A Communication Optimization Scheme for Basis Computation of Krylov Subspace Methods on Multi-GPUs

Abstract

Krylov Subspace Methods (KSMs) are widely used for solving large-scale linear systems and eigenproblems. However, the computation of Krylov subspace bases suffers from the overhead of performing global reduction operations when computing the inner vector products in the orthogonalization steps. In this paper, a hypergraph based communication optimization scheme is applied to Arnoldi and incomplete Arnoldi methods of forming Krylov subspace basis from sparse matrix, and features of these methods are compared in a analytical way. Finally, experiments on a CPU-GPU heterogeneous cluster show that our optimization improves the Arnoldi methods implementations for a generic matrix, and a benefit of up to 10x speedup for some special diagonal structured matrix. The performance advantage also varies for different subspace sizes and matrix formats, which requires a further integration of auto-tuning strategy.