Statistical methods for automatic performance bottleneck detection in MPI based programs

Abstract

With increasing number of processors the visualization of trace data for MPI communication primitives becomes harder due to many limitations. An automatic analysis of the recorded data should be able to identify some clues for the optimization of the source code. The approach presented with this paper is able to divide the communication time for each communication relationship into a necessary part and the overhead. We use statistical methods to define a time for each single communication event that we will accept as the 'usual' time for this event. Additional runtime will be considered as overhead. At the end of this article, the benefits from this method are demonstrated on the Sweep3D benchmark. © Springer-Verlag Berlin Heidelberg 2005.