Experimental Evaluation of Array Caches

Cache memories have a dual objective: bridging the gap between memory and CPU speeds and reducing the demand bandwidth on the main memory. These related objectives are achieved by exploiting the locality of access inherent in programs. Locality can be either temporal (when the same location is accessed repeatedly within the a window of references) or spatial (when contiguous locations are accessed with a same window.) This dual nature of locality is not categoric: most accesses are both temporal and spatial in nature; however, accesses to some data structures such as arrays tend to be more often spatial while accesses to scalar variables tend to be temporal. Obviously, spatial locality bene ts from large but fewer cache blocks while temporal locality is best exploited by many but smaller cache blocks. Modern cache organizations attempt to strike a compromise between these designs. This paper reports on the experimental evaluation of an architecture with separate array and scalar caches. The objective is to evaluate the potential bene ts of tailoring a cache organization to a speci c type of locality.