ABSTRACT
Power Management 188 6.2 Principles of Coordinated Federated Management 190
6.2.1 High-Level Problem Formulation: Constrained Optimization 190
6.2.2 Understanding the System: Interaction between the Knobs and Metrics 192 6.2.2.1 Power Consumption as a Function of
Workload and Server Power Management 192
6.2.2.2 Application Performance Model 196 6.2.2.3 Server Temperature Model 197
6.2.3 Understanding the Systems: Temporal and Spatial Variance 198
6.2.4 High-Level Principles of Architecture Design 200 6.3 Cross-Layer Power Management Architectures 203
6.3.1 Coordinated Server-Level Power Management: Power Efficiency and Power Capping 203
6.3.2 Multilayer Power Management Solution 205 6.3.3 SLA Power Management 208 6.3.4 Unified Power and Cooling Management 209
6.4 Cross-Layer Power Management Solution Designs 210 6.4.1 Feedback Control and Case Studies 210
6.4.1.1 An Adaptive Power Efficiency Control Algorithm 210
6.4.1.2 An Adaptive Server-Capping Algorithm 212 6.4.1.3 Group Power Capper 216
6.4.2 Optimization and Case Studies 217 6.4.2.1 Optimization Problem for Power
Management 217 6.4.2.2 Constrained Optimization 221
6.4.3 Case Study: Evaluation of the Integrated Management Solutions 221 6.4.3.1 Utilization Trace-Based Simulation 221 6.4.3.2 Basic Results: Uncoordinated versus
Coordinated Control 224 6.5 Concluding Comments 226 Acknowledgment 227 References 227
6.1 INTRODUCTION Power delivery, electricity consumption, and heat management are becoming critical in data center environments that happen at multiple layers: in individual servers, across a collection of servers such as a blade enclosure or a server rack, at the entire data center level, or even across multiple geographically distributed data centers. In addition, at each of these layers, power control can be done at the hardware, the firmware, the operating system (OS), or the application level. A key challenge arises when these multiple controllers interfere with one another due to lack of adequate coordination. Such interference can compromise both the management objectives and the system performance. This chapter addresses the cross-layer active power management problem. We discuss the challenges and summarize the key principles to consider for coordinated power control. We then present a few examples that illustrate the problem formulation, the models, the architectures, the algorithms and policies, and the benefits of the cross-layer power management solutions.
