VSCP: A Cache Controlling Method for Improving Single Thread Performance in Multicore System

Single thread performance is still a matter of concern even for multicore systems. Making good advantage of on-chip cache memory is critical in improving single thread performance. However, as the distributed layout of on-chip cache, single-threaded applications only use limited cache capacity of the whole system, while leaving others idle. Besides, data sets of a thread, which present strong localities, may not be allocated with sufficient cache spaces due to user transparent characteristic of cache memory. To address these two critical issues, we proposed a new cache controlling mechanism, named Virtual Shared Cache Partitioning (VSCP), which enables single thread to enjoy the whole on-chip cache spaces, and also enables programmers to control cache space allocation. We evaluated VSCP on an Intel multicore system with a set of memory intensive applications selected from SPEC2006 and NAS; the experiment results showed that VSCP achieved an average speedup of 43% and reduced the last level cache miss rate by 53% on average.