Design of fast collective communication functions on clustered workstations with ethernet and myrinet

Abstract

In cluster computing, current communication functions under MPI library are not well optimized. Especially, the performance is worse if there are multiple sources and/or destinations involved, which are the cases of collective communication. Our algorithms uses multidimensional factorization and pairwise exchange communication/dissemination methods to improve the performance. They deliver better performance than previous algorithms such as ring, recursive doubling and dissemination algorithms. Experimental results show the improvement of 50% or so over MPICH version 1.2.6 on a Linux cluster. © Springer-Verlag Berlin Heidelberg 2008.