Machine Learning-Based Energy Optimization for Parallel Program Execution on Multicore Chips

Abstract

Energy is increasingly becoming the major constraint in designing multicore chips. Power and performance are the main components of energy and are inversely correlated. In this paper, we study the energy optimization of multicore chips that process parallel workloads using either power or performance optimization. To do so, we propose novel machine learning-based global and dynamic power management controller. The controller is used either to maximize performance within a fixed power budget or to minimize the consumed power to achieve the same baseline performance. The controller is also scalable, as it does not incur significant overhead as the number of cores or demands increases. The technique was evaluated using the PARSEC benchmark suite on a full-system simulator. The experimental results show that our global power controller outperforms, in terms of the EDP metric, the non-DVFS baseline by 28 and 35.5%, when optimized for performance and power, respectively. This suggests that optimizing power is more related to energy efficiency than optimizing performance.