Performance of a Newton–Krylov–Schur Algorithm for Solving Steady Turbulent Flows

A methodology is presented for characterizing flow solver performance. The methodology can be applied to assess the efficiency of a given approach, where efficiency is defined in terms of accuracy per unit cost measured in central processing unit time. The procedure is presented by demonstrating its application to the parallel Newton–Krylov–Schur finite difference flow solver known as Diablo. The benchmark cases to which the procedure is applied are two-dimensional turbulent flows modeled using the Reynolds-averaged Navier–Stokes equations on three families of NACA 0012 grids with three sets of operating conditions. Performance statistics are presented in a variety of ways that show the relationships between central processing unit time, grid spacing, and accuracy in ways that are informative for both flow solver users and developers.