Event Flow Graphs for MPI Performance Monitoring and Analysis

High Performance Computing (HPC) infrastructures provide an enormous amount of computational power, however, this power comes with an increase in software complexity. Applications have to be tuned to highly-complex underlying hardware, combine several programming models such as MPI and OpenMP, or utilize accelerator specific directives to exploit parallelism. As a result, application performance is usually far from ideal. Therefore, performance analysis of scientific parallel applications is essential to use HPC infrastructures efficiently.

In this talk, we explore the use of Event Flow Graphs for performance data collection and analysis. Event Flow Graphs in the context of performance monitoring combine the scalability and low overhead of profiling methods together with lossless information recording of tracing tools. In addition, these graphs require significantly less storage space than regular event traces, while still being able to reconstruct the full ordered sequence of events performed by the application. Moreover, graphs can also be used to detect application structure, and to visually explore performance data.