high performance computing on graphics processing units: hgpu.org

We propose an adaptability judging method applied to sparse matrices and the target cache memory using two metrics based on spatial locality and temporal locality. For indirect access sequences of sparse matrix-vector multiplications, one metric is the number of valid data within a cache line, and another metric is average reference interval. We also develop a set of analysis tools to generate the above performance metrics, histograms of reference intervals and theoretical cache hit rates. As an experimental result, a cache memory behavior which was difficult to explain from the view point of spatial locality becomes explicable from that of temporal locality. Outputs of the tool show that return on investment is too thin to increase the cache memory capacity unless all the elements of a column vector with the same size as the number of columns of the sparse matrix are stored in the cache memory.