Parallelizable approximate solvers for recursions arising in preconditioning (Q1386504)

The aim of this paper is to show that it is possible to replace the modified incomplete LU recursions by corresponding restarted recursions which are efficiently implemented on parallel and vector computers. Two types of domain decomposition, suitable for two types of parallel architectures are introduced. The first one is based on overlapping square subdomains and is suitable for message-passing parallel architecture with mesh-connected processor arrays. The second type uses subdomains consisting of several consecutive data fronts, suitable for line-connected arrays of vector or SIMD processors. Finally, numerical examples with the standard and parallelizable algorithms are presented.