Parallel algorithms of the Shur complement method in software model CUDA+OpenMP

Abstract

Implementation of the Shur complement method on the hybrid (CPU/GPU) architecture is effective because of quality of computing distribution between CPU and GPUs. We have show that the formation of the Shur complement matrix can be performed efficiently on the GPU for matrix consisting of several thousand rows and columns. To solve the interface system, we proposed parallel algorithm of the conjugate gradients method with explicit preconditioning. This helps to achieve significant acceleration of the computing on several GPUs.