The evolution of computing brings up two very important trends: the proliferation of cores in processors and generalization of SIMD processing units. To exploit the full computing power provided by these developments the use of fine grain parallelism is essential. The problem is that many of parallel computing algorithms are not designed for fine-grained parallelization and the question arises as to develop these algorithms or find new ones to effectively leverage these processors. The proposed research project to study fine grained parallelization of ILU methods, used as preconditioners in iterative linear solvers. The research work will focus on the permutations of linear systems, which introduce fine grain parallelism, and their impact on the convergence of the iterative solver. The aim will be to propose new permutations for obtaining the degree of parallelism required for the targeted computing units while maintaining a fast convergence of the iterative solver.
Keywords: HPC, parallelism, linear algebra, incomplete factorizations, OpenMP, MPI, GPU.
Academic supervisor Pr, Giraud Luc, INRIA Bordeaux
Doctoral School EDMI/Ecole Doctorale Mathématique et Informatique https://ed-mi.u-bordeaux.fr/
IFPEN supervisor GUIGNON Thomas, Scientific Computing Department, thomas.guignonifpen.fr
PhD location Scientific computing Department, IFPEN, Rueil Malmaison, France & INRIA Bordeaux, Talence, France
Duration and start date 3 years, starting preferably on November 1, 2019
Employer IFPEN, Rueil malmaison
Academic requirements University Master degree or Engineering degree in High performance computing, linear algebra.
Language requirements Fluency in English, Fluency in French or willingness to learn French
Other requirements MATLAB, Algorithmic, C++, Software Engineering, Computing systems