Knowledge based automatic scalability analysis and extrapolation for MPI programs

Abstract

The question how well a MPI program is scaling with an increasing number of processors becomes more and more interesting, especially when these number grows to 10.000 or even 100.000 with IBM's 'Blue Gene' this year. The approach presented with this paper is able to identify locations within the source code of an application where the communication effort does not scale well with the growing number of processors. We show how traces for the same program generated with different numbers of processors can be inspected and compared automatically. An analytical approach will then identify the points within the source that do not scale as expected. At the end of this article, the benefits from this method are demonstrated on an ASCI benchmark. © Springer-Verlag Berlin Heidelberg 2005.