An Analysis of Data Access Patterns in Integer Benchmarks

We present memory reference data for the SpecInt95 benchmarks that can be used in evaluating the potential for prefetching spatially correlated data. In contrast with previous work, the data presented are for complete program runs and trace all executed instructions. The data are classified according to their access patterns and we show how types of access patterns arise from the executed benchmark code. Finally, we examine the accuracy of several published beliefs concerning memory access patterns. One common belief is that only scientific code displays sufficiently “regular” behavior for techniques such as prefetching to be of benefit. Our conclusion is that there is significant amount of “regular” behavior even in non-scientific code.