Malleability Implementation in a MPI Iterative Method

Abstract

In this poster is evaluated the data redistribution stage for two malleable versions of the Conjugate Gradient. One version is based on synchronous communications, while the other one uses asynchronous communications to overlap computation and data redistribution. Both improve execution time when adding more processes, but there is not a noticeable difference between them, because the asynchronous method lowers the performance of the iterations due to the method’s own communications. When both versions are compared, the synchronous version is preferred when resizing to more processes, while the asynchronous one achieves better times when resizing to fewer processes.