An Approach for Fine-Grained Profiling of Mesh-Based Parallel Programs

Characterizing the dynamic behavior of parallel programs in terms of their execution profile helps to understand their behavior and suggest optimization strategies to improve the performance. Traditional event tracing techniques write the profiled data to trace files. Using the traditional approach for fine grained profiling not only yields large unwieldy trace files but often also gives skewed results due to the inaccuracies introduced by the profiling. This paper describes an approach to profile mesh-based parallel programs at a very fine level of granularity by measuring performance metrics at the level of each mesh element. The approach described in this paper is novel in that profile data is associated with mesh elements, not processors, so the profile data can be used, for example, to develop adaptive load balancing policies. A tool implementing the idea described in this paper is developed which provides an easy-to-use C API with Fortran 90 wrappers to enable fine grained profiling of mesh-based parallel applications.