Tailoring a vanilla finite-difference solver for high-fidelity simulations of wall turbulence at extreme scales

Despite the leading role of Direct Numerical Simulations (DNS) in wall turbulence research, several fundamental challenges still stand. Many of these concern the flow dynamics at high Reynolds numbers that are just becoming in reach of DNS, thanks to the ever-increasing computing power and improvement of computational tools. This work focused on developing a fast and versatile numerical solver for canonical turbulent flows that can harness modern high-performance computing power while securing the fidelity of the numerical simulation. To achieve this, several changes in the algorithm, implementation, and computational setup choice have been required. Among other things, we (1) use a novel adaptive pencil domain decomposition library for distributed-memory calculations on thousands of GPUs; (2) solve the Navier-Stokes equations in a mixed-precision mode, which allowed for halving the amount of communicated data while ensuring the simulation fidelity; and (3) leveraged a physics-based natural grid and solved the equations on a moving reference frame, which greatly reduced the numerical error and computational effort. Numerical simulations of turbulent channel flow at friction Reynolds numbers of O(10 000) were performed to illustrate the fidelity and high efficiency of the resulting tool for DNS of very large systems. These improvements have been implemented in a new version of the DNS code CaNS.