A New Locality Metric and Case Studies for HPCS Benchmarks

We propose in this paper a new approach to study the temporal and spatial locality of codes using a plot of cache miss bandwidth as a function of cache size and line size for a fully associative LRU cache. We apply this new approach to the study of locality for several High-Performance benchmarks. We show that this plot capture fine behavior of these benchmarks and explain some of the difficulties that recent attempts to characterize locality using a few parameters are facing: Codes can exhibit different levels of temporal or spatial locality for different cache sizes; averaging these different behavior requires to weight properly the cost of misses at different levels of the memory hierarchy. We propose such a scheme, for an average measure of temporal locality.