SC Conference - Activity Details

Serban Georgescu  (University of Tokyo)

Yohei Sato  (University of Tokyo)

Hiroshi Okuda  (University of Tokyo)

Tuesday,  05:15PM - 07:00PM

Room Rotunda Lobby

Krylov-type iterative solvers are widely used for solving large systems of linear equations. Severely memory bounded, Krylov solvers have difficulty achieving even 10% of CPU peak. With the large and widening performance gap between CPUs and GPUs, employing GPUs as accelerators looks indeed promising. In this poster we present a new open-source software package named CUKr (CUDA Krylov) intended as a scalable framework for building Krylov solvers on GPU clusters. Additionally, the implementation style required for working on multiple GPUs makes it suitable for multicore/manycore. The same solver can be run multi-node / multi-GPU / multi-core, with or without iterative refinement and so on, using simple command-line options. We will be showing performance and scalability results obtained for a Conjugate Gradient solver built using this framework on a collection of real-world matrices on a GPU cluster and a multicore system. For large matrices, up to 15x speedup has been obtained.