High-level topology-oblivious optimization of mpi broadcast algorithms on extreme-scale platforms

Abstract

There has been a significant research in collective communication operations, in particular in MPI broadcast, on distributed memory platforms. Most of the research works are done to optimize the collective operations for particular architectures by taking into account either their topology or platform parameters. In this work we propose a very simple and at the same time general approach to optimize legacy MPI broadcast algorithms, which are widely used in MPICH and OpenMPI. Theoretical analysis and experimental results on IBM BlueGene/P and a cluster of Grid’5000 platform are presented.