Implementing a 3D multigrid algorithm on Fujitsu's vector parallel supercomputer

摘要

Multigrid algorithms are well adopted by many engineering applications programs on the traditional sequential computers due to their good convergence behavior. In order to convert those existing resources to the field of parallel computing in a timely manner, an effective parallelization methodology should be set up to ease the conversion efforts and to take advantage of the power of parallel processing at the same time. We demonstrate how domain decomposition can be applied to the multigrid algorithm by means of a simple three-dimensional problem which is solved by a V-cycle scheme. The sequential code in FORTRAN provided by NASA NAS, known as the NAS multigrid parallel benchmark test, is being used in our current study. We have concluded that the parallelization methodology presented here would greatly reduce conversion efforts in the process of parallelization and would obtain very good performance and scalability on the Fujitsu's VPP500 computers.