Exploiting GPU-Based HPC Architectures to Accelerate an Unsteady CFD Solver for Turbomachinery Applications

Aircraft engine designers are nowadays facing more and more challenges: they strive to reduce fuel consumption, obtain better engine performance, and create quieter, safer, and environmentally friendlier products. One of the key factors to achieve these goals is the availability of numerical simulation tools able to accurately predict engine behavior and of hardware/software platforms where the tools can successfully run. However, accurate numerical simulations, particularly unsteady Computational Fluid Dynamics (CFD) ones based on sophisticated solvers, are time-consuming and demanding in terms of hardware resources. This may limit the industrial applicability of these methods. A possible strategy to overcome this problem is the acceleration of numerical solvers on advanced High Performance Computing (HPC) architectures, in order to reduce the execution time down to values compatible with industrial needs. Traf is a CFD solver for steady/unsteady three-dimensional Reynolds-averaged Navier-Stokes equations. It is developed at the University of Florence, with a special focus on turbomachinery applications. The current production release is a parallel code that runs on CPU-based platforms. A new version of Traf has been ported to and optimized for GPU-based HPC architectures, in order to dramatically accelerate CFD analyses. The code has been tested on an industrial-grade use case concerning a low-pressure turbine module for aeronautical applications in the context of the EU H2020 funded project LEXIS (Large-Scale Execution for Industry & Society, GA 825532), comparing its performance with the CPU-based release and obtaining promising results. To this aim, the speedup weighted to account for the different hardware cost is selected as a meaningful Key Performance Indicator (KPI).