Online Thread Assignment for Heterogeneous Multicore Systems

Abstract

As computing devices absorb more of our computing needs and the energy crisis continues, specialized hardware is being built with energy conservation in mind. In particular, processor manufacturers keep increasing the number of cores and are now moving to heterogeneous multi-core processors. In this paper we propose PATA (Performance-Asymmetric Thread Assignment) algorithm for heterogeneous core systems composed of ``big'' high-performance and ``small'' low-power cores. The PATA algorithm makes thread-to-core assignments decisions at runtime based on the threads' IPS (Instructions committed Per Second), periodically allocating the available big cores to threads with current highest IPS, in order to exploit the different threads' execution phases. We implemented and evaluated PATA in a real multi-core system running Linux, where core heterogeneity was emulated using different clock speed (dynamic frequency scaling). Our results show that PATA provides the best performance when compared with standard Linux scheduler and the state-of-art IPS-driven algorithm.