SAGA: A Stream Attribute Guided Cache Allocation Policy for Microprocessors

Abstract

Traditional cache replacement schemes such as the commonly used LRU policy often fail to exploit reuse of stream data when the working set size of the application is bigger than the cache size,resulting in poor cache performance.In this paper,the data cache performance is improved by enhancing LRU policy with a novel Stream Attribute Guided Cache Allocation(SAGA) policy,which dynamically utilizes streaming information in applications detected by stream engines on microprocessors to guide whether allocate a new cache line or not when a cache miss occurs.Experiments show that SAGA outperforms LRU by 31% in terms of cache misses and 4% in terms of CPI for SPEC2000FP benchmark on a 1MB cache.