Automatic mapping of Fortran programs onto clusters with graphics processing units

Abstract

The paper shows results of exploitation of automated parallelization system SAPFOR for sequential programs parallelization for cluster with GPUs, including programs with regular data dependencies. The system translates Fortran program into Fortran DVMH program. Produced program can be executed on cluster. Fortran DVMH language, its compilers and debugging tools are bundled within DVM-system. Performed transformations with original programs are considered. Programs which use a variety of parallel programming technologies are emitted by the system. Source code characteristics are considered. Performance analysis is done on CPUs and GPUs of the K-100 computational cluster.