Research on Cache Partitioning and Adaptive Replacement Policy for CPU-GPU Heterogeneous Processors

Abstract

Heterogeneous multicore processors integrate CPU and GPU cores which use a common last-level cache (LLC). However, it puts more pressure on cache management algorithm. Since GPU cores have higher number of threads, most of the LLC space will be dominated by GPU application, leaving limited space for CPU application. Because of this reason, it seriously affects the overall system performance. Aiming at the unfair utilization of GPU and CPU cores for shared cache resource, this paper mainly proposes a novel cache management method: cache partition combined with the adaptive replacement policy. We first split the cache capacity to adjust the ratio of CPU and GPU cores for shared LLC resource and then use adaptive replacement policies for CPU and GPU applications to access LLC. Experimental results show that our scheme can make GPU applications in the case of minimal loss of performance, improve the performance of CPU applications by 16% on average (up to 33%), the overall performance improved by 6 %( up to 19%).