GPU-accelerated CFD Simulations for Turbomachinery Design Optimization

Design optimization relies heavily on time-consuming simulations, especially when using gradient-free optimization methods. These methods require a large number of simulations in order to get a remarkable improvement over reference designs, which are nowadays based on the accumulated engineering knowledge already quite optimal.
High-Performance Computing (HPC) is essential to reduce the execution time
of the simulations. While parallel programming using the CPU is established since more than two decades, the use of accelerators, such as the Graphics Processing Unit (GPU), is relatively recent in design optimization. The GPU has actually a huge computational power comparable to a many-core cluster but concentrated in one device. This raw power is not easy to utilize as entire code parts have to be rewritten using a GPU programming language. Even though high-level standards (e.g. openACC) are able to bring a basic acceleration with a low development effort, it is not simple to get large speedups with these methods. Low-level programming languages are more efficient but different speedups are reported and there is a need for
a deep analysis to make the GPU potential more transparent to scientists especially non-experts in HPC.
In order to study the GPU acceleration for CFD steady simulations, two in-house CFD solvers have been ported to the GPU; one with explicit and the second with implicit time-stepping. After the porting and the validation of the GPU solvers, the GPU code optimization leads to the identification of a set of key parameters affecting the GPU efficiency. At the same time, both methods have been compared resulting into a performance model and a classification of the GPU acceleration of some CFD operations. The purpose is to enable scientists to take an educated decision concerning the GPU porting of their CPU applications by providing an expected GPU speedup.
In addition to the two GPU CFD solvers that are now integrated into the in-
house design optimization software package, this research provided key elements to reduce the ambiguity about the GPU potential, namely a qualitative analysis and a classification. These tools can help selecting the best candidate for a breakthrough in CFD acceleration. At the same time, this work identified serious limitations in the preconditioning of a linear system of equations and the limit of today iterative matrix factorization methods in terms of stability and convergence. There is a need for a paradigm shift toward inherently parallel preconditioners. The developed tools have been used for the optimization of a compressor and a turbine cascade resulting
into a faster optimization process on the GPU.