Improving Performance in Sub-Block Caches with Optimized Replacement Policies

Abstract

Recent advances in computer processor design have led to the introduction of sub-blocking to cache architectures. Sub-block caches reduce the tag area and power overhead in caches without reducing the effective cache size by using fewer tags to index the full data RAM array. In spite of achieving reduced area and power overhead, sub-block caches suffer performance degradation due to cache trashing. This occurs when a wider cache line (super-block), made up of multiple valid cache lines (sub-blocks), is replaced or evicted when only a sub-block is to be allocated into the wider super-block. To address this problem, we propose cache replacement policies as they relate specifically to sub-block caches. We propose new replacement policies that are tuned for sub-block caches by adding more intelligence based on the valid state of individual sub-blocks of a super-block. We also investigate the effect of using a few level-0 registers to bypass a few level-1 cache pipe stages on sub-block cache performance. To evaluate the performance improvement offered by our proposed replacement policies and the use of level-0 registers, we developed a sub-block cache simulator based on the Simplescalar toolset for single-core evaluations and the Sniper Simulator for multicore evaluations. We show that, with minimal architectural updates to existing conventional cache replacement policies, we are able to improve level-1 cache hit rates by up to 4.17% using our proposed policies alone on SPEC2006 benchmarks and up to 14% in shared level-2 caches using multicore benchmark suites: PARSEC and SPLASH2.