DynamicPower Optimization for Higher

DynamicPower Optimization for Higher

DynamicPower Optimization for Higher

ServerDensity Racks – A Baidu CaseStudy with Intel® DynamicPower Technology Executive Summary Intel’s Digital Enterprise Group partnered with Baidu.com conducted a proof of concept (POC) project using the Intel® Dynamic Power Node Manager Technology (Node Manager) to dynamically optimize server performance and power consumption to maximize the server densit...y of a rack. The POC engineers used Node Manager to identify optimal control points, which became the basis to set power optimization policies at the node level. A management console – Intel® Datacenter Manager (Datacenter Manager) was used to manage servers at rack- level to coordinate power and performance optimization between servers to ensure maximumserverdensityandperformyieldforgivenpowerenvelopefortherack. The POC was conducted in Q1’ 2008 at Intel-Baidu joint lab under a simulated Baiduproductionenvironment.ThePOCresultsillustratedthat: � Atasinglenodelevel, upto40W savings/systemwithoutperformance impactwhenaoptimalpowermanagementpolicyisapplied � Atracklevel,upto20% additionalcapacityincreasecouldbeachieved withinthesamerack-levelpowerenvelopewhenaggregated optimalpower managementpolicyisapplied � Comparingwithtoday’sdatacenteroperationatBaidu, byusingIntelNode Manager,therecouldbearack densityincrease20~40%improvement Moredetailedfindings fromthisPOCarepublishedinthispaper. Node Manager POC withBaidu Contents ExecutiveSummary................................................................................................1 BusinessOverview..................................................................................................3 Top BusinessIssues.......................................................................................................3 IntelTechnologyandSolution..................................................................................4 Intel®DynamicPowerNode Manager(Node Manager)........................................................4 Intel®DatacenterManager(DatacenterManager)..............................................................4 POC UseCases.......................................................................................................4 POC Architecture....................................................................................................5 POC Results...........................................................................................................6 Single Node Testing.......................................................................................................6 Rack-leveltesting..........................................................................................................9 Conclusion...........................................................................................................10 2 Node Manager POC withBaidu BusinessOverview BaiduisthebiggestsearchcompanyinChina,accounting forover60%ofsearchmarketshare in China.Itsmarketshare inthe Chinesedomesticmarkethasgrownsteadily,overthe pastyears.It alsoextendstointernationalmarketsand hasstartedbranchofficesinJapan,the U.S.and other countries. CurrentlyBaiduleasesracksatdatacentertoprovide theirservicesonthe interent.Itpaysbyrack andeachrackispowerlimited.Increasing serverdensityonarackisofgreatinterestforBaidu. 3 Figure 1: Baidu.com Search Portal TopBusiness Issues Datacenterhosting isaprimarycostforBaidubusinessoperation,whichisproportionaltothe numberofleased racks.Currently,the numberofserversoneachrackislowdue topower constraints.Baiduisinterested inincreasing serverdensityonarackwithinthe powerconstraints setbythe datacenterserviceprovider,inordertomaximize the performance oneachrack. Incurrentdatacenteroperationtoday,there isnodynamicpowermanagementtechnologywhich allowsBaidutooptimizepowerutilization.Asaresult,Baiduisfacing followingpowermanagement challenges: • Over-allocationofpower:Powerallocationtoserversdoesnotmatchactualserverpower consumption.Poweristypicallyallocated forworstcase scenariobased onserver nameplate.Staticallocationofpowerbudgetbasedonworstcasescenarioleadsto inefficienciesand doesnotmaximize use ofavailablepowercapacityand rackspace. • Under-populationofrackspace:Asadirectresultofthe over-allocationproblem,there isa lotofemptyspace onracks.WhenBaiduneedsmorecompute capacity,theyhave topay more foradditionalracks.Availabledatacenterspace isalimiting factortoBaidu’sbusiness growth. Node Manager POC withBaidu • Capacityplanning:There islackofeffective meanstooptimize the powerand performance dynamicallyatracklevel.Toimprove powerutilization,datacentersneedstotrackactual powerand cooling consumptionand dynamicallyadjustworkload andpowerdistributionfor optimalperformance atrackanddatacenterlevels. Intel Technologyand Solution Intel®DynamicPower Node Manager (Node Manager) Node Managerisanout-of-band (OOB)powermanagementpolicyengine thatisembedded inIntel serverchipset.ItworkswithBIOSand OSpowermanagement(OSPM)todynamicallyadjust platformpowertoachieve maximumperformance/poweratnode (server)level.Node Managerhas the following features: • DynamicPowerMonitoring:Measuresactualpowerconsumptionofaserverplatformwithin acceptableerrormarginof+/-10%.Node ManagergathersinformationfromPSMI instrumented powersupply,providesreal-timepowerconsumptiondata(pointintime,or average overaninterval),and reportsthroughIPMIinterface. • PlatformPowerCapping:Setsplatformpowertoatargeted powerbudgetwhile maintaining maximumperformance forthegivenpowerlevel.NodeManagerreceivespowerpolicyfrom anexternalmanagementconsole throughIPMIinterface and maintainspowerattargeted levelbydynamicallyadjustingCPUp-states. • PowerThreshold Alerting:NodeManagermonitorsplatformpoweragainsttargeted power budget.Whenthe targetpowerbudgetcannotbe maintained,Node Managersendsout alertstothe managementconsole Intel®Datacenter Manager (Datacenter Manager) Intel®DatacenterManagerisasoftware add-ontoNode Managertomonitorand controlpowerfor agroup ofservers.Intel®DatacenterManagerdependsonIntel®DynamicPowerNode Manager. Itisasoftware developmentkit(SDK)designed toplug-intosoftwaremanagementconsole products.Italsohasareference userinterface whichwasused inthisPOC asproxyfora managementsoftware product.KeyIntel®DatacenterManagerfeaturesare: • Group levelmonitoring ofpower • Log andqueryfortrend data • Grouppowerlimiting • Group levelpoweralertsand notifications • Supportofdistributedarchitectures(acrossmultiple racks) POCUse Cases InthisPOC we focused onusecasestotestNode Managerfeaturesatnode levelfirst.Abaseline testisneeded toidentifythe optimalcontrolpointsatthe node level.We thenused these optimal controlpointsasthebase forracklevelpolicydefinition. Asummaryofuse casesislistedbelow: 4 Node Manager POC withBaidu Use Case Title Description Get power consumption Using the Intel Node Manager features to dynamically gather on each server point in time power consumption from each server on the rack Estimate total power Estimate rack level power consumption by summing up node consumption of a rack level power consumption; display on, and notify console as appropriate. Optimize rack level At rack level, analyze the power consumption of each server, policy within a given overall power consumption, rack level power envelope, and power envelop and targeted performance goals (utilization, response time, query server workload queue length, etc.) as well as other factors important to Baidu to determine the optimal power distribution policy. Baidu will set the policy and optimization strategy based on their work load and priority. Set policy to servers Read the full DynamicPower Optimization for Higher.

Related Videos