Improving Uniformity of Cache Access Pattern using Split Data Caches

In this paper we show that partitioning data cache into array and scalar caches can improve cache access pattern without having to remap data, while maintaining the constant access time of a direct-mapped cache and improving the performance of L-1 cache memories. By using 4 central moments (mean, standard-deviation, skewness and kurtosis) we report on the frequency of accesses to cache sets and show that split data caches significantly mitigate the problem of non-uniform accesses to cache sets for several embedded benchmarks (from MiBench) and some SPEC benchmarks