Download - Capacity Planning for Cloud Computing
CapacityPlanningintheCloudanini2alpeekintoanewworld
CMG08PanelSessionAdrianCockcro=‐Ne@lix
PaulStrong–eBay
WhatisCloudCompu2ng?hGp://www.slideshare.net/StuC/cloud‐compu2ng‐for‐architects‐qcon‐2008‐tutorial‐presenta2on
WhatisCapacityPlanning
• WecareaboutCPU,Memory,NetworkandDiskresources,andApplica2onresponse2mes
• Weneedtoknowhowmuchofeachresourceweareusingnow,andwilluseinthefuture
• Weneedtoknowhowmuchheadroomwehavetohandlehigherloads
• Wewanttounderstandhowheadroomvaries,andhowitrelatestoapplica2onresponse2mesandthroughput
CapacityPlanningNorms
• Capacityisexpensive• Capacitytakes2metobuyandprovision
• Capacityonlyincreases,can’tbeshrunkeasily• Capacitycomesinbigchunks,paidupfront
• Planningerrorscancausebigproblems
• Systemsareclearlydefinedassets
• Systemscanbeinstrumentedindetail
CapacityPlanninginClouds
• Capacityisexpensive• Capacitytakes2metobuyandprovision
• Capacityonlyincreases,can’tbeshrunkeasily• Capacitycomesinbigchunks,paidupfront
• Planningerrorscancausebigproblems
• Systemsareclearlydefinedassets
• Systemscanbeinstrumentedindetail
CapacityisexpensivehGp://aws.amazon.com/s3/&hGp://aws.amazon.com/ec2/
• Storage(AmazonS3)– $0.150perGB–first50TB/monthofstorageused– $0.120perGB–storageused/monthover500TB
• DataTransfer(AmazonS3)– $0.100perGB–alldatatransferin– $0.170perGB–first10TB/monthdatatransferout– $0.100perGB–datatransferout/monthover150TB
• Requests(AmazonS3StorageaccessisviahGp)– $0.01per1,000PUT,COPY,POST,orLISTrequests– $0.01per10,000GETandallotherrequests– $0perDELETE
• CPU(AmazonEC2)– Small(Default)$0.10perhourtoExtraLarge$0.80perhour
• Network(AmazonEC2)– Inbound/Outboundaround$0.10perGB
Capacitycomesinbigchunks,paidupfront
• Capacitytakes2metobuyandprovision– Nominimumprice,monthlybilling– “AmazonEC2enablesyoutoincreaseordecreasecapacitywithinminutes,nothoursordays.Youcancommissionone,hundredsoreventhousandsofserverinstancessimultaneously”
• Capacityonlyincreases,can’tbeshrunkeasily– Payforwhatisactuallyused
• Planningerrorscancausebigproblems– Sizeonlyforwhatyouneednow
Systemsareclearlydefinedassets
• Youarerunningina“stateless”mul2‐tenantedvirtualimagethatcandieorbetakenawayandreplacedatany2me
• Youdon’tknowexactlywhereitis• Youcanchoosetolocate“USA”or“Europe”• Youcanspecifyzonesthatwillnotsharecomponentstoavoidcommonmodefailures
Systemscanbeinstrumentedindetail
• Needtousestatelessmonitoringtools• e.g.Ganglia–automa2cconfigura2on– Mul2castreplicatedmonitoringstate
– Noneedtopre‐definemetricsandnodes
December11,2008 AdrianCockcro=andMarioJauvin
Ganglia–www.ganglia.info
• WebbasedRRDtoolGUI• Goodmanagementofclustersofsystemsanddevices,useful
forhundredstothousandsofnodesinahierarchyofclusters• Providesmanysummarysta2s2cplotsatclusterleveland
collectsdetailedconfigura2ondata• XMLbaseddatarepresenta2on• Useslowoverheadnetworkprotocol(mul2castorunicast)• IncommonuseathundredsoflargeHPCGridsites,lessvisibly
inuseatsomelargecommercialsites• hGp://wiki.apache.org/hadoop/AmazonEC2includesganglia
asastandardfeatureofHadooponEC2.