Summary

Communication behavior is one of the most important aspects of parallel applications. The Splash-2 paper, unfortunately, chooses to characterize such an important issue with a single umbrella parameter called ``traffic per instruction''. We have demonstrated in this section how inadequate this methodology is. In particular, we have shown that issues such as communication granularity, the ability of overlapping communication with computation, and traffic bustiness can result in drastically different program behavior and performance. To address the inadequacy of the Splash-2 effort, we propose mainly two sets of histograms for modeling communication behavior. One measures the amount of communication before each instance of computation; and the second measures the amount of computation that can still make progress after each instance of communication.