Performance of ILU factorization preconditioners based on multisplittings (Q1412188)

The author investigates the parallelizability of incomplete LU (ILU) preconditioners for sparse block tridiagonal H matrices. Preconditioned Krylov subspace methods are generally necessary to speed up convergence. But the ILU factorization does not lend itself to parallelization. However, if the multisplitting method converges, then \(P^{-1}\) (with \(PA = I-H\) and an iteration matrix \(H\) with spectral radius less than 1) is a good preconditioner for the Krylov method. Such a preconditioner \(P\) is derived from the multisplitting of the block diagonal H-matrix and it can be used in parallel in the preconditioner solver step to advantage. The convergence of such a scheme is studied, as well as the parallel performance for the ILU preconditioned BiCGSTAB algorithm.