Timely for Siggraph 2014 (because animations use meshes) and food-for-thought for CFD (Computational Fluid Dynamics) research, the paper by A. Gorobets, F.X. Trias, R. Borrell, G. Oyarzún and A. Oliva, “Direct Numerical Simulation of Turbulent Flows with Parallel Algorithms for Various Computing Architectures” considers structured and unstructured meshes for incompressible flows and overlapped MPI communication patterns. The authors are quite clear in their performance conclusions about their new algorithm, “Performance comparison for basic OpenCL kernels of the algorithm on unstructured meshes showed that the different GPUs considered substantially outperform Intel Xeon Phi accelerator. Also, the AMD GPU tends to be more efficient than NVIDIA on heavy computing kernels.” Such performance comparisons -when fairly performed on generally usable kernels – provide valuable information to help cut through all the marketing. For more information see http://termofluids.com/.
The purpose of the work is twofold. Firstly, it is devoted to the development of efficient parallel algorithms for large-scale simulations of turbulent flows on different
supercomputer architectures. It reports experience with massively-parallel accelerators including graphics processing units of AMD and NVIDIA and Intel Xeon Phi coprocessors.
Secondly, it introduces new series of direct numerical simulations of incompressible turbulent flows with heat transfer performed with the considered algorithms.
Parallel finite-volume algorithms for large-scale simulations of turbulent flows have been considered. Simple approaches for OpenMP parallelization aimed at hundreds of threads have
been presented for both structured and unstructured meshes. High internal speedups ensuring good level of parallelism demonstrated on Intel Xeon Phi surprisingly coexist with relatively
poor net performance comparable with 8-core CPU. Implementation of overlapped communications with an OpenCL task scheduler infrastructure has demonstrated promising
results. Performance comparison for basic OpenCL kernels of the algorithm on unstructured meshes showed that the different GPUs considered substantially outperform Intel Xeon Phi
accelerator. Also, the AMD GPU tends to be more efficient than NVIDIA on heavy computing kernels.
For more information: