Cache and asymmetry aware scheduling for big.LITTLE architecture

Abstract

Clustered asymmetric multi-cores group big and little cores to different shared cache clusters to further reduce the energy consumption when no big cores are used. However, when all cores are active, prior scheduling schemes for asymmetric multi-cores may no longer be effective with separate shared caches for big or little cores, due the cache sharing and interference effect. This paper investigates whether, for cluster symmetric multi-cores, the cache sharing effect is as important as core asymmetry for scheduling by comparing prior asymmetry-oriented scheduling and cache-oriented scheduling. Since neither cache nor asymmetry-oriented scheduling schemes works best for combinations of applications,this paper proposes a dynamic scheduling scheme based on a hill-climbing algorithm. The experimental results show that the scheduling scheme can improve the overall throughput by 0.9% and 3.5% compared to the prior asymmetry-oriented scheduling schemes.