A GMRES method on distributed parallel computers for unsteady flow simulation

To improve the computational efficiency and applicability for high performance parallel computing of CFD, a parallel full-implicit method based on GMRES+LU-SGS is developed for unsteady flow simulation. The convergence and parallel property of the method have been studied. The restartable GMRES method is improved to save computing time in the process of GMRES restart. The domain decomposition and Hybrid LU-SGS preconditioner have been studied for parallel implementation. A robust Negative-SA turbulence model for large CFL number, computing and storing the Jacobian matrix and grid reorder methods have been studied to improve the computational efficiency. The methods have been used to compute the plane flow, the flow around NACA0012 airfoil, wing-body configuration, F-16 fighter aeroelastics, and rotor unsteady flow simulation. The numerical results indicate that these methods on distributed parallel computers lead to 20%~200% increase in performance over LU-SGS method. Moreover, remarkable super-linear speedup is achieved on 240 processors.