Implementation and Usage of the PERUSE-Interface in Open MPI

This paper describes the implementation, usage and experience with the MPI performance revealing extension interface (Peruse) into the Open MPI implementation. While the PMPI-interface allows timing MPI-functions through wrappers, it can not provide MPI-internal information on MPI-states and lower-level network performance. We introduce the general design criteria of the interface implementation and analyze the overhead generated by this functionality. To support performance evaluation of large-scale applications, tools for visualization are imperative. We extend the tracing library of the Paraver-toolkit to support tracing Peruse-events and show how this helps detecting performance bottlenecks. A test-suite and a real-world application are traced and visualized using Paraver.