CloudCmpComparingPublicCloudProviders

2024年3月17日发(作者：负摩阻力计算)

CloudCmp:ComparingPublicCloudProviders

,,a,

KatrinaLaCurts,MIT(6.897Spring2011)

Choosingapubliccloudproviderisadifficustomermeanttocomparepublic

cloudproviderswheneachoffersamyriadofrvices,instacetypes,andpricingschemes?Evenifa

customercouldmakeasystematiccomparison,wouldtheybeabletodiscernwhichcloudprovider

wasbest—intermsofcost-effectivenessorotherwi—foraparticularapplication?

CloudCmp,publishedatIMC2010,providesasystematicframeworkforcomparingcloudproviders.

heirmethodsandcompareAmazonAWS,RackspaceCloudServers,GoogleApp-

Engine,andMicrosoftAzure

1

.Ratherthancomparethediff,whichpro-

gramminglanguagesareofferedbyeachprovider,thetypeoftechsupportavailable),CloudCmp

measurestheperformanceofthecorervicesthatarecommontothefourproviders:theelas-

ticcomputecluster,persistentstorage(tablestorage,whichofferskey/valuefunctionality;blob

storage,whichallowsurstostoreunstructureddata;andqueuestorage,whichisdesignedto

storemessagesthatarelaterpasdbetweencloudisntances),andnetworking(bothintra-cloud

andWAN).Theirmainfindingisthatnoonecloudisbest;differentcloudsaresuitedfordifferent

ecifically,theyfindthefollowing:

•Badonfinishingtimesforwell-knownJavabenchmarks(SPECjvm2008),differentproviders

offerdifferentlevelsofcost-effmple,alow-endinstanceofMicrosoftAzure

is30%moreexpensivethanthecomparableAmazonAWSinstance,butabletofinishtasks

er,higher-endinstanceswithinoneclouddonotnecessarilyoffer

ult,parallelapplicationsmightdowellto

umorelow-endinstancesratherthanafewhigh-endones.

•Theperformanceofdifferentstoragetypes(table,blob,andqueue)

example,AWS’srespontimefortablequeriesisordersofmagnitudefasterthananyother

provider,butittakeslongerfortheirtablestoreachconsistency(around100msonaverage,

butsometimesashighas500ms;otherprovidernosuchinconsistencies).Blobstorage

andqueuestorageesimilarvariability.

•Thewide-areanetworklatencyroughlycorrespondstogeographiclocation,suggestingthat

themoregeographicallydiveracloudprovideris,hput,bothintra-

cloudandwide-area,isrelativelyconsistentacrossallproviders,exceptforRackspace,which

emstohavearatherunder-provisionednetworkinthemeasurements;itsintra-datacenter

TCPthroughputisbetween600and700Mbpslowerthanallotherproviders,anditsWAN

throughputisbetween100and200Mbpslowerthanallotherproviders.

Giventheresults,it’sstilldifficulttopicka“winning”iderisbest

overall,andit’r,theauthors

applytheresultsoftheirstudytospecifimple,ane-commercewebsitedeals

withmanytableoperations,

ontheresultsfromCloudCmp,e-commercesitesshoulduAWS;ithasthebestperformanceon

tablestorage(thepaperverifiesthisconclusionbyrunningatofe-commerce-specificbenchmarks

oneachcloud).Thepapergivessimilarexamplesforparallelscientificcomputingapplicationsand

TheauthorslisttheirresultsintermsofcloudsC

1

,C

2

,C

3

,andC

4

.Wegiveourowninferencesastowhich

providereachC

i

correspondsto.

1

1

CloudCmp:ComparingPublicCloudProviders

,,a,

KatrinaLaCurts,MIT(6.897Spring2011)

ificcomputingapplicationsareCPU-intensive,andthusshouldu

alow-endinstanceofMicrosoftAzure;Latency-nsitivewebsitesshoulduGoogleAppEngine,

whichhasthelowestwide-areanetworklatencydistribution.

Overall,CloudCmpisagoodfituresit

measuresarelikelytobecommonamongstallcloudproviders,notjustthefourstudiedinthis

paper,andthecomparisontoreal-worldapplicationsishelpfulingivingmoreinsightintotheir

r,thepaperlacksthe“stresstest”,

MapReduce)’salsonotclearhowfarinthefuturethemeasurementswillhold,or

whetheraverageconsumerswouldfindthisperformancecomparisonasufulasasimplefeature

comparisonbetweencloudproviders.

2