Fast High Order Large Eddy Simulations on Many Core Computing Systems for Turbomachinery

Abstract

The overall aim of our research is to enable overnight high fidelity LES for realistic industry problems on affordable computing resource. We have adopted a “3E” approach: high spatial discretization Efficiency on general unstructured meshes, high Efficiency accurate time integration and high computing Efficiency on modern low cost HPC hardware. Our approach is centered on high order Flux Reconstruction with local time stepping — the STEFR algorithm [1]. In this paper, an offload-mode version of this code is described targeted at a heterogeneous many-core computing system based on low cost commodity hardware — Intel PHI cards. Three key techniques are introduced to achieve high FLOP rates — and optimal usage of non-equilibrium memory of both CPU and the many core coprocessor — with three levels of parallelization, multi-level nonequilibrium mesh partition and an asynchronous computing structure. A series of high order LES runs for a high lift low pressure turbine blade and a transonic turbine blade, with different order of accuracy, both fully wall-resolved and wall-modelled, were performed, analyzed and presented. This work demonstrates that the high order STEFR method has the potential to support over-night LES for realistic industrial problems on affordable computing resource.