Options
2014
Journal Article
Titel
Hybrid-parallel sparse matrix-vector multiplication and iterative linear solvers with the communication library GPI
Abstract
We present a library of Krylov subspace iterative solvers built over the PGAS-type communication layer GPI. The hybrid pattern is here the appropriate choice to reveal the hierarchical parallelism of clusters with multiand many- core nodes. Our approach includes asynchronous communication and differs in many aspects from the classical one. We first present the GPI-based implementation of the sparse matrix-vector multiplication and then, using as a benchmark the numerical solution of a Poisson boundary value problem in a unit cube, we compare the performance on Intel/Infiniband and CRAY XE6 architectures of our GPI-based Conjugate Gradients and Richardson methods against the ones available in PETSc. The results show good scalability and performance of our approach, at least comparable to these of PETSc.
Author(s)