Optimizing processes mapping for tasks with non-uniform data exchange run on cluster with different interconnects

The problem of mapping the parallel task to the nodes of computing cluster is considered. MPI software with non-uniform communication and heterogeneous interconnect of computing cluster require to appropriate parallel processes mapping for optimization of data exchange. The graph mapping algorithm is developed. It uses parallel program representation as a task graph and cluster topology representation as system graph. The proposed optimization technique is tested on synthetic benchmark and on real QBox software to study its efficiency on large number of computing cores. The positive results of optimization are achieved and the summary is presented in the paper. Speedup of 17-20 % is obtained on scalable benchmarks using 300-4800 parallel processes.