Power Optimization - a Reality Check

A flurry of recent work has examined the interaction between system design and power optimization, with the promise of energy savings. We find that the power consumed by current commodity hardware can be well-approximated by a model with two components: a large always-present constant power and a linear power-performance tradeoff. As a result, many proposed optimizations currently have limited value and the best possible approach is the so-called “race to sleep” model of computation. We survey several recently-released computer systems in support of this conclusion and briefly examine the most important targets for future work.