A Study On GPU/CPU Based Multi-Level Optimization Strategies For CFD Method

Use of CFD and CAA methods can significantly improve the design quality of aircraft projects. While high order CFD and CAA methods can provide high fidelity results, it generally suffers from high computational complexities and computational costs. Kriging model proves to be valuable in optimization problems involving expensive computational simulations such as CFD and CAA. However, as the parameters to be optimized in modern aircrafts design increases when more features were considered, the method suffers efficiency problems when dealing with large amount of high dimensional dataset. The key issue lies in the hyper-parameter tuning in the process of building the Kriging model. This paper proposes the uses of GPU-CPU in hyper-parameter tuning.Interests in using GPU to accelerate specific programs have been rapidly increasing recently. In this paper, we propose a multi-level optimization method using the heterogeneous programming architecture: GPU+CPU. By devising high efficient method, we achieve a speedup of more than 200 in tuning Kriging models. We also reach a speedup of 8 in tuning Memetic algorithms. At the same time, we devise a parallel compact scheme used in CAA using GPU and cluster separately. The results also show that using GPU can significantly improve the performance of the algorithm for achieving results of similar accuracy, while reaching the same speed up, GPU can consume less power comparing to commonly used MPI approach. The current work provides basis further research in the use of GPU/CPU in CAA and CFD aircraft optimization.