Optimization of Mapping Programs to Computing Resources

E. A. Kiselev, Researcher, e-mail: kiselev@jscc.ru, Joint Supercomputer Center of the Russian Academy of Sciences, Moscow, V. V. Korneev, Professor, Director for Research, e-mail: korv@rdi-kvant.ru, State Enterprise "Research and Development Institute "Kvant", Moscow

Possibility to decline execution time of parallel programs with optimization of mapping programs to supercomputer resources is considered. The optimization process performs mapping application graph to parallel computer system graph. An application graph represents aggregate information on communication operations between parallel application's processes. It contains aggregate information about number of operations transmit/receive and the amount of data transferred for each processes pair. A system graph discloses a parallel computer system's communication structure. For each communication channel it comprises a latency values and bandwidth. Parallel implementation of the simulation annealing algorithm is proposed. The basic idea of the algorithm is the localization of communication exchanges between the intensity of the interaction program branches and reduce competition for communication resources. Building application graph and a parallel system graph are performed automatically using the developed software. Results of optimization test High Performance Conjugate Gradient execution on "MBC-10p" supercomputer resources using created software tools are presented. Time evaluation of mapping algorithm is done.

Keywords: supercomputer, test HPCG, mapping of program, information graph, latency of communication fabric, analysis of communication fabric structure, analysis of parallel program structure, simulated annealing tool, MPI-program

pp. 3–8