Parallelization on GPUs of NPB 3.3 NAS tests on language Fortran DVMH

Authors

  • Valeriy Fedorovich Aleksahin
  • Vladimir Aleksandrovich Bahtin
  • Olga Fedorovna Zhukova
  • Aleksandr Sergeevich Kolganov
  • Viktor Alekseevich Kryukov
  • Nataliya Viktorovna Podderyugina
  • Mihail Nikolaevich Pritula
  • Olga Antonievna Savickaya
  • Aleksey Vladimirovich Shubert

Keywords:

DVM; DVMH; CUDA; OpenCL; NASA NPB 3.3.

Abstract

The article describes the transformation of serial versions of the NAS tests package NPB3.3.1 (EP, MG, BT, LU, SP) and the specification of the parallel execution by DVMH-directives necessary for their effective work on clusters with GPUs. The impacts of the various options are examined of parallelization of program performance. The characteristics of tests developed in a high level language Fortran-DVMH (FDVMH), are compared with OpenCL implementation, performed by researchers at Seoul National University.

Downloads

Published

2018-07-12

Issue

Vol. 19 No. 1 (67) (2015): Вестник УГАТУ

Section

INFORMATICS, COMPUTER ENGINEERING AND MANAGEMENT