Yousef Saad (Université du Minnesota, http://www-users.cs.umn.edu/~saad/) fera une présentation le mardi 2 juin à partir de 10h00 salle Ada Lovelace.
Titre : « Parallel Multilevel Low-Rank approximation preconditioners »
Résumé:
This presentation will discuss a class of preconditioning methods for solving linear systems of equations that are based on exploiting low-rank approximations to certain matrices. These methods have a number of appealing features. Because they are essentially approximate inverse techniques, they handle indefiniteness quite well. Furthermore, they are amenable to SIMD compuations such those inherent to GPUs. The talk will first describe a recursive divide and conquer approach geared toward Symmetric Positive Definite model problems issued from Finite Difference discretizations of PDEs. Then two extensions of this general approach will be described. The first exploits Schur complements in a parallel computing Domain Decomposition (DD) framework. The second extends this DD approach further by considering so-called `hierarchical interface decomposition further by considering so-called `hierarchical interface decomposition orderings’ which are essentially algebraic generalizations of `wirebaskets’ techniques used in Domain Decomposition methods.