An Infrastructure for the Analysis of Communication Patterns in Virtual Topologies

The virtual topology of a parallel application is the neighborhood relationship between communicating processes developed due to specific communication patterns resulting from domain decomposition. We present an infrastructure that allows the usage of topological information for the performance analysis of a parallel application. For this purpose we have implemented an easy to use extension of the KOJAK performance analysis toolkit.

The KOJAK toolkit defines communication patterns for parallel applications which describe inefficient behavior. The performance analysis is carried out by calculating the effect of these inefficiency patterns on the application's performance. The distribution of these inefficiency patterns is studied across a three-dimensional performance space. The knowledge of virtual topology can be exploited to explain the occurrence of these inefficiency patterns in terms of higher-level events related to the parallel algorithm implemented in the application. Also, it can be used to visualize the relationships between pattern occurrences and the topological characteristics of the affected processes, To prove these principles, we have used our extensions to KOJAK to analyze two realistic MPI applications.