Tracing MPI binaries for knowledge discovery

Abstract

Our research project intends to build knowledge about HPC problems to be able to help local researchers. In the general case, researchers who consult us do not own the sources to the parallel programs they run. In order to advise users in choosing parallel machines to run their applications, we are developing a simple methodology to extract knowledge related to performance and scalability, from runs of pure MPI applications on multicore clusters. For the applications under study, our methodology allows us to obtain a reduced description of the program which gives an idea of the original source code.