Towards Adapting Parallel Programs to Different Platforms: Identifying Interaction Patterns

Modern parallel computing platforms exhibit substantial variation in communication performance between on-socket, on-node and inter-rack locations. When application interaction patterns are irregular, communication-aware process placement on such platforms can be critical for overall runtime. Understanding program interaction patterns can be valuable in order to minimize the impact of capacity variations in communications parameters. Such knowledge can also be useful in selecting the best execution platform from the available options. In this paper we propose an approach based on source code analysis which identifies parallel application communication patterns such as star, ring, mesh, or torus. Our implementation based on ROSE framework was tested on various parallel programs that exhibit different communication patterns. We present the accuracy as well as the limitations of our static identification approach. We envisage augmenting our approach with trace information from pilot runs and best effort approaches to determine a process interaction graph.