Hybridní paralelizace TFETI-1 metody

Abstract

The thesis deals with the hybrid parallelization of the TFETI-1 method which is implemented in the library PermonFLLOP. At first, the thesis presents various parallel programming models and ways of the hybrid parallelization. There are listed the possible benefits of the hybrid parallelization compared with the pure MPI parallelization. Then the thesis analyzes the packages providing direct solvers, which have important role in the TFETI-1 implementation, in terms of suitability for hybrid parallelization. New implementation extends the existing pure MPI implementation. With this extension, data of more subdomains can be now stored per one MPI process. This allows the use of good properties of the numerical scalability of the FETI methods. In numerical experiments, the numerical scalability and the impact of hybrid parallelization are tested.