Accelerating simulations of turbulent flows over waves leveraging GPU parallelization

Abstract

We present a highly efficient solver, accelerated by graphics processing units (GPUs), for simulations of turbulent flows over wave surfaces. The solver employs a boundary-fitted curvilinear grid, which is horizontally discretized by a Fourier-based pseudospectral scheme, enabling accurate and efficient resolution of turbulence motions and wave geometry effects across scales. Our GPU implementation incorporates optimizations including kernel fusion for computing derivatives using the fast Fourier transform and a mixed-precision iterative solver for the pressure Poisson equation. A parallel tridiagonal solver is developed to solve the batched systems arising from the iterative Poisson solver on multiple GPUs. Our GPU solver is validated against a canonical problem of turbulence over waves, demonstrating the solver's accuracy. Performance tests indicate substantial speed improvement of up to 80% enabled by the proposed optimizations. Good parallel scalability suggests the solver's capability for large-scale simulations of turbulent flows over waves.