Thread Owned Block Cache: Managing Latency in Many-Core Architecture

Abstract

Shared last level cache is crucial to performance. However, multithread program model incurs serious contention in shared cache. In this paper, to reduce average cache access latency, we propose two schemes. First, an implicitly dynamic cache partitioning scheme, i.e. block agglutinating. The purpose is to isolate conflicting data blocks. Second, a novel hardware buffer, called thread owned block cache, i.e. TOB Cache. The purpose is to store conflicting data blocks. Extensive analysis of the proposed schemes with Splash2 benchmarks and Bioinformatics workloads is performed using a cycle accurate many-core simulator. Experimental results show that the proposed schemes make conflict miss rate of shared cache reduced by 40% compared to traditional shared cache. Compared with victim cache, average load latency of shared cache and primary data cache is reduced by about 26% and 12%, respectively; primary data cache miss penalties are reduced by about 14%, and IPC is improved by 17%.