The impact of a realistic cache structure on a high performance Superscalar architecture

Despite the widely held belief that the most limiting factor in processor performance is the memory hierarchy, much of the recent research into multiple instruction issue techniques assumes a perfect cache structure with a 100% hit rate. This paper attempts to rectify this imbalance by quantifying the performance impact of if incorporating a realistic cache structure into a high-performance superscalar architecture. A highly parameterised cache simulator is integrated into a minimal superscalar architecture, the Hatfield Superscalar Architecture (HSA), that uses static instruction scheduling and in-order instruction issue. Two main studies are presented. First, the impact of a cache on unscheduled code is compared to the impact of a cache on scheduled code. Second, the speedup achieved through static instruction scheduling with a perfect cache is compared to the speedup achieved with a series of cache sizes.