Mapping DVMH-programs with regular data dependencies on clusters with graphics processors

Authors

V.A. Bakhtin, A.S. Kolganov, V.A. Krukov, N.V. Podderugina, M.N. Pritula

Annotation

DVMH model (DVM for Heterogeneous systems) was introduced at Keldysh Institute of Applied Mathematics, Russian Academy of Sciences in 2011. High level programming languages which are standard FORTRAN and C languages, extended with directives to map on a parallel machine and implemented as special comments (or pragmas) were developed for new heterogeneous and hybrid supercomputer systems. The problems and methods of mapping of the loops with dependencies on graphics processors are described, and performance of the parallel Fortran DVMH programs with the regular data dependencies is demonstrated in the paper.

Key words

DVM for Heterogeneous systems, Fortran DVMH, hybrid systems with accelerators, graphics processors, CUDA.

Language

English

Reference

V.A. Bakhtin, A.S. Kolganov, V.A. Krukov, N.V. Podderugina, M.N. Pritula. Mapping DVMH-programs with regular data dependencies on clusters with graphics processors // Proceedings of the International supercomputing conference Scientific service in the Internet: all verges of parallelism, Moscow: Moscow University Press, 2013, P. 250-257