Algorithmic Support for Power Management

点击翻译按钮获取中文摘要

Power is now widely recognized as a first class architectural design constraint at all levels of computing, from the level of processor chips to the level of data centers. There are at least three distinct common goals for power management: energy conservation, temperature management, and limiting maximum power. The most commonly used power management technique is speed scaling, which involves changing the speed of the processor. Essentially all currently produced general purpose processors may be run at multiple speeds, and all leading processor manufacturers produce associated software that manages power by scaling processor speed. However, speed scaling as a power management technique, has to date mostly been applied in a reactive fashion in that it is usually only invoked in response to abnormal conditions, and its application is largely decoupled from the scheduling policies in that they are working independently of each other. It is widely accepted that integrated proactive power management and scheduling policies will result in better power management than simple reactive strategies.The proposed research investigates algorithmic problems related to integrated power management and and scheduling. The resulting optimization problems have dual objectives, since one seeks to optimize both some quality-of-service measure of the schedule, and some power related criteria. In general these objectives are conflicting in that the more power that one uses, the better quality of service that can be provided. The algorithmic solutions to these problem then involve increasing power when the improvement in the scheduling objective justifies the increased cost in the power management objective. The goals of the research are threefold. The first goal is to develop algorithms that could form the basis for future power management software. The second goal is to build a toolkit of widely applicable algorithmic methods for power management problems. The third goal is help increase the computing community's ability to abstractly reason about power, energy and temperature.