Accelerating steady-state CFD simulations by leveraging non-converged linear solutions

Kurzfassung

Aircraft simulations along the flight envelope require many extensive steady-state RANS simulations. Each simulation is solved through an inexact-Newton iteration, where for each iteration a linear and a non-linear system have to be solved. Solving the linear system accounts for the dominant share of total wall-clock time in this solution process, making the optimization of its linear solver, particularly through preconditioning, very important. \\ This paper evaluates the preconditioning strategies particularly for the Generalised Minimum RESidual (GMRES) linear solver available in the CFD solver \enquote{CFD for ONERA, DLR and Airbus} (CODA)\footnote{CODA is the computational fluid dynamics (CFD) software being developed as part of a collaboration between the French Aerospace Lab ONERA, the German Aerospace Center (DLR), Airbus, and their European research partners. CODA is jointly owned by ONERA, DLR and Airbus. }. algorithms presented, while the library PETSc's~\cite{PETSc} implementation in CODA uses the partition based approach. Many different preconditioning approaches of the two linear solver libraries Spliss and PETSc are investigated, ranging from nested preconditioners such as Jacobi with LU factorisation to direct solvers such as ILU and different multigrid approaches. Furthermore, two Spliss implementations in CODA regarding hybrid parallelisation and the usage of mixed precision are investigated. \\ The high-speed wing-body NASA Common Research Model is used for evaluating the solver combinations, offering a well established simulation framework and the ability to evaluate scaling behaviour across different mesh sizes. Best-practice configurations are identified and key pitfalls, such as partition sensitivity and solution sensitivity, are highlighted.