Abstract:
—Serverlesscomputinghasbecomeincreasinglypopularforcloudapplications,duetoitscompellingpropertiesof high-levelabstractions,lightweightruntime,highelasticityand pay-per-usebilling.Inthisrevolutionarycomputingparadigm shift,challengesarisewhenadaptingdataanalyticsapplications totheserverlessenvironment,duetothelackofsupport forefficientstatesharing,whichattractever-growingresearch attention.Inthispaper,weaimtoexploittheadvantagesoftasklevelorchestrationandfine-grainedresourceprovisioningfor dataanalyticsonserverlessplatforms,withthehopeoffulfilling thepromiseofserverlessdeploymenttothemaximumextent. Tothisend,wepresentACTS,anautonomouscost-efficient taskorchestrationframeworkforserverlessanalytics.ACTS judiciouslyschedulesandcoordinatesfunctiontaskstomitigate cold-startlatencyandstatesharingoverhead.Inaddition,ACTS explorestheoptimizationspaceoffine-grainedworkloaddistributionandfunctionresourceconfigurationforcostefficiency. WehavedeployedandimplementedACTSonAWSLambda, evaluatedwithvariousdataanalyticsworkloads.Resultsfrom extensiveexperimentsdemonstratethatACTSachievesupto 98%monetarycostreductionwhilemaintainingsuperiorjob completiontimeperformance,incomparisonwiththestate-ofthe-artbaselines.
