The International Conference for High Performance Computing, Networking, Storage and Analysis
PyFR: An Open Source Python Framework for High-Order CFD on Heterogeneous Platforms.
Authors: Freddie D. Witherden (Imperial College London), Brian C. Vermeire (Imperial College London), Peter E. Vincent (Imperial College London)
Abstract: Computational fluid dynamics underpins several high-tech industries. However, the numerics utilised by most industrial codes are not suitable candidates for acceleration. In this work we present a framework for solving the compressible Navier-Stokes equations efficiently across a range of hardware platforms. The framework, PyFR, is open source and written entirely in Python. By employing the flux reconstruction approach of Huynh PyFR is able to combine the accuracy of spectral schemes with the geometric flexibility of low-order finite volume schemes.
Our results show that PyFR is both accurate and performance portable across CPUs and GPUs from Intel, AMD and NVIDIA. We also demonstrate the heterogeneous capabilities of PyFR which permit a simulation to be decomposed across a range of platforms. Finally, we show the scalability of PyFR on up to 104 NVIDIA M2090 GPUs.