Optimization of MPI-Process Mapping for Clusters with Angara Interconnect

摘要

An algorithm of MPI processes mapping optimization is adapted for supercomputers with interconnect Angara. The mapping algorithm is based on partitioning of parallel program communication pattern. It is performed in such a way that the processes between which the most intensive exchanges take place are tied to the nodes/processors with the highest bandwidth. The algorithm finds a near-optimal distribution of its processes for processor cores to minimize the total execution time of exchanges between MPI processes. The analysis of results of optimized placement of processes using proposed method on small supercomputers is shown. The analysis of the dependence of the MPI program execution time on supercomputer parameters and task parameters is performed. A theoretical model is proposed for estimation of effect of mapping optimization on the execution time for several types of supercomputer topologies. The prospect of using implemented optimization library for large-scale supercomputers with the interconnect Angara is discussed.