大数据基础设施中心成立.pdf

数据中心基础设施

与运维的思考

清华大学交叉信息研究院

徐葳

AboutMe-WeiXu徐葳

•TsinghuaUPenn(B.S)1999-2003

•Berkeley(M.S.andPh.D.)2003-2010

•Advisors:DavidPattersonandArmandoFox

•LongtermvisitingresearcheratGoogle’sSREteam

•Google2010–2013

•MonitoringandDebuggingtools:Logginglibraries,Dapper

(Tracing)

•Tsinghuafrom2013

•InstituteforInterdisciplinaryInformationSciences(IIIS)

•ResearchArea

•Distributedsystems+Machinelearning

•Interdisciplinary“Bigdata”Applications22

MyweirdjobdescriptionatTsinghua

•AssistantProfessor

•Director,OpenComputeCertificationLabatTsinghua

•AdvisoryCommitteememberforDataScienceinstitute

•Associatedirectorforonlineeducation

•...…

•(unofficiallyandunfortunately)SystemAdministrator

•sshintomachinestofixingmisconfiguredOpenstack+

networkat2am…

3

Ourproductionclusterand

OpenComputeLab

~300servers

（200*2Uservers

100*FacebookOCP）

OpenStack(125nodes)

Ceph(60nodes)

Hadoop

Spark

…

4

Everythingmanagedby…

5

But…

everythingsmartrunsonadumb

datacenterinfrastructure

needsconstantbaby-sitting!

6

Outline

•Thenewtrendsofdatacenterfailures

•G.Wanget.al.WhatCanWeLearnfromFourYearsof

DataCenterHardwareFailures?DSN’2017[BestPaper]

•Thenewtrendsofinfrastructuretechnologies

•S*:Ops-DevOps

•Scripts/commands-AIandnaturallanguage

•Consensuseverywhere

EvolvingFailureModel

•Today’sdatacentersaredifferent

-ὠBetterfailuredetectionsystems,experienced