Online Power and Performance Estimation for Dynamic Power Management

Power management is at the forefront of challenges facing next-generation computing systems. Multiple mechanisms for dynamic regulation of processor power can be used aggressively to address power and thermal issues in real time; however, to do so effectively requires knowledge of the power and performance impact of the power-regulation choices. The problem is complex since the impact of each mechanism is both systemand workload-dependent. In this paper, we present our methodology and models for online predictions driven by data from processor performance counters. They estimate power consumption and performance when multiple power-management features are used simultaneously. Our goal is to provide effective, efficient power and performance prediction capabilities, which allow OS schedulers and workload/power managers to gauge the power and performance impact of a state change before making it. We focus on near-term estimation over 10s of millisecond intervals, rather than long-term averages. This approach allows power managers to fine-tune the power state of the processor for scheduling changes and workload variations. It also offers fast response for controlling the demand on the power supply and the cooling subsystem. We implement our methodology in a software service running on a Pentium M system and evaluate the power and performance models using the SPEC CPU2000 benchmarks. We also demonstrate their use with two new dynamic power management applications: Power Saver that optimizes for saving power within a specified bound on performance loss and a Performance Maximizer that optimizes for best performance while meeting a specified power limit.