A parallel GMRES version for general sparse matrices (Q1920179)

An implementation of a parallel variant of the generalized minimal residual (GMRES) algorithm on Paragon is described, based on two steps: it first builds a Newton basis which is then orthogonalized. This approach requires the parallelization of two steps: the basis formation which relies on matrix vector products and the basis QR factorization. Numerical results on this parallel version of the GMRES algorithm show good performances, even on small matrices.