Using GPU computing in iterative solvers for sparse linear systems.
We illustrate how Graphics Processing Units (GPU) can be used to accelerate iterative solvers based on Krylov subspace methods. First we consider sparse linear systems from computational fluid dynamics (CFD) simulations. We evaluate and optimize the performance of Conjugate Gradient routines designed for GPU accelerator and compare against an industrial CPU-based implementation. Then we show how the distributed parallel Algebraic Recursive Multilevel solver (pARMS) based on MPI can be adapted for heterogeneous CPU/GPU architectures. The preconditioning of each part of the distributed matrix (local preconditioning) is performed on a GPU and is based on the randomization of the last Schur complement system in the multilevel recursive process.