ENERGY-EFFICIENT DISTRIBUTED COMPUTING SYSTEMS -- CONTENTS -- PREFACE -- ACKNOWLEDGMENTS -- CONTRIBUTORS -- 1 POWER ALLOCATION AND TASK SCHEDULING ON MULTIPROCESSOR COMPUTERS WITH ENERGY AND TIME CONSTRAINTS -- 1.1 Introduction -- 1.1.1 Energy Consumption -- 1.1.2 Power Reduction -- 1.1.3 Dynamic Power Management -- 1.1.4 Task Scheduling with Energy and Time Constraints -- 1.1.5 Chapter Outline -- 1.2 Preliminaries -- 1.2.1 Power Consumption Model -- 1.2.2 Problem Definitions -- 1.2.3 Task Models -- 1.2.4 Processor Models -- 1.2.5 Scheduling Models -- 1.2.6 Problem Decomposition -- 1.2.7 Types of Algorithms -- 1.3 Problem Analysis -- 1.3.1 Schedule Length Minimization -- 1.3.1.1 Uniprocessor computers -- 1.3.1.2 Multiprocessor computers -- 1.3.2 Energy Consumption Minimization -- 1.3.2.1 Uniprocessor computers -- 1.3.2.2 Multiprocessor computers -- 1.3.3 Strong NP-Hardness -- 1.3.4 Lower Bounds -- 1.3.5 Energy-Delay Trade-off -- 1.4 Pre-Power-Determination Algorithms -- 1.4.1 Overview -- 1.4.2 Performance Measures -- 1.4.3 Equal-Time Algorithms and Analysis -- 1.4.3.1 Schedule length minimization -- 1.4.3.2 Energy consumption minimization -- 1.4.4 Equal-Energy Algorithms and Analysis -- 1.4.4.1 Schedule length minimization -- 1.4.4.2 Energy consumption minimization -- 1.4.5 Equal-Speed Algorithms and Analysis -- 1.4.5.1 Schedule length minimization -- 1.4.5.2 Energy consumption minimization -- 1.4.6 Numerical Data -- 1.4.7 Simulation Results -- 1.5 Post-Power-Determination Algorithms -- 1.5.1 Overview -- 1.5.2 Analysis of List Scheduling Algorithms -- 1.5.2.1 Analysis of algorithm LS -- 1.5.2.2 Analysis of algorithm LRF -- 1.5.3 Application to Schedule Length Minimization -- 1.5.4 Application to Energy Consumption Minimization -- 1.5.5 Numerical Data -- 1.5.6 Simulation Results -- 1.6 Summary and Further Research -- References.

2 POWER-AWARE HIGH PERFORMANCE COMPUTING -- 2.1 Introduction -- 2.2 Background -- 2.2.1 Current Hardware Technology and Power Consumption -- 2.2.1.1 Processor power -- 2.2.1.2 Memory subsystem power -- 2.2.2 Performance -- 2.2.3 Energy Efficiency -- 2.3 Related Work -- 2.3.1 Power Profiling -- 2.3.1.1 Simulator-based power estimation -- 2.3.1.2 Direct measurements -- 2.3.1.3 Event-based estimation -- 2.3.2 Performance Scalability on Power-Aware Systems -- 2.3.3 Adaptive Power Allocation for Energy-Efficient Computing -- 2.4 PowerPack: Fine-Grain Energy Profiling of HPC Applications -- 2.4.1 Design and Implementation of PowerPack -- 2.4.1.1 Overview -- 2.4.1.2 Fine-grain systematic power measurement -- 2.4.1.3 Automatic power profiling and code synchronization -- 2.4.2 Power Profiles of HPC Applications and Systems -- 2.4.2.1 Power distribution over components -- 2.4.2.2 Power dynamics of applications -- 2.4.2.3 Power bounds on HPC systems -- 2.4.2.4 Power versus dynamic voltage and frequency scaling -- 2.5 Power-Aware Speedup Model -- 2.5.1 Power-Aware Speedup -- 2.5.1.1 Sequential execution time for a single workload T1(w, f ) -- 2.5.1.2 Sequential execution time for an ON-chip/OFF-chip workload -- 2.5.1.3 Parallel execution time on N processors for an ON-/OFF-chip workload with DOP = i -- 2.5.1.4 Power-aware speedup for DOP and ON-/OFF-chip workloads -- 2.5.2 Model Parametrization and Validation -- 2.5.2.1 Coarse-grain parametrization and validation -- 2.5.2.2 Fine-grain parametrization and validation -- 2.6 Model Usages -- 2.6.1 Identification of Optimal System Configurations -- 2.6.2 PAS-Directed Energy-Driven Runtime Frequency Scaling -- 2.7 Conclusion -- References -- 3 ENERGY EFFICIENCY IN HPC SYSTEMS -- 3.1 Introduction -- 3.2 Background and Related Work -- 3.2.1 CPU Power Management -- 3.2.1.1 OS-level CPU power management.

3.2.1.2 Workload-level CPU power management -- 3.2.1.3 Cluster-level CPU power management -- 3.2.2 Component-Based Power Management -- 3.2.2.1 Memory subsystem -- 3.2.2.2 Storage subsystem -- 3.2.3 Thermal-Conscious Power Management -- 3.2.4 Power Management in Virtualized Datacenters -- 3.3 Proactive, Component-Based Power Management -- 3.3.1 Job Allocation Policies -- 3.3.2 Workload Profiling -- 3.4 Quantifying Energy Saving Possibilities -- 3.4.1 Methodology -- 3.4.2 Component-Level Power Requirements -- 3.4.3 Energy Savings -- 3.5 Evaluation of the Proposed Strategies -- 3.5.1 Methodology -- 3.5.2 Workloads -- 3.5.3 Metrics -- 3.6 Results -- 3.7 Concluding Remarks -- 3.8 Summary -- References -- 4 A STOCHASTIC FRAMEWORK FOR HIERARCHICAL SYSTEM-LEVEL POWER MANAGEMENT -- 4.1 Introduction -- 4.2 Related Work -- 4.3 A Hierarchical DPM Architecture -- 4.4 Modeling -- 4.4.1 Model of the Application Pool -- 4.4.2 Model of the Service Flow Control -- 4.4.3 Model of the Simulated Service Provider -- 4.4.4 Modeling Dependencies between SPs -- 4.5 Policy Optimization -- 4.5.1 Mathematical Formulation -- 4.5.2 Optimal Time-Out Policy for Local Power Manager -- 4.6 Experimental Results -- 4.7 Conclusion -- References -- 5 ENERGY-EFFICIENT RESERVATION INFRASTRUCTURE FOR GRIDS, CLOUDS, AND NETWORKS -- 5.1 Introduction -- 5.2 Related Works -- 5.2.1 Server and Data Center Power Management -- 5.2.2 Node Optimizations -- 5.2.3 Virtualization to Improve Energy Efficiency -- 5.2.4 Energy Awareness in Wired Networking Equipment -- 5.2.5 Synthesis -- 5.3 ERIDIS: Energy-Efficient Reservation Infrastructure for Large-Scale Distributed Systems -- 5.3.1 ERIDIS Architecture -- 5.3.2 Management of the Resource Reservations -- 5.3.3 Resource Management and On/Off Algorithms -- 5.3.4 Energy-Consumption Estimates -- 5.3.5 Prediction Algorithms.

5.4 EARI: Energy-Aware Reservation Infrastructure for Data Centers and Grids -- 5.4.1 EARI's Architecture -- 5.4.2 Validation of EARI on Experimental Grid Traces -- 5.5 GOC: Green Open Cloud -- 5.5.1 GOC's Resource Manager Architecture -- 5.5.2 Validation of the GOC Framework -- 5.6 HERMES: High Level Energy-Aware Model for Bandwidth Reservation in End-To-End Networks -- 5.6.1 HERMES' Architecture -- 5.6.2 The Reservation Process of HERMES -- 5.6.3 Discussion -- 5.7 Summary -- References -- 6 ENERGY-EFFICIENT JOB PLACEMENT ON CLUSTERS, GRIDS, AND CLOUDS -- 6.1 Problem and Motivation -- 6.1.1 Context -- 6.1.2 Chapter Roadmap -- 6.2 Energy-Aware Infrastructures -- 6.2.1 Buildings -- 6.2.2 Context-Aware Buildings -- 6.2.3 Cooling -- 6.3 Current Resource Management Practices -- 6.3.1 Widely Used Resource Management Systems -- 6.3.2 Job Requirement Description -- 6.4 Scientific and Technical Challenges -- 6.4.1 Theoretical Difficulties -- 6.4.2 Technical Difficulties -- 6.4.3 Controlling and Tuning Jobs -- 6.5 Energy-Aware Job Placement Algorithms -- 6.5.1 State of the Art -- 6.5.2 Detailing One Approach -- 6.6 Discussion -- 6.6.1 Open Issues and Opportunities -- 6.6.2 Obstacles for Adoption in Production -- 6.7 Conclusion -- References -- 7 COMPARISON AND ANALYSIS OF GREEDY ENERGY-EFFICIENT SCHEDULING ALGORITHMS FOR COMPUTATIONAL GRIDS -- 7.1 Introduction -- 7.2 Problem Formulation -- 7.2.1 The System Model -- 7.2.1.1 PEs -- 7.2.1.2 DVS -- 7.2.1.3 Tasks -- 7.2.1.4 Preliminaries -- 7.2.2 Formulating the Energy-Makespan Minimization Problem -- 7.3 Proposed Algorithms -- 7.3.1 Greedy Heuristics -- 7.3.1.1 Greedy heuristic scheduling algorithm -- 7.3.1.2 Greedy-min -- 7.3.1.3 Greedy-deadline -- 7.3.1.4 Greedy-max -- 7.3.1.5 MaxMin -- 7.3.1.6 ObFun -- 7.3.1.7 MinMin StdDev -- 7.3.1.8 MinMax StdDev -- 7.4 Simulations, Results, and Discussion -- 7.4.1 Workload.

7.4.2 Comparative Results -- 7.4.2.1 Small-size problems -- 7.4.2.2 Large-size problems -- 7.5 Related Works -- 7.6 Conclusion -- References -- 8 TOWARD ENERGY-AWARE SCHEDULING USING MACHINE LEARNING -- 8.1 Introduction -- 8.1.1 Energetic Impact of the Cloud -- 8.1.2 An Intelligent Way to Manage Data Centers -- 8.1.3 Current Autonomic Computing Techniques -- 8.1.4 Power-Aware Autonomic Computing -- 8.1.5 State of the Art and Case Study -- 8.2 Intelligent Self-Management -- 8.2.1 Classical AI Approaches -- 8.2.1.1 Heuristic algorithms -- 8.2.1.2 AI planning -- 8.2.1.3 Semantic techniques -- 8.2.1.4 Expert systems and genetic algorithms -- 8.2.2 Machine Learning Approaches -- 8.2.2.1 Instance-based learning -- 8.2.2.2 Reinforcement learning -- 8.2.2.3 Feature and example selection -- 8.3 Introducing Power-Aware Approaches -- 8.3.1 Use of Virtualization -- 8.3.2 Turning On and Off Machines -- 8.3.3 Dynamic Voltage and Frequency Scaling -- 8.3.4 Hybrid Nodes and Data Centers -- 8.4 Experiences of Applying ML on Power-Aware Self-Management -- 8.4.1 Case Study Approach -- 8.4.2 Scheduling and Power Trade-Off -- 8.4.3 Experimenting with Power-Aware Techniques -- 8.4.4 Applying Machine Learning -- 8.4.5 Conclusions from the Experiments -- 8.5 Conclusions on Intelligent Power-Aware Self-Management -- References -- 9 ENERGY EFFICIENCY METRICS FOR DATA CENTERS -- 9.1 Introduction -- 9.1.1 Background -- 9.1.2 Data Center Energy Use -- 9.1.3 Data Center Characteristics -- 9.1.3.1 Electric power -- 9.1.3.2 Heat removal -- 9.1.4 Energy Efficiency -- 9.2 Fundamentals of Metrics -- 9.2.1 Demand and Constraints on Data Center Operators -- 9.2.2 Metrics -- 9.2.2.1 Criteria for good metrics -- 9.2.2.2 Methodology -- 9.2.2.3 Stability of metrics -- 9.3 Data Center Energy Efficiency -- 9.3.1 Holistic IT Efficiency Metrics -- 9.3.1.1 Fixed versus proportional overheads.

9.3.1.2 Power versus energy.