MPI Process Mapping Optimization on Infiniband Clusters

摘要

Communication congestion on Infiniband clusters for MPI parallel applications is considered in this paper. Standing on the point of supercomputer system management and administration, we propose an automatic loading scheme for putting MPI processes onto the nodes with optimized mapping. An indicator is introduced to evaluate the quality of process mapping scheme. Simulated annealing algorithm is then adopted to optimize the mapping scheme according to the indicator. Experiments on different hardware configurations are presented. The results demonstrate that the optimized mapping will benefit the communication of MPI programs.