Large scale simulation of turbulence using a hybrid spectral/finite difference solver

Abstract

Performing Direct Numerical Simulation (DNS) of turbulence on large-scale systems (offering more than 1024 cores) has become a challenge in high performance computing. The computer power increase allows now to solve flow problems on large grids (with close to 10^9 nodes). Moreover these large scale simulations can be performed on non-homogeneous turbulent flows. A reasonable amount of time is needed to converge statistics if the large grid size is combined with a large number of cores. To this end we developed a Navier-Stokes solver, dedicated to situations where only one direction is heterogeneous, and particularly suitable for massive parallel architecture. Based on an hybrid approach spectral/finite-difference, we use a volumetric decomposition of the domain to extend the FFTs computation to a large number of cores. Scalability tests using up to 32K cores as well as preliminary results of a full simulation are presented.