MECHANISMS FOR COORDINATED POWER MANAGEMENT WITH APPLICATION TO COOPERATIVE DISTRIBUTED SYSTEMS A Thesis Presented to The Academic Faculty by Ripal Nathuji In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the School of Electrical and Computer Engineering Georgia Institute of Technology August 2008 MECHANISMS FOR COORDINATED POWER MANAGEMENT WITH APPLICATION TO COOPERATIVE DISTRIBUTED SYSTEMS Approved by: Karsten Schwan, Advisor Henry Owen College of Computing School of Electrical and Computer Georgia Institute of Technology Engineering Georgia Institute of Technology Sudha Yalamanchili, Co-Advisor Gabriel Loh School of Electrical and Computer College of Computing Engineering Georgia Institute of Technology Georgia Institute of Technology Hsien-Hsin Sean Lee Vijay Madisetti School of Electrical and Computer School of Electrical and Computer Engineering Engineering Georgia Institute of Technology Georgia Institute of Technology Date Approved: April 21, 2008 To my parents for their years of support, my loving wife for being my best friend and my better half, and our daughter Elena. iii ACKNOWLEDGEMENTS There are are countless people who have been instrumental in helping me complete my Ph.D. by providing support and advice. I would like to take the opportunity to express my gratitude and appreciation for a few of them here. First, I would like to thank my adviser, Dr. Karsten Schwan. He was always available to discuss ideas and provide suggestions for how to develop my Ph.D. research. Most, if not all, of the ideas presented in this thesis came from time spent in his office bouncing ideas around, hours which proved to be invaluable. During my graduate years, I spent three summers with the Systems Technology Lab at Intel collaborating with the power management group. I’d like to thank some of the members of that team, including Ram Chary, Eugene Gorbatov, and Pradeep Sebestian for the great experiences, the hardware resources they always made available, and for the opportunity to learn from them while working hands on in an industry research lab. I would also like to thank my co-advisor Dr. Sudha Yalamanchili, as well as the other members of my committee including Dr. Hsien-Hsin Sean Lee, Dr. Henry Owen, Dr. Gabriel Loh, and Dr. Vijay Madisetti. They were a source of tough questions and useful thoughts that helped improve the contents of this document. I’ve been fortunate to work with, and sometimes just kill time with, some great students that have been around during the years. I thank them for the great discussions, constructive feedback, and sometimes random conversations. Finally, I’d like to thank the members of my family who have been there for me: My parents for their support and advice throughout the years. My wife for her love and friend- ship that have helped me endure the ups and downs of graduate school. Last, but not least, the four legged canine members of our household, Raja, Sonya, and Ravi, for magically sensing when I needed a break from work, and always reminding me to have fun in life. iv TABLE OF CONTENTS DEDICATION . ..................................... iii ACKNOWLEDGEMENTS................................ iv LISTOFTABLES .................................... ix LISTOFFIGURES.................................... x SUMMARY ......................................... xiii I INTRODUCTION.................................. 1 1.1 Thesis Statement and Contributions . .................. 4 1.2ThesisOrganization.............................. 5 II POWERMANAGEMENT:TRENDSANDOPPORTUNITIES........ 7 2.1ThePowerProblem.............................. 7 2.1.1 MobileTrends............................. 8 2.1.2 ServerTrends............................. 9 2.2MethodstoImprovePowerConsumption.................. 10 2.2.1 Hardware Management Capabilities . .............. 11 2.2.2 ResourceAllocationinTime..................... 12 2.2.3 ResourceAllocationinSpace..................... 12 2.2.4 Power Management Methods in Thesis Contributions . ...... 13 III EESI:DEVICEWINDOWSENABLEDOSSCHEDULING.......... 14 3.1Motivation:AggregatingDeviceAccesses.................. 15 3.2AProcessSchedulingApproach....................... 17 3.2.1 Scheduling for Access Bursting . .................. 17 3.2.2 Device Windows - Data Structures for Device Usage Correlation . 19 3.3EESISystemDesign.............................. 20 3.3.1 EESIwithDynamicState....................... 21 3.3.2 Real-TimeEESIwithDynamicState................ 22 3.3.3 UseofSourceAnnotationswithEESI................ 26 3.4ExperimentalEvaluation........................... 30 3.4.1 IntelSitsangEvaluationPlatform.................. 30 v 3.4.2 PowerMeasurementMethodology.................. 31 3.4.3 EESIResultswithDynamicState.................. 32 3.4.4 EESIResultswithStaticHints.................... 34 3.5RelatedWork.................................. 37 3.6Summary.................................... 38 IV POWER EFFICIENT MOBILE WORKLOADS WITH COMPATPM . 39 4.1EnergyEfficientDeploymentofMultimediaTasks............. 41 4.1.1 Multimedia Workloads in Cooperative Distributed Systems . 41 4.1.2 PowerManagementArchitecture................... 43 4.1.3 EvaluationInfrastructure....................... 44 4.2EnergyAnalysisofMultimediaApplications................ 45 4.3EvaluationofCompatPM........................... 52 4.3.1 CompatPM Attributes . ...................... 52 4.3.2 Results................................. 53 4.4RelatedWork.................................. 55 4.5Summary.................................... 56 V HETEROGENEITY AWARE DATACENTER POWER MANAGEMENT . 58 5.1Motivation................................... 61 5.1.1 DatacenterCompositionandExploitingHeterogeneity....... 61 5.1.2 BenefitsofHeterogeneity-AwareManagement........... 62 5.2ScalableEnterpriseandDatacenterManagement.............. 64 5.3Methodology.................................. 67 5.3.1 PlatformHardware.......................... 67 5.3.2 ApplicationModel........................... 69 5.4WorkloadBehaviorEstimation........................ 71 5.5ManagementPolicies............................. 77 5.5.1 HALMAllocationPolicy....................... 77 5.5.2 HALMPowerBudgetingPolicy................... 78 5.6ExperimentalEvaluation........................... 78 5.6.1 Increasing Power Efficiency ...................... 78 vi 5.6.2 MaximizingPerformanceUnderPowerBudgets.......... 80 5.7RelatedWork.................................. 82 5.8Summary.................................... 84 VI VIRTUALPOWER: POWER MANAGEMENT IN VIRTUALIZED SYSTEMS 85 6.1ManagingVirtualizedSystems........................ 87 6.1.1 Scalable Enterprise and Virtualized Resources . .......... 87 6.1.2 LeveragingGuestVMPolicies.................... 89 6.1.3 LimitationsofHardwareManagement................ 90 6.2VirtualPowerArchitecture.......................... 91 6.2.1 VPMStates.............................. 92 6.2.2 VPMChannels............................. 93 6.2.3 VPMMechanisms........................... 94 6.3EvaluationMethodology............................ 96 6.3.1 ExperimentalSetup.......................... 96 6.3.2 GuestApplicationsandPolicies................... 97 6.4 Viability of Soft Scaling . .......................... 99 6.5PolicyBasedCoordination.......................... 102 6.5.1 PM-LPolicies:PlatformManagement................ 103 6.5.2 PM-GPolicies:GlobalCoordination................. 112 6.6RelatedWork.................................. 116 6.7Summary.................................... 118 VII VIRTUAL MACHINE-AWARE POWER BUDGETING WITH VPM TOKENS 119 7.1Motivation................................... 121 7.1.1 VM-CentricBudgeting........................ 122 7.1.2 Application-awareManagement................... 123 7.1.3 CompensatingBudgetedApplications................ 124 7.1.4 Budgeting Heterogeneous Resources . .............. 126 7.2PowerBudgetingArchitecture........................ 127 7.2.1 SystemManagers........................... 128 7.2.2 VPMTokens.............................. 130 vii 7.3 Token Currency Exchange . .......................... 131 7.3.1 BudgetTokensandPower...................... 131 7.3.2 BudgetTokensandVMManagement................ 132 7.4ExperimentalMethodology.......................... 133 7.4.1 PlatformsandPowerMeasurements................. 133 7.4.2 ExperimentalApplications...................... 134 7.5ImplementationandEvaluation....................... 136 7.5.1 ManagerImplementations....................... 136 7.5.2 ExperimentalResults......................... 140 7.6RelatedWork.................................. 145 7.7Summary.................................... 146 VIIICONCLUSIONSANDFUTUREWORK..................... 147 REFERENCES....................................... 150 VITA............................................ 160 viii LIST OF TABLES 1 System Coordination in Thesis Contributions. .............. 5 2 SummarizedUsageofPowerManagementMethods.............. 13 3 DeviceWindowParameterDefinitions...................... 20 4 DWCS Precedence Rules. .......................... 22 5 Modified DWCS Precedence Rules. ...................... 24 6 EESIExperimentalApplicationswithDefaultLinuxScheduler........ 33 7 EESIExperimentalScenarioswithDefaultLinuxScheduler.......... 33 8 EESIExperimentalApplicationswithDWCSScheduler............ 34 9 EESIExperimentalScenarioswithDWCSScheduler.............. 34 10 EESIExperimentalApplicationswithSourceAnnotationHints........ 36 11 PXA255OperatingPoints............................ 44 12 Sitsang Platform Power Consumption Overview. .............. 45 13 Sitsang CPU Power Consumption Overview.