Alternatives for parallel Krylov subspace basis computation (Q4383433)

[Shortened review. The full version is available at Zbl.] Several alternatives for the parallel implementation of the Arnoldi-based Krylov subspace methods are examined and experimentally tested in distributed memory multiprocessor environment. The Arnoldi procedure for constructing an orthonormal basis of the Krylov subspace is frequently one of the building blocks in iterative methods for the solution of linear systems or eigenvalue problems such as GMRES, FOM or the Arnoldi method. On serial computers, the modified Gram-Schmidt implementation of the Arnoldi procedure is widely used and considered as a reliable technique. In parallel environment, the Newton basis Arnoldi implementation proposed by \textit{Z. Bai, D. Hu} and \textit{L. Reichel} [IMA J. Numer. Anal. 14, No. 4, 563-581 (1994; Zbl 0818.65022)] represents a successful alternative. In this implementation, a monomial basis of the Krylov subspace (Newton basis) shifted by the Leja point ordering of the spectrum of the associated Hessenberg matrix is constructed first using the iterated matrix-vector multiplication and then orthogonalized using some standard orthogonalization technique. In the paper, a parallel template for the efficient implementation of such algorithm is presented and studied in a distributed memory multiprocessor environment. Several aspects such as load-balancing, minimization of the communication and data distribution among processors are considered.