大图数据上顶点驱动的并行最小生成树算法.pdfVIP

计算 机研 究 与发展 DOI：10．7544／issnl000l239．2014 JournalofComputerResearchandDevelopment 51(12)：2688—2701，2014 大图数据上顶点驱动的并行最小生成树算法 谷 峪 杨佳学 鲍玉斌 于 戈 (东北大学信息科学与T程学 院 沈 阳 110819) (guyu@ ise．neu．edu．cn) Vertex—DrivenParallelM inimum SpanningTreeAlgorithmsonLargeGraphs GuYu，YangJiaxue，BaoYubin，andYuGe (Collegeo／InformationSc、ienceandEngineering，NortheasternUniversity，Shenyang110819) Abstract Theminimum spanningtree(M ST)algorithm isoneofthemostclassicalgorithmsin the graphtheory．Somecomplexgraph algorithmsbasedon MST includingclustering，classificationand shortestpath querieshavebeen improved significantlyin termsofefficiencyand quality．However， withtherapid developmentofInternet，thescalesofgraphshavebeen becoming largerand larger ． Largescalegraphswhich containmillionsoreven billionsofverticeshavebecomemore common． Therefore，how toimplementqueryprocessinganddataminingalgorithmson largescalegraphshas becomeaproblem tobesolvedurgently．Inaddition，becauseofthedynam icpropertiesoflarge—scale graphs，how to maintain resultsdynamically hasalso becomeoneofthemostattractiveproblems ． However，thestateoftheartM ST algorithmscan’thandlesuchmassiveanddynamicgraphdata．In thispaper，weproposeavertex-driven parallelM ST algorithm calledPB basedon apartition Prim algorithm named PP，and demonstratethecorrectnessofPB．Moreover，we implementthewhole processofPB algorithm on theMapReduceandBSP framework respectively．Taking deletion—only graphsinto consideration，we put forward a maintenance algorithm for M ST which can conduct efficientincrementalevaluation．A1lthecomputation andmaintenancecostsarefurtheranalyzedand c