OPTIMIZATION ALGORITHMS OF PARALLEL PROGRAMS EXECUTION ON HIGH-PERFORMANCE SYSTEMS WHEN SOLVING PROBLEMS OF MODELING PHYSICAL PROCESSES
K.V. Pavsky1,2, M. G. Kurnosov1,2, A. V. Efimov1,2, K. E. Kramarenko1,2, E. N. Peryshkova1,2, A. Yu. Polyakov3
1Rzhanov Institute of Semiconductor Physics, Siberian Branch, Russian Academy of Sciences, Novosibirsk, Russia
2Siberian State University of Telecommunications and Information Sciences, Novosibirsk, Russia
3NVIDIA Corporation, Santa Clara, California, USA
Keywords: high-performance systems, parallel programs, collective operations, synchronization of memory access, read-write locks, MPI, PMIx
Abstract
The paper proposes algorithms that allow to increase the efficiency of parallel programs execution on high-performance computer systems, in particular, when solving problems of modeling physical processes. The developed algorithms are focused on reducing the execution time of collective operations for computer systems with SMP/NUMA node architecture in standard MPI libraries. The developed read-write blocking algorithms increase the efficiency of synchronization of access to shared memory, relative to the algorithms used in the Open PMIx library.
|
