MPI Performance Characterization on InfiniBand with Fine-Grain Multithreaded Communication
Shane Farmer, Anthony Skjellum, Ryan Eric Grant, Ron Brightwell · 2016
This paper describes work with end goal of quantifying the impact of threading on MPI performance models in order to enable justifications of model construction methods. To do so, it evaluates benchmarks on a specific, representative networked platform, and makes these contributions: 1) it evaluates the performance of point-to-point transmission between two multithreaded Message Passing Interface (MPI) processes over InfiniBand, 2) it elucidates strengths and weaknesses of Intel MPI under multiple multithreaded modes, 3) it provides indications as to where predictive modeling of such multithreaded performance would prove fruitful. Two established benchmarks are used to obtain performance information and they are compared and contrasted as needed to explain the results described here. In one case, we enhanced a benchmark to enable threading for study, justification for these changes are provided.