A GraphHunter software tool for mapping parallel programs to a supercomputer system structure

One of well-known problems in high-performance computing is optimal mapping of parallel program processes to supercomputer system nodes. A solution for this problem minimizes the overhead for information exchanges between the processes of a parallel program and thus increases the performance of calculations. When solving a mapping problem, both a supercomputer system and a parallel program are represented as graphs. The paper shows solving the mapping problem in relation to a system for collective use of a supercom-puter that handles a queue of parallel programs. After passing the queue, a new previously unknown subset of supercomputer nodes is allocated to the parallel program. In this case, it is necessary to construct a graph of a selected subset of nodes and find a suitable mapping of the parallel program onto this graph in a reasonable time. It is suggested to run parallel mapping algorithms on the supercomputer nodes allocated for parallel program. To study the properties of mapping algorithms, the GraphHunter software tool was developed. This tool makes it possible to conduct experiments with three parallel algorithms: simulated annealing, genetic algorithm, and their combination. This article discusses the structure of the GraphHunter software tool, and pre-sents the results of experiments with GraphHunter runs on the MVS-10P OP supercomputer at the Joint Su-percomputing Center of the Russian Academy of Sciences.