Performance and power management for multi-core processors
MetadataShow full item record
This dissertation addresses the problem of power and performance management for various computing systems, from single voltage island multicore processors to power constrained extreme scale cloud systems. Balancing power and performance in modern computing systems is a complex optimization problem. This challenge is addressed by the statement of this thesis: Improving performance and power consumption in modern computing systems will require new techniques, and the body of control theories can provide the basis for such solutions. This thesis developed dynamic models for throughput and power that adjust well to workload variations. Those models are general and can be applied to various kinds of computing frameworks. Based on those models, we use feedback controllers for throughput regulation and power regulation. The controllers are based on integrators for variable gain designed for stabilizing the closed-loop system as well as for rapidly responding to changing workload in short time frames. The feedback control is robust with respect to model uncertainties and computing errors in the loop, and they exhibit fast convergence despite such errors. This thesis addresses the performance and power management through three main contributions: 1. Effective and efficient power & performance management techniques in a single voltage island multi-core processor. 2. Maximizing power efficiency under a power cap in a multi-core processor that is composed of several voltage islands. 3. A hierarchical power management technique to improve performance and energy efficiency under power budgets in a cloud system.