1995 · 37 citations · 14 references
ComputingComputational EngineeringMechanical EngineeringComputational Fluid DynamicsFlow VisualizationFluid DynamicsConvergence AnalysisFinite Difference MethodsNavier-stokes EquationsNumerical ConvergenceCompressible Navier-stokes EquationsEigenvaluesMultigrid ConvergenceNavier-stokes SolversPreconditioningPreconditioning TechniquesUpwind Finite Difference SchemesUpwind Schemes
Semi-implicit and AD1 preconditioners based on classic point-Jacobi and line-Jacobi relaxation are analyzed for the linearized 2-D Navier-Stokes equations discretized using three upwind difference operators: firstorder upwinding and two stencils that are second-order accurate. The preconditioners are intended to be combined with optimal-damping multistage updates for construction of relaxation schemes for multigrid. The purpose of each preconditioner is to tightly cluster residual eigenvalues for required Fourier error modes that are identified for both classic multigrid and semi-coarsening multigrid strategies. Analysis is conducted for both the scalar advec tion/diffusion equation and the linearized Navier-Stokes equations. Results of the analysis show that the proposed preconditioners can be successfully applied to cluster residual eigenvalues for all required error modes over a wide range of governing flow parameters for most choices of discretization. In particular, all analyzed cases of consistently discretized residuals and preconditioners produce successful eigenvalue clustering. When first-order upwinding is used to construct the preconditioner and second-order upwinding used in the residual, then one of the second-order stencils produces successful eigenvalue clustering, while the other stencil has inadequate clustering. A numerical test case is presented that demonstrates the preconditioners in a multistage scheme that effectively damps the desired error modes. Introduction and Motivation Multigrid's success a t speeding up convergence rates of structured-grid compressible Euler and Navier-Stokes codes is well documented. However, the true potential of multigrid for these applications has yet to be realized. Typical asymptotic convergence rates for Navier-S tokes Principal Engineer, Senior Member AIAA Copyright @American Institute of Aeronautics and Astronautics, Inc., 1995. All rights reserved. applications remain in the neighborhood of 0.99 reduction per cycle, far from the arbitrarily low rates achieved for elliptic model problems. Among the many hurdles to be overcome in multigrid methods is development of effective relaxat ion methods for damping high-frequency error modes. A promising approach proposed by van Leer, et a1 (1, 2, 3, 41 is to combine a local matrix preconditioner with an optimized multistage Runge-Kutta scheme. Following this same approach, Allmaras [5] analyzed and tested a local matrix preconditioner based on pointJacobi relaxation for its effectiveness a t clustering residual eigenvalues of the compressible 2-D Navier-Stokes equations discretized using upwinding. The analysis demonstrated that over a wide range of governing flow parameters, the preconditioner successfully clusters residual eigenvalues for high-frequency error modes where they can be damped. The analysis also showed that scalar preconditioning performs abysmally in comparison. Damping of high-frequency error modes by local matrix preconditioners have also been studied in Refs. [6, 71. Related preconditioning research for viscous flows can be found in Refs. [8, 91. The present paper is a continuation of the research presented in Ref. [5], extending the analysis to semiimplicit preconditioners for damping of error modes that contain both lowand high-frequency components. We propose using either local matrix preconditioners or semi-implicit preconditioners combined with an optimized multistage Runge-Kutta update to construct relaxation schemes for multigrid methods. The proposed preconditioners are based on classic relaxation methods. The key to success of all these relaxation schemes is the effectiveness of the preconditioner at tightly clustering the residual eigenvalues away from the origin in Fourier space; Runge-Kutta methods have poor damping characteristics for eigenvalues near the origin. The preconditioner must perform its function for all necessary error modes over a wide range of governing flow parameters. The objectives of this paper are first to identify the appropriate error modes to damp and then to analyze the performance of our preconditioners. After identification of the appropriate error modes to damp, analysis proceeds with the linear 2-D NavierStokes equations. Fourier analysis is conducted for each of the proposed preconditioners to determine the envelope of all possible residual eigenvalues. These envelopes are shown to be bounded away from the origin and are compared to the corresponding envelopes for the scalar advec tion/diffusion equation. Finally, numerical tests of the convergence for preconditioned multistage relaxation schemes are performed to corroborate the findings of the Fourier analysis. Both conventional multigrid and the semi-coarsening multigrid of Mulder [lo, 111 are discussed to identify which error modes must be damped on each grid to ensure that all error modes are effectively damped on a t least one grid during the multigrid cycle. For conventional multigrid we propose an alternating direction implicit (ADI) preconditioner for use on all grid levels (except the coarsest). For semi-coarsening multigrid we present the explicit preconditioner (i.e. point-Jacobi or point implicit) of Ref. [5] for use on all grids that can be coarsened in both directions; we also propose semi-implicit preconditioners based on lineJacobi relaxation for grids that can be coarsened in only one direction. Our intent is to show that relaxation schemes based on classic point-Jacobi and line-Jacobi can form the basis of successful multigrid schemes. However, we do not imply that these are the only relaxation schemes that will meet our damping objectives. Variants of GaussSeidel or approximate factorization met hods will likely also prove successful. We do want to suggest that it may not always be necessary to construct special purpose preconditioners from first principles; suitably modified classic relaxation schemes may be adequate for the task. Multigrid Strategies and Fourier Analysis One interpretation of the multigrid process is that error modes are divided into fineand coarse-grid modes and each are damped on different multigrid levels. The distinction between fineand coarse-grid error modes has a convenient representation in Fourier space. For each grid in a multi-level scheme, fine-grid modes are those modes that cannot be resolved without aliasing on coarser grids. Typically, fine-grid modes are high frequency and coarse-grid modes are low frequency. A robust multigrid strategy requires that all finegrid error modes on each grid level must be effectively damped by the relaxation method used on that grid. However, we also want to minimize the amount of work Figure 1: Fourier Space and Frequency Quadrants that is required on the finest grids. This means that we do not want to waste work damping coarse;grid modes that could be more efficiently damped by relaxation on coarser grids. If any error mode is not effectively damped on a t least one grid, then the asymptotic convergence rate of the multigrid method will be dominated by the slow decay of that mode. Consider a 2-D multigrid method where the finest grid is m x n and coarsening is by halving the grid in each direction. A Fourier error mode on the finest grid is given by Cijj = exp[i(iO, + jO,)], where i, j are the grid indices, and i a. The Fourier angles in x and y have the principal domain 8, E [-n, n] and 8, E [-n, n]. For the analysis presented in this paper, we distinguish different regions of Fourier space (Ox, 8,). Figure 1 shows separation of lowand highfrequency modes into four quadrants: quadrant LL contains modes of low frequency in both x and y directions (i.e. 8, E [0, n/2) and Oy E [0, ~ 1 2 ) ) ; quadrant HH contains modes of high-frequency in both x and y (8, E [n/2, n] and 8, E [a/2, n]); quadrant HL contains modes of high frequency in x and low frequency in y (8, E [n/2, n] and Oy f [0, ~ 1 2 ) ) ; and quadrant LH contains modes of low frequency in x and high frequency in y. The discussion presented below applies equally well to negative Fourier angles. For coarse grids, the principal domain of the Fourier angles is reduced. For example, on the x grid constructed by deleting every other grid line in each coordinate direction, only modes in the domain 0, E [-n/2, n/2] and 0, E [-n/2, n/2] can be resolved without aliasing. Thus the resolvable domain on the 7 x 5 grid is equivalent to the LL quadrant of the finest grid. On each coarse grid we can recursively distinguish four lowand high-frequency quadrants in the same manner as done on the finest grid. For example, on the 7 x 5 grid the HH contains the error modes 0, E [n/4, */2] and 8, E [*/4, */2]. Explicit methods (or point-implicit methods for systems such as the Navier-Stokes equations) can be constructed to effectively damp all HH modes as demonstrated in Ref. [5]. Identification of fineand coarse-grid modes with Fourier angle quadrants depends on the coarsening strategy used, as explained in the following sections.
14
Journal of Computational Physics
Journal of Computational Physics · 2014 · 1.2K citations
Review of preconditioning methods for fluid dynamics
Eli Turkel · Applied Numerical Mathematics · 1993 · 280 citations