Kasia Swirydowicz

Ali Karakus, Noel Chalmers, Kasia Swirydowicz et al.

We present a GPU-accelerated version of a high-order discontinuous Galerkin discretization of the unsteady incompressible Navier-Stokes equations. The equations are discretized in time using a semi-implicit scheme with explicit treatment of the nonlinear term and implicit treatment of the split Stokes operators. The pressure system is solved with a conjugate gradient method together with a fully GPU-accelerated multigrid preconditioner which is designed to minimize memory requirements and to increase overall performance. A semi-Lagrangian subcycling advection algorithm is used to shift the computational load per timestep away from the pressure Poisson solve by allowing larger timestep sizes in exchange for an increased number of advection steps. Numerical results confirm we achieve the design order accuracy in time and space. We optimize the performance of the most time-consuming kernels by tuning the fine-grain parallelism, memory utilization, and maximizing bandwidth. To assess overall performance we present an empirically calibrated roofline performance model for a target GPU to explain the achieved efficiency. We demonstrate that, in the most cases, the kernels used in the solver are close to their empirically predicted roofline performance.

Kasia Swirydowicz, Julien Langou, Shreyas Ananthan et al.

Communication-avoiding and pipelined variants of Krylov solvers are critical for the scalability of linear system solvers on future exascale architectures. We present low synchronization variants of iterated classical (CGS) and modified Gram-Schmidt (MGS) algorithms that require one and two global reduction communication steps. Derivations of low synchronization iterated CGS algorithms are based on previous work by Ruhe. Our main contribution is to introduce a backward normalization lag into the compact $WY$ form of MGS resulting in a ${\cal O}(\eps)κ(A)$ stable GMRES algorithm that requires only one global synchronization per iteration. The reduction operations are overlapped with computations and pipelined to optimize performance. Further improvements in performance are achieved by accelerating GMRES BLAS-2 operations on GPUs.