基于机器学习的并行文件系统性能预测.pdf

ISSN 计算机研究与发展 1000—1239／CN11—1777／TP and Journalof Research Computer Development 基于机器学习的并行文件系统性能预测 赵铁柱1’2 董守斌2 VerdiMarch3’4SimonSee3J 1(五邑大学计算机学院广东江门529020) 2(广东省计算机网络重点实验室(华南理T大学)广州 510641) 3(甲骨文公司技术与云计算中心新加坡) 4(新加坡国立大学计算机科学系新加坡) 5(南洋理工大学机械与航空航天工程系新加坡) (zhao．tiezhu@man．scut．edu．cn) thePara¨eIFile PerformanceViaMachine Predicting System Learning Zhao Shoubin2，VerdiMarch3’1，andSimonSee3’5 Tiezhul仙，Dong 529020) 1(Sc_hooZ U咒iz肥，百打j，，．，in挖g，n￡n，(施竹gdDng o，(二D7挖户MfPrSfiP咒c已，W么yi 2(G如ngdD咒譬KPyL口60mtD删o，C0优p“￡盯NPf伽心(So“咖C矗i撇U行iw邝ityc∥n如挖oZogy)，G眦挖g拍D510641) n雄dCZD“d 3(丁-c．jl行if口Z CDmp“fi以gCP咒fPr，Or口fZF(乃r户or口￡iD，z，Si珂gn户o，v) SciP竹fe，．NⅡ￡io咒口Z 4(De户口r￡，雄P咒fD，C07，lp“￡Pr U打i伽，苫iryo，Si珂g口户Dre，Si，lgnpo，t) Ekgi咒PPri咒g，N口竹y口珂g7■flIl行oZDgi纯ZUkit盱，百ify，Si花g口夕Dr已) 5(D已户口，0聊P竹fo，M已f^盘雄ic口f＆AP，．0s户acP AbstractParallelfile can solvethe ofmassivedata and effectively I／0 system problems storage bottleneck．Becausethe onthe isnot understood，howtoevaluateand potentialimpact systemcIearly of file becomesthe and Inthiswork， predictperformanceparallelsystem potentialchallengehotspot． the evaluationand of file After weaimtoresearch performance predictionparallelsystem． studying thearchitectureand factorsofsuchfile we a modeof performance system，designpredictive parallel file basedonmachine Weusefeatureselection toreducethe learning