会议专题

ParallelSubgraphListinginaLarge-ScaleGraph

  Subgraphlistingisafundamentaloperationtomanygraphandnetworkanalyses.Theproblemitselfiscomputationallyex pensiveandiswell-studiedincentralizedprocessingalgorithms.However,thecentralizedsolutionscannotscalewelltolarg egraphs.Recently, severalparallelapproachesareintroducedtohandlethelarge graphs.Unfortunately, theseparallelappr oachesstillrelyontheexpensivejoinoperations,thuscannotachievehighperformance.Inthispaper, wedesign anovelparallelsubgraphlisting framework,namedPSgL.ThePSgLiterativelyenumeratessu bgraphinstancesandsolvesthesubgraphlistinginadivide-and-conquerfashion.Theframeworkcompletelyreliesonthegr aphtraversal,andavoidstheexplicitjoinoperation.Moreover,inordertoimproveitsperformance,weproposeseveralsoluti onstobalancetheworkloadandreducethesizeofintermediateresults.Specially,weprovetheproblemofpartialsubgraphins tancedistributionforworkloadbalanceisNP-hard,andcarefullydesignasetofheuristicstrategies.Tofurthermeducetheenor mousintermediateresults,weintroducethreeindependentmechanisms,whichareautomorphismbreakingofthepatterngr aph,initialpatternvertexselectionbasedonacostmodel,andapruningmethodbasedonalight-weightindex.WehaveimplementedtheprototypeofPSgL,andruncomprehensive experimentsof various graphlisting operations on diverselargegraphs.TheexperimentsclearlydemonstratethatPSgLisrobustandcanachieveperformancegainovert hestate-of-the-artsolutionsupto90%.

Graph GraphAlgorithm SubgraphListing GraphEnumeration

YingxiaSHAO BinCUI LeiCHEN LinMA JunjieYAO NingXU

国际会议

第十二届全国博士生学术年会——计算机科学与技术专题

昆明

英文

275-292

2014-05-01(万方平台首次上网日期,不代表论文的发表时间)