Parallel Preconditioning with Lower-Upper Approximate Inverse Factors

Edmond Chow, Center for Applied Scientific Computing Lawrence Livermore National Laboratory
Yousef Saad, Department of Computer Science and Engineering, University of Minnesota

Abstract:

Much research in sparse approximate inverse preconditioning is converging to triangular factorized forms. A new method for computing these factors based on bordering will be presented and evaluated. Compared to other methods, the new method is both more parallel and offers the possibility of updating the sparsity pattern of the approximate inverse dynamically. Issues including approximate inverse structure prediction and the smoothing property of approximate inverses will also be presented.