Recent Development of Fast, Parallel Solvers on GPUs

Hui Liu, Song Yu
Supervisor: Dr. Zhangxing (John) Chen
115

This research focuses on the development of fast Krylov solvers, algebraic multi-grid solvers, and commonly used pre-conditioners, accelerating linear solvers using GPUs and accelerating simulators using the author’s solvers. For the parallel linear solver packages, there are PASS (PArallel Simulation Solvers) and MPASS (Multi-GPUs PArallel Simulation Solvers); where all solvers and pre-conditioners have CPU and GPU versions.

The feature of PASS includes single GPU, seven Krylov solvers and two AMG solvers, seven preconditioners, problem size of millions of grid blocks, up to 15 times faster than CPU sovlers, and so on. And the feature of MPASS includes multi-GPUs, four Krylov solvers, the problem size of 5 millions X #GPUs, upto 50 using 4 GPUs.