Parallel DNS algorithms on unstructured grids

Abstract

With the prospect of Petaflop computing to be realized in the near future, we present two different parallel algorithms suitable for simulating turbulent flows in non-separable and multiply-connected computational domains. The algorithms are based on a new class of hierarchical spectral methods appropriate for tensor-product representations in hybrid subdomains, i.e., tetrahedra, hexahedra, prisms and pyramids. We review the numerical implementation of the spectral method and subsequently present two parallel paradigms, the first for a spectral element/Fourier algorithm, and the second for a fully 3D algorithm based on geometric domain decomposition. Emphasis is placed on the communication patterns of these algorithms in conjunction with the features of current or upcoming computer models. Examples of turbulent parallel simulations are included and limitations in currently achieving high parallel efficiencies are discussed. A perspective on the future of DNS on the emerging distributed shared memory (DSM) computer architectures is presented.