Using crowdsourcing, big data and machine (2013)

《Using crowdsourcing, big data and machine (2013)》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、CollectiveMind:cleaninguptheresearchandexperimentationmessincomputerengineeringusingcrowdsourcing,bigdataandmachinelearningGrigoriFursinINRIA,FranceGrigori.Fursin@cTuning.orgAbstractSoftwareandhardwareco-designandoptimizationofHPCsystemshasbe-comeint

2、olerablycomplex,ad-hoc,timeconsuminganderrorproneduetoenor-mousnumberofavailabledesignandoptimizationchoices,complexinteractionsbetweenallsoftwareandhardwarecomponents,andmultiplestrictrequirementsplacedonperformance,powerconsumption,size,reliability

3、andcost.Wepresentournovellong-termholisticandpracticalsolutiontothisproblembasedoncustomizable,plugin-based,schema-free,heterogeneous,open-sourceCollectiveMindrepositoryandinfrastructurewithunifiedwebinterfacesandon-lineadvisesystem.Thiscollaborativef

4、rameworkdistributesanalysisandmulti-objectiveoff-lineandon-lineauto-tuningofcomputersystemsamongmanypar-ticipantswhileutilizinganyavailablesmartphone,tablet,laptop,clusterordatacenter,andcontinuouslyobserving,classifyingandmodelingtheirrealisticbehav

5、-ior.AnyunexpectedbehaviorisanalyzedusingshareddataminingandpredictivemodelingpluginsorexposedtothecommunityatcTuning.orgforcollaborativeexplanation,top-downcomplexityreduction,incrementalproblemdecompositionanddetectionofcorrelatingprogram,architect

6、ureorrun-timeproperties(features).hal-00850880,version1-10Aug2013Graduallyincreasingoptimizationknowledgehelpstocontinuouslyimproveop-timizationheuristicsofanycompiler,predictoptimizationsfornewprogramsorsuggestefficientrun-time(online)tuningandadapta

7、tionstrategiesdependingonend-userrequirements.Wedecidedtoshareallourpastresearchartifactsinclud-inghundredsofcodelets,numericalapplications,datasets,models,universalex-perimentalanalysisandauto-tuningpipelines,self-tuningmachinelearningbasedmetacompi

8、ler,andunifiedstatisticalanalysisandmachinelearningpluginsinapublicrepositorytoinitiatesystematic,reproducibleandcollaborativeresearch,developmentandexperimentationwithanewpublicationmodelwhereexperi-mentsandtechniquesarevalidated,rankedandimprovedbyt