Influence of Time Integration Method on GPU Performance for Industry Relevant CFD Simulations

Massive parallel computation on GPUs allows promising speedups for high performance computing, especially for CFD. However, simulations with different algorithms on different industry relevant cases may not exhibit the same performance on the same GPU. Using JAXA's FaSTAR CFD solver, accelerated with OpenACC, the performance of two time integration methods, the LU-SGS and the DP-LUR methods, was investigated on two different cases. For a simple case such as the NASA CRM geometry with low angle of attack at cruise flight, the DP-LUR method is 1.37 times faster than the LU-SGS method on GPU. For a more complex case such as the 30P30N geometry with higher angle of attack at takeoff, the method is 1.46 times slower.