Investigation of Shared L2 Cache on Many-Core Processors

Abstract:
Several research works point out to share L2 cache for two or more cores. Their main goals are to improve the data access through shared addresses. However, it is hard to define the best organization and sharing model, since shared cache can improve the performance of sharing data, but it can also increase the number of conflict and misses. In this way, shared cache needs more study and evaluation considering different workloads. Therefore, the objective of this paper is to investigate the impact of multiple processing cores sharing different organizations of L2 cache. Based on these organizations, a set of scientific parallel applications is evaluated in order to tune the best cache organization on a simulated 32-core chip. Results point out to the increase of the final performance varying L2 cache organization with the increase of the L2 line size, which achieves better results than cache size. Moreover, applications with memory random access increase its performance just varying the L2 cache sharing.