GPUs for statistical data analysis in HEP: a performance study of GooFit on GPUs vs. RooFit on CPUs

摘要

InordertotestthecomputingcapabilitiesofGPUswithrespecttotraditionalCPUcoresahigh-statisticstoyMonteCarlotechniquehasbeenimplementedbothinROOT/RooFitandGooFitframeworkswiththepurposetoestimatethestatisticalsignificanceofthestructureobservedbyCMSclosetothekinematicalboundaryoftheJψinvariantmassinthethree-bodydecayB+→JψK+.GooFitisadataanalysisopentoolunderdevelopmentthatinterfacesROOT/RooFittoCUDAplatformonnVidiaGPU.TheoptimizedGooFitapplicationrunningonGPUshostedbyserversintheBariTier2providesstrikingspeed-upperformanceswithrespecttotheRooFitapplicationparallelisedonmultipleCPUsbymeansofPROOF-Litetool.Theconsiderablyresultingspeed-up,whilecomparingconcurrentGooFitprocessesallowedbyCUDAMultiProcessServiceandaRooFit/PROOF-LiteprocesswithmultipleCPUworkers,ispresentedanddiscussedindetail.BymeansofGooFitithasalsobeenpossibletoexplorethebehaviourofalikelihoodratioteststatisticindifferentsituationsinwhichtheWilksTheoremmayapplyordoesnotapplybecauseitsregularityconditionsarenotsatisfied...