Parallel hybrid sparse solvers through flexible incomplete cholesky preconditioning

Abstract

We consider parallel preconditioning schemes to accelerate the convergence of Conjugate Gradients (CG) for sparse linear system solution. We develop methods for constructing and applying preconditioned on multiprocessors using incomplete factorizations with selective inversion for improved latency-tolerance. We provide empirical results on the efficiency, scalability and quality of our preconditioners for sparse matrices from model grids and some problems from practical applications. Our results indicate that our preconditioners enable more robust sparse linear system solution. © Springer-Verlag Berlin Heidelberg 2006.