Optimization of All-to-All Communication on Multi-core Cluster Systems
Daisuke Takahashi · 2011
In this paper, we propose an optimization method for all-to-all communication on multi-core cluster systems. In the all-to-all communication, increasing of the start up time prevents the scalability. We show that a two-phase all-to-all algorithm improves performance effectively by reducing the start up time of communication for a larger number of MPI processes. We also improve the performance of all-to-all communication by automatic tuning. The automatically tuned all-to-all communication achieves up to approximately 5.4x speedup over the native MPI All to all routine on 1024 cores. The performance results demonstrate that the automatically tuned all-to-all communication is most advantageous on multi-core cluster systems.