Quantitative study of data caches on a multistreamed architecture

In this paper, we quantify the effect that fine grained multistreamed interaction of threads within a shared cache has on the miss rate. By concentrating on the miss rate, we focus on just the cache performance and separate ourselves from a given system architecture. We show the effects of cache capacity, associativity, and line size on the miss rates of multistreamed workloads of two, three and four streams. We then separate the individual components of the miss rate into misses that are intrinsic to the workload (i.e., misses present in a non-multistreamed processor) and extrinsic (i.e., misses introduced due to the multistreamed nature of the processor). Our preliminary results show that for two streamed workloads the increase in miss rate due to multistreaming is less than 60% for caches with some associativity. We observe that for larger cache capacities, direct mapped caches and caches with larger line sizes exhibit poor performance due to a significant increase in the number of extrinsic misses.