On a Hierarchical Parallel Algebraic Domain Decomposition Linear Solver
Résumé
The solution of linear systems is often the most computational consuming kernel in large complex numerical simulations. In this talk, we will describe a parallel algebraic hierarchical linear solver for sparse linear systems. The numerical scheme based on a partition of the adjacency graph of a sparse matrix, that leads to the solution of a Schur complement system, will be presented as well as the related preconditioning technique. Parallel numerical experiments of the hybrid direct/iterative technique will be described on 3D examples from both academic and industrial relevance. Prospective for implementations on many- cores heterogeneous systems on runtime systems will be discussed.