Improving Energy Efficiency by Phase-Grained Migration for Asymmetric Multicore

Asymmetric multicore architecture that consists of multiple cores with different power and performance features has the potential for improving energy efficiency. Migrating threads with different characteristics to appropriate cores can reap the full benefits of the asymmetry to improve energy efficiency. A key challenge in the use of asymmetric processors is to determine the reassignment of threads to asymmetric processor cores according to changing program phases within threads. In this paper we propose a scheme to improve the energy efficiency by phase-grained thread migration for asymmetric multicore. Our scheme mainly includes three new techniques. Firstly, we employ a linear estimation model with high accuracy to directly estimate performance per watt of the current thread phase on each core type online. These information of performance per watt decides the affinities of threads to cores. Secondly, according to the affinities of threads to cores and the overhead of migration, an Energy Efficiency Optimization Model (EEOM) is proposed to clarify the problem of thread-to-core assignment. Lastly, we present a scheduling algorithm based on the Kuhn-Munkres (KM) algorithm to find an optimal solution of the EEOM. We compare our proposed scheme with the state-of-the-art static assignment, and the result shows that our scheme can reach up to 9.5% improvement of performance per watt.