FullSystemPowerAnalysisandModeling forServerEnvironments D.Economou, - - PowerPoint PPT Presentation

▶
full system power analysis and modeling for server
SMART_READER_LITE
LIVE PREVIEW

FullSystemPowerAnalysisandModeling forServerEnvironments D.Economou, - - PowerPoint PPT Presentation

Jan 23, 2024 9 likes •282 views

FullSystemPowerAnalysisandModeling forServerEnvironments D.Economou, ,C.Kozyrakis,P.Ranganathan

slide-1
SLIDE 1

FullSystemPowerAnalysisandModeling forServerEnvironments

D.Economou,,C.Kozyrakis,P.Ranganathan

  • WorkshoponModeling,Benchmarking,andSimulation(MoBS)
slide-2
SLIDE 2
  • Motivation
  • Costsofpowerandcooling

Electricitynow~50%ofdatacentercosts( !,4/06) Datacentercoolingconsumes~1WperWconsumedbysystem

  • Powerdensityandcompaction
  • Thermalfailures

10Ctemperatureincrease→ 50%reliabilitydecrease

  • Environmentalissues

EnergyStar EnterpriseServerandDataCenterEfficiencyInitiative,2006

slide-3
SLIDE 3
  • Goals:PrerequisitestoOptimizingPower
  • Understandserverpower

Acrossdifferenttypesofsystems Componentbreakdowns Temporalvariation Withinandbetweenworkloads

  • Developmodelforserverpower

Fast,onlinemodeldeployableinadatacenterscheduler Zerohardwarecosttotheenduser Input:accessibleOSmetrics;Output:“goodenough”(within510%) estimateofpower

slide-4
SLIDE 4
  • Outline
  • Motivation
  • Experimentalsetup
  • Powercharacterization
  • Powermodeling
  • Futurework
  • Conclusions
slide-5
SLIDE 5
  • TestMachines
  • bladeserver

Lowpowerprocessorstates

  • Itaniumserver

Zeropowersavingtechnologyinprocessors Resourcesimbalancedinfavorofprocessors

10/100Ethernet 10/100Ethernet

  • 1HDD,36GB,3.5”

1HDD,40GB,2.5”

  • 1GBDDR

512MBSDRAM

  • 4*Itanium2,1.5GHz

1*AMDTurion,2.2GHz

  • !"
slide-6
SLIDE 6

#

MeasurementInfrastructure

slide-7
SLIDE 7

$

MeasurementInfrastructure

  • SystemUnderTest:BladeorItaniumserver
  • Runs% &' +lowoverhead( & ) (e.g.sar,caliper)at1

sample/sec

slide-8
SLIDE 8

*

MeasurementInfrastructure

Insertmeasurementbetweenmachineandwalltomeasureoverallpower

  • Bladeserver:1sample/sec
  • Itaniumserver:Currently20sample/sec
slide-9
SLIDE 9

+

MeasurementInfrastructure

  • Wecutintoandinstrumentedtheindividual"! oftheservers,tocapture

componentlevelDCpower(~20samples/sec)

  • ThisisNOTrequiredforourmodel
slide-10
SLIDE 10

,-

MeasurementInfrastructure

  • PC:synchronizesmeasurements,collectsdata

Performancemetricsfromsystemundertest

  • OverallpowerfromACpowermeter
  • ComponentpowerfromADC
slide-11
SLIDE 11

,,

  • !

" "

  • "
  • #
  • $
  • #
  • #%

& '

  • PowerCharacterization
  • AverageDCpowerofcomponents
  • Benchmarks:!###$#"%

!!

slide-12
SLIDE 12

,

  • !

" "

  • "
  • #
  • $
  • #
  • #%

& '

  • PowerCharacterization
  • &',,,and( components
  • Nonnegligiblecontributorstopower
  • Smallvariationinaveragepowerconsumption(occasional

spikes)

slide-13
SLIDE 13

,

  • !

" "

  • "
  • #
  • $
  • #
  • #%

& '

  • PowerCharacterization
  • Blade( isthesinglelargestconsumerofpower,although

isclosebehind

  • Highvariationinprocessorpowerconsumptionshowsthatbladeis
  • ptimizedforpower
slide-14
SLIDE 14

,

  • !

" "

  • "
  • #
  • $
  • #
  • #%

& '

  • PowerCharacterization
  • 100Wwhen!))
  • Notmuchvariation(30%)betweenidleandmaxpowerinItanium
  • Sothe4processorsdominate
  • Highvariationinmemory,percentagewise
slide-15
SLIDE 15

,

PowerCharacterizationConclusions

  • Conventionalwisdom

AfterCPU,memoryisthenextbottleneck LotsofvariationinCPUpowerifchipisoptimizedforpower;otherwise runsnear100%atalltimes

  • Moresurprising

Theassorted“misc”components– thearcanecircuitsondifferentpower planes– reallymatter(~20%ofbladepower).Optimizingthesemaybe worthwhile Diskcontributionisrelativelysmall EnormousidlepowerontheItaniumsystem

slide-16
SLIDE 16

,#

PowerModeling

  • Goal:Developanonlinemodelforuseindatacenterschedulers
  • Modelrequirements

Fullsystem Nonintrusive;easyforenduser Fastenoughforonlineuse Reasonablyaccurate(within510%) Inexpensive Generic(applicabletodifferenttypesofsystems)

slide-17
SLIDE 17

,$

PowerModeling:PastApproaches

  • Simulationbaseddetailedmodels

Inexpensive,arbitrarilyaccurate Notfullsystem Tailoredspecificallytoparticularsystems&components

  • Directhardwaremeasurements

Accurate,fast,easy Expensive(especiallyovermanymachines)

  • TheMantisQuestion

Canhighlevelcombinedmetricsgiveagoodapproximation?

slide-18
SLIDE 18

,*

PowerModeling

  • Run calibrationscheme

(possiblyatvendor)

*:performancemetrics,AC powermeasurements Workloadsthatstressindividual components:CPU,memory,disk, network

  • Fitmodelparameterstocalibration

data

Linearmodelforsimplicity

  • Usemodeltopredictpower

Inputs:performancemetrics(asfrom sar orcaliper)ateachpointintime Output:estimationofACpowerat eachpointintime

slide-19
SLIDE 19

,+

Calibration

  • Stresseachsystemcomponentinisolationtodevelopamodel
  • Used+ program(J.Moore,2005)tostressCPU,memory,

disk,networkatvaryingdegreesofutilization

Coulduseanyprogramthatcanselectivelystresscomponents , can’talwaysstresseachcomponenttotheabsolutemaximum

  • +./(!(!.

." ,+".!%01(.(! '"!+ /.2+/3%+

slide-20
SLIDE 20
  • ModelCreation
  • GOAL:Predictinstantaneouspowerwithin10%usingasimple,

fastmodel

Inputs:OSlevelutilizationmetrics+ACpowerforcalibrationsuite Output:Anequationwhichrelatespowertothesemetrics

  • INPUT:Utilizationmetrics

( =CPUutilization(%) =Offchipmemoryaccesscount ' =HarddiskI/Orate =NetworkI/Orate

  • OUTPUT:Forlinearmodel,anequationofform
  • '
  • (
  • #
  • &
  • 4

5

  • ,

, , , ,

* * * * + + + + =

slide-21
SLIDE 21

,

ModelInputs

  • Inputisamatrix6,e.g.:
  • Andavectore.g.:

... 1 1 1

2 , 2 , 2 , 2 , 1 , 1 , 1 , 1 , , , , , = = = = = = = = = = = =

  • '
  • (
  • '
  • (
  • '
  • (
  • '
  • (
  • !

...

2 , 1 , , = = =

slide-22
SLIDE 22
  • ModelCreation
  • !:avectorofweightsforeachutilizationmetric
  • #
  • /$(:minimizeabsoluteerrorofmodelsoverallcalibrationprograms
  • 6
  • =

) ( min

1 − = + −

∑

  • 7
  • ,

, , −

= ε

slide-23
SLIDE 23
  • ModelsDeveloped

0.0 0.00405 4.05*107 0.1108 635.62 Itanium 3.1*108 0.00281 4.47*108 0.236 14.45 Blade E(net) D(disk) C(mem) B(cpu) A(const)

  • '
  • (
  • #
  • &
  • 4

5

  • ,

, , , ,

* * * * + + + + =

Powerpredictionequation:

slide-24
SLIDE 24
  • Evaluation
  • (
  • #
  • $
  • #
  • )
  • *
  • !

" " )

  • *
  • "

)

  • *
  • )
  • *
  • #$

# )*!"" )*" )* )*

Mean%Error 90th PercentileAbsoluteError

slide-25
SLIDE 25
  • #
  • $
  • #
  • )
  • *
  • !

" " )

  • *
  • "

)

  • *
  • )
  • *
  • (
  • Evaluation

#$ # )*!"" )*" )* )*

Mean%Error 90th PercentileAbsoluteError

Genericmodelworks(within10%)on2verydifferentsystemsoveravariedsetof benchmarks

slide-26
SLIDE 26

#

ApplicationsandFutureWork

  • Improvingmodels

Componentlevelmodelingandvalidation Exploringnonlinearmodels Adding/replacingCPUutilization%withagenericmeasurementof ILP

  • Datacenterresourceprovisioning

Estimatepowercostsatdifferentgranularities(server,enclosure,rack…) Powerawareschedulingandmapping

  • Datacenterthermaloptimizations

ReplaceexpensiveexternalthermalsensorswithMantisestimates Generatedatacenterthermalmap

  • Fancontrol

Dynamicallysetfanspeedinresponsetoestimatedpower Withcomponentlevelmodels,turnonfansaimedathighpower components

slide-27
SLIDE 27

$

Conclusions

  • Goals:

Understandserverpowerconsumption Developpowermodelthatcanbeusedonlineindatacenters

  • Understandingserverpower

Quantitativecomponent/temporalpowerbreakdown Confirmingconventionalwisdom:CPUisbiggestconsumer,memoryis next Needcooperationofsoftwareforlowpower “Misc”componentisworthpayingattentionto

  • Developingapowermodel

Highlevelmetricsgiveareasonableapproximationofpower

  • Futurework

Improvemodel(ILPmetrics,nonlinearmodels…) Usemodelinadatacenterscheduler