηzτ H and is the accepted normalization integralT Rn independent calculationG IzτHG is a sum of amplitudesG squared to account for interference with the variable τ representing the set of variables necessary to define a configuration of the final stateT Within IzτHG the variables V and RG represent the production and decay amplitudes respectivelyT The subscripts k and β for V and RG are the parameters that describe the partial wave decomposition we are usingG α specifying properties of the different intermediate states that do not interfereG such as the spin states of the incoming or outgoing particles in the detectorT The subscript βG on the other handG represents the properties whose differing values do interfereG for instance the spin states of broad resonances produced as intermediate states in a sequential decayT Using the calculations from aboveG PWR is able to extract the properties of the individual particles as seen belowT Figure 5G represents the raw mass spectrum of π 2 π 2 π V from electron scattering on the protonT WhileG Figure I represents the extracted properties from the raw mass spectrumT Rs seen belowG using PWR allows us to find the πFz0P3,H particle that was hidden behind the aFz05F,H in unprocessed dataT FigT M FigT P
Transcript
Page 1: wwww www - University of Richmondggilfoyl/research/justin2013DNP.pdfwwww www wwww goratorgzR9Hggggggggggg gggggVgggTgggggg ggggggggjromodynamigzj+H ggggggggonmeTggggg technologgggggggggggggggge

ApplicabilitywofwParallelwComputingwtoPartialwWavewAnalysisw(PWA)J.wRugerw-wChristopherwNewportwUniversityG.P.wGilfoylew-wUniversitywofwRichmondD.P.wWeygandw-wJeffersonwLab

wwwwwwwwwwwwwwwwwwwwwwwwwwwwIntroductionJeffersongLaboratorygzJLR9HgisgagUSgnationalglaboratorygbuiltgtogprobegthegatomicgnucleusgandgtogunravelgthegquarkVgluongnaturegofgmatterTgOurgbestgtheorygofgthegstrongginteractiongthatgbindsgnucleigtogethergisgQuantumgjhromodynamicsgzQj+HwhichgisgnotgwellgunderstoodgingthegnucleargenvironmentTgWegaregdevelopinggnewgtechnologiesgtoganalyzegtheglargegdatagsetsgthatgwillgbegproducedgbygonegofgtheglargedetectorsgundergconstructiongingHallg9gasgpartgofggtheg0FGeVgUpgradeT

ggggggggggggggggggggggggggggjLRS0FThegprimaryggoalgofgexperimentsgusinggtheggjLRS0Fgdetectorgatgenergiesgupgtog0FgGeVgisgthegstudygofgtheginternalgnucleongdynamicsTgTowardsgthisgendGgthegdetectorghasgbeengtunedgforgstudiesgofgreactionsgingagwidegkinematicgrangegsuchgasginclusivegprocessesTgTheglargegacceptancegofgjLRS0FgleadsgtoghighgdatagratesgsincegmostgofgthegdebrisgfromgthegcollisiongofgthegelectronbeamgwithgthegtargetgisgdetectedTgThegcomputationalgrequirementsgtogacquireGgreconstructGganalyzeGgandgsimulategthesegdatagareghighTgWegexpectgtogcollectgaboutg0,G,,,geventsgpergsecondgwheregeachgeventgcontainsg0,gk9ytesgofgdataTgForgtypicalgrunningGgthisgrategmeansgwegwillggathergaboutgMG,,,gG9ytesgpergdaygandghavegtogstoregaboutgonegmilliongG9ytesgeachgyearTgItgisgtheggoalgofgourggroupgtoghavegthegsoftwaregingplacegongthegdaygofgfirstgbeamsgingjLRS0FgtogcalibrateGgreconstructGgandganalyzegthesegnewgdatagandgtogkeepgpacegwithgtheghighgratesT

ggggggggggggggggggggggggMotivationgggTogunderstandgthegstrongginteractionGgorgQuantumgjhromodynamicsgzQj+HitgisgnecessarygtogunderstandgthegspectrumgofgboundgstatesgofgtheginteractionGthatgisgmesonsgandgbaryonsTgOnegofgthegbestgtoolsgforgtheganalysisgofgthesereactionsgisgPartialgWavegRnalysisgzPWRHTgUsinggPWRgwegcangextractgthegpropertiesgofgindividualgparticlesgfromgoverlappinggmassgspectragasgseengingFigTg5gandgFigTgITgWeGgwhenganalyzinggjLRS0FgdataGgwillghavegtoganalyzegmillionsgofgeventsgthatgmaygrequiregthesegcomputationallygintensivegcalculationsTggggggPWRgrequiresgfittinggangulargmomentumgstatesgtoglargegsetsgofgdataTgRtg3,,gmsgpergeventgpergwavegitgwillgtakegroughlygninegmonthsgtogdogonegfitgongthegdatagscalegthatgjLRS0FgacquiresTgThisgtimingganalysisgtakesgintogaccountgthegnumbergofgwavesGgnumbergofgeventsgandglikelihoodgcalculationsTgggggggThegmotivationgforgpursuinggresearchgwithgXeongPhiWsgandgitsgusegingPartialgWavegRnalysisgisgtogminimizegthegoverallgtimegrequiredgtogperformgcalculationsT9ygexploitinggthegvectorgandgmultiVthreadinggcapabilitiesgofgthegMIjgprocessorgavailablegfromgIntelGgtheglikelihoodgfunctiongzFigTgIHgisgmultiVthreadedgtogenhancegperformancegofgPWRgcalculationsTgg

ggggggggggggggIntelgXeongPhi1gJeffersongLabghasgag0PVnodegIntelgXeongPhigzMIjHgcluster1gEachgnodegcontainsgIgIntelgXeongPhigM00,PgjoVprocessors1gIntelgXeongPhigM00,Pgggggg1gP,gcoresgforgagtotalgofgFI,gthreadsgatg0T,M5gGHzgggggg1gSupportsg0Pgsinglegprecisiongorg4gdoublegprecisiongcalcT1gJeffersongLabgoffersguniquegcapabilitygofgrunninggggapplicationsgeithergingnativegmodegongthegjPUgorgggthroughgoffloadingggfromgtheghostT1gjreatinggmultiVthreadedgsoftwaregpackagesgisgeasygggwithgthegPhiWsgusinggOPENVMPgandgOPENVMPITg1gjodegcreatedgforgthegXeongPhigcangbegrangongsingleggthreadedgplatformsgorgGPUgincreasinggvalueT

gg

ggggggggjomputationgResultsRgrepresentationgofgthegprocessinggpowerggainedgfromgtheginclusionofgagXeongPhigisgagcomparisongofgthegrungtimegforgPWRglogVlikelihoodcalculationgversusgthegnumbergofgcalculationsgdeployedgongthegXeongPhigandgHostgmachineTgFiguregMgshowsgthatgusinggthegXeongPhigdecreasedgthegprocessinggtimegbygagfactorgofg0PTgThisgwillgreducegthetimegneedgtogdogtheg7gmonthgPWRgfitgtalkedgaboutgingmotivationgtogaboutgtwogweeksT

gg

TogshowgthegabilitygofgthegXeongPhigtogmultiVthreadglogVlikelihoodgcalculationsGgFiguregPgrepresentsgthegparallelgscalinggovergavailablegthreadsTgThegdatagshowedgthatgmaximumgefficiencygwasgfoundgusinghalfgthegavailablegthreadsTgThisgwillglimitgthegamountgofgtimegspentginthreadgmanagementghavingg0F,gthreadsgalwaysgavailableT

ThegresultsgshowgpromisegingusinggthegIntelgXeongPhigforcomputationallygintensivegcalculationsTgThegabilitygtogdecreasegrungtimegbygagfactorgofg0PgallowsgdatagprocessinggtoghappengquickergandgminimizesgthegamountgofgtimegspentgpergfitT

gggggggggggggggggggggRcknowledgmentsIgwouldglikegtogthankg+rTgJerrygGilfoylegandgthegUniversitygofgRichmondgforgthegfundinggandgopportunitygtogresearchgPWRTgSecondGgIgwouldglikegtogthankgg+rTg+ennisgWeygandgforgthegcomputationalgguidancegreceivedgduringgtheprojectTgLastGgIgwouldglikegtogthankg+rTg9alintgJoogforghisgknowledgegandguidancegwhengworkinggwithgthegXeongPhigandgJeffersongLabWsgXeongPhigjlusterT

π2gπ2gπVgmassgspectrum

a0z0FP,HaFz05F,H

piFz0P3,H

FigTg5

FigTgI

FigTg0 FigTgF

Phaseg+ifference

ThegModel

isobar

rawgmassgspectrum

gggOurggoalgisgtogfindgagmathematicalgparametrizationgthatgexplainsgthegexperimentalgobservationgthatgisgdifferentialgcrossgsectionsgzFigTg0HTgRgmajorgportiongofgthisgparametrizationgisgthegcalculationgofgthegLikelihoodgfunctiongzLHGgwhichgisgdefinedgasgagproductgofgprobabilitiesTgLgiscalculatedgusinggthegquotientgofgthegIntensityg+istributiongzIzτHHgandgnormalizationgintegralTgIngLGthegtermgoutsidegthegproductgingFigTgFgisgthegPoissongprobabilitygofgobservinggngeventsgandgreflectsgthegfactgthatgwegusegthegextendedgmaximumglikelihoodgmethodTgThegintegralgingthegdenominatorgofgthegsummedgtermgcontainsgthegacceptancegηzτgHgandgisgthegacceptedgnormalizationgintegralTggggRngindependentgcalculationGgIzτHGgisgagsumgofgamplitudesGgsquaredgtogaccountgforginterferencegwithgthegvariablegτgrepresentinggthegsetgofgvariablesgnecessarygtogdefinegagconfigurationgofgthefinalgstateTgWithingIzτHGgthegvariablesgVgandgRGgrepresentgthegproductiongandgdecaygamplitudesgrespectivelyTgThegsubscriptsgkgandgβgforgVgandgRGgaregthegparametersgthatgdescribegthegpartialgwavegdecompositiongwegaregusingGgαgspecifyinggpropertiesgofgthegdifferentgintermediategstatesgthatgdognotginterfereGgsuchgasgthegspingstatesgofgthegincominggorgoutgoinggparticlesgingthegdetectorTgThegsubscriptgβGgongthegotherghandGgrepresentsgthegpropertiesgwhosegdifferinggvaluesgdoginterfereGgforginstancegthegspingstatesgofgbroadgresonancesgproducedgasgintermediategstatesingagsequentialgdecayTgggUsinggthegcalculationsgfromgaboveGgPWRgisgablegtogextractgthegpropertiesgofgthegindividualgparticlesgasgseengbelowTgFigureg5Ggrepresentsgthegrawgmassgspectrumgofgπ2gπ2gπVgfromgelectrongscatteringgongthegprotonTgWhileGgFiguregIgrepresentsgthegextractedgpropertiesgfromgthegrawgmassgspectrumTgRsgseengbelowGgusinggPWRgallowsgusgtogfindgthegπFz0P3,HgparticlegthatgwasghiddengbehindgthegaFz05F,HgingunprocessedgdataTg

g

gggggg

FigTgM

FigTgP

Recommended