Tracing Data Movements within MPI Collectives
Kevin A. Brown, Jens Domke, Satoshi Matsuoka · 2014
We propose extending common performance measurement and visualization tools to identify network bottlenecks within MPI collectives. By creating additional trace points in the Peruse utility of Open MPI, we track low-level InfiniBand communication events and then visualize the communication profile in Boxfish for a more comprehensive analysis. The proposed tool-chain is non-intrusive and incurs less than 0.1% runtime overhead with the NPB FT benchmark.