Measuring the potential benefits of a dynamically adaptive cache line size

Among the most important design parameters in cache memories are storage capacity, associativity, and line size. Conventional caches are tuned to provide fast performance across a variety of representative applications; however, there is no fixed cache configuration that best fits the varying memory requirements of every application. In this paper we study the potential performance benefits of using an adaptive cache that dynamically adjusts its line length to better match the spatial locality of any memory access of a running application. In our L2 cache model, a group of fixed-size cache lines can be concatenated to form longer lines called superlines. We develop an optimistic reference lookahead technique to determine the optimal superline size for every cache miss. The effectiveness of alternative superline length adjustment strategies could then be measured against this theoretical "best case" strategy. Our results show that a cache with adaptive line size can improve the hit rate in up to 3.25%, and produce speedups of up to 14%