Per-core Power Estimation and Power Aware Scheduling Strategies for CMPs

Abstract

The problem of accurately estimating the processor power consumption has generated significant interest among computer architects in the last decade. With the focus on green computing intensifying, increasing number of task management applications have become power aware in last few years. Hence, the need for a fast and accurate power model is greater than ever. In addition, today’s multi-core processors demand task schedulers to balance the performance requirements, power budget and thermal constraints. This thesis addresses this requirement by presenting a percore power model based upon performance monitoring counters and temperature data. PMC based power models provide a straightforward and fast way of analyzing the activity of processor’s underlying microarchitecture. The advantage of our model is that it is general enough to be ported and scaled across different platforms with ease, fast enough to be used online by task schedulers, and it requires no knowledge of individual applications. During this thesis work, we validated the model on three different (two- to eight-core) platforms. The model accurately estimates core power consumption, exhibiting 1.8%-4.8% per-suite median error on the NAS , SPEC OMP , and SPEC 2006 benchmarks (and 1.6%-4.4% overall).