Francis 99 CFD through RapidCFD accelerated GPU code

Abstract

Francis turbines research and development (R&D) requires performance assessment through hydraulic laboratory model testing which can be assisted by auxiliary tools like computational fluid dynamics (CFD), widely used recently. CFD has a history of seeking and requiring ever higher computational performance (HPC) because of the parallelism where Graphics Processor Units (GPUs) have emerged as a major paradigm for solving complex computational problems. However, their implementation to CFD solvers is still a challenge and the tremendous computational power of the GPUs has been wasted. This work presents how the open source RapidCFD code, based on OpenFOAM and ported to Nvidia CUDA, enabled GPUs to be able of running almost entire simulations in thousands of parallel stream cores packed in small form factor hardware in order to solve the incompressible Reynolds-Average-Navier-Stokes (RANS) equations. The simulations were based on a full 3D Francis turbine case which consisted of a grid domain of 23 million cells, including spiral case with stay vanes, distributor, runner and draft tube. CFD results of shaft torque, static pressure and velocity components in steady state deploying a multiple reference frame (MRF) motion approach were compared with available experimental data for main operation conditions at different distributor opening angles: best efficiency operation point (BEP), part load operation point (PL) and full load operation point (HL). The obtained data showed that by transferring directly all the computations to the GPUs, it is possible to make CFD simulations faster compared with central processing units (CPUs). Thus, it is expected to obtain an affordable low computational cost in optimization processes or full range performance evaluations.