PERFORMANCE STUDY OF A PARALLEL DOMAIN DECOMPOSITION METHOD

摘要

This paper studies the influence of various parameters, in order to improve the performances of a parallel Domain Decomposition Method (aka DDM). If introducing more parallelism represents an opportunity to heighten the performance of deterministic schemes, substantial modifications of their architecture are required. In this context, DDM has been implemented into the Apollo3~®multigroup S_n solver, Minaret. The fundamental idea involves splitting a large boundary value problem into several independent subproblems, that can be computed in parallel. Two DDM algorithms are considered. The first one solves a one-group problem per subdomain. The second one is a multigroup block-Jacobi algorithm. To improve performances of these DDM, various parallelism strategies are implemented and compared, depending on the internal structure of the DDM algorithm, the technology chosen (MPI or OpenMP), and the variable parallelized (angular direction or subdomain). Based on these considerations, an efficient hybrid parallelism, suitable for HPC is built: a parallel multigroup Jacobi iteration algorithm, using a two layer MPI/OpenMP architecture, gives the best performances for the reactor configuration studied.