Developing a CPU-GPU LES Parallel Solver for Canonical Turbulent Flows

Parallel programming on CPU and GPU has accelerated CFD simulations. Recently a fast solver of incompressible canonical flow has been released as open-source code called CaNS (Costa in Comput Math Appl 76(8):1853-1862, 2018). The code is massively parallel, and the Poisson equation is solved with eigenfunction expansions based on FFT. This work presents this solver's extensions for CPU and GPU versions (Costa et al. in Comput Math Appl 81:502-511, 2021), adding the large eddy simulation (LES) numerical technique. Here, the first step aims to validate the code's LES extension to predict the details of the turbulent flows in comparison with DNS data. Considering the constrained Cartesian domain of the code, validation has been performed on three geometries: channel, duct, and cavity. Computed mean-velocity and turbulence fluctuations show good agreement with DNS data. The needed computational time is one-four hundred of that for DNS. At the next step, LES of turbulent channel flow for several Re numbers was done to investigate the speed-up measurements for the CPU and GPU versions. Computations on the GPU perform 1.7 to 6.5 times faster than on the CPU version. Furthermore, the limitation of memory accesses was discussed as the problem size increased. Our results shed light on how to get the maximum acceleration of a fast solver with an available processor.