|
Abstract:
|
Power and thermal effects have emerged as serious problems for computing systems by limiting performance, degrading reliability, and imposing a high cost in energy resources. A fundamental problem with power and thermal management is the difficulty of reducing power and heat output without sacrificing performance, which creates a complex web of inter-related constraints and requirements. Meeting these multiple, potentially conflicting objectives simultaneously is a difficult challenge, exacerbated by shifting environmental conditions and variable workloads, yet essential for future generations of high-performance systems. We propose a comprehensive, goal-oriented management framework that sorts priorities and balances conflicting goals, named PET for performance, power, energy, and temperature. The approach provides a level of indirection between macro objectives, such as reducing operating cost or increasing performance, and micro directives, including voltage and frequency settings and other power-management choices. Goal-driven decisions reflect relevant run-time conditions, rather than pre-defined policies, and a concise specification for desired outcome provides an opportunity to customize operation to conserve or spend power resources as situations warrant, delivering performance on demand. We demonstrate the feasibility and benefit of the PET approach with a prototype implementation developed in software, executing on an instrumented Pentium M system. First, we present a detailed characterization of the system response to power management mechanisms to identify the timescales and magnitude of expected response to management decisions. Second, we illustrate PET operation with realistic workloads and usage scenarios, demonstrating that the prototype achieves the desired ranges of operation with dynamic run-time control. |
