Applying statistical machine learning to multicore voltage & frequency scaling

Abstract

Dynamic Voltage/Frequency Scaling (DVFS) is a useful tool for improving system energy efficiency, especially in multi-core chips where energy is more of a limiting factor. Per-core DVFS, where cores can independently scale their voltages and frequencies, is particularly effective. We present a DVFS policy using machine learning, which learns the best frequency choices for a machine as a decision tree. Machine learning is used to predict the frequency which will minimize the expected energy per user-instruction (epui) or energy per (user-instruction)2 (epui2). While each core independently sets its frequency and voltage, a core is sensitive to other cores' frequency settings. Also, we examine the viability of using only partial training to train our policy, rather than full profiling for each program. We evaluate our policy on a 16-core machine running multiprogrammed, multithreaded benchmarks from the PARSEC benchmark suite against a baseline fixed frequency as well as a recently-proposed greedy policy. For 1ms DVFS intervals, our technique improves system epui2 by 14.4% over the baseline no-DVFS policy and 11.3% on average over the greedy policy.