DOEE: dynamic optimization framework for better energy efficiency.

The growing adoption of mobile devices powered by batteries along with the high power costs in datacenters raise the need for energy efficient computing. Dynamic Voltage and Frequency Scaling is often used by the operating system to balance power-performance. However, optimizing for energy-efficiency faces multiple challenges such as when dealing with non-steady state workloads. In this work we develop DOEE - a novel method that optimizes certain processor features for energy efficiency using user-supplied metrics. The optimization is dynamic, taking into account the runtime characteristics of the workload and the platform. The method instruments monitoring code to search for per-program-phase optimal feature-configurations that ultimately improve system energy efficiency. We demonstrate the framework using the LLVM compiler when tuning the Turbo Boost feature on modern Intel Core processors. Our implementation improves energy efficiency by up to 23% on SPEC CPU2006 benchmarks, outperforming the energy-efficient firmware algorithm. This framework paves the way for auto-tuning additional CPU features.