Parallel Multi-Projection type methods on hybrid CPU/MIC cluster

Abstract

Many problems occurring in various scientific fields require the efficient solution of large sparse linear systems derived from the discretization of partial differential equations. Preconditioned Krylov subspace iterative methods based on domain decomposition techniques are suitable for solving large sparse linear systems on parallel systems. A parallel preconditioned iterative method in conjunction with semi-aggregation based algebraic domain decomposition method for symmetric sparse linear systems is presented. The proposed method is designed for distributed memory systems with multicore nodes, equipped with many integrated core architecture co-processors (Intel© Xeon Phi™). Utilizing the MIC architecture co-processors, concurrently with existing CPUs, for solving the local linear systems results in accelerating the solution process. Moreover, for large number for subdomains the proposed parallel scheme has improved convergence behavior. The convergence behavior and the scalability of the proposed scheme are examined and numerical results are given.