Multigrid finite element methods on semi-structured triangular grids

Similar documents
Aspects of Multigrid

University of Illinois at Urbana-Champaign. Multigrid (MG) methods are used to approximate solutions to elliptic partial differential

Kasetsart University Workshop. Multigrid methods: An introduction

Multigrid Methods and their application in CFD

c 2014 Society for Industrial and Applied Mathematics

1. Fast Iterative Solvers of SLE

Finite Element Multigrid Framework for Mimetic Finite Difference Discretizations

Numerical Solution I

Comparison of V-cycle Multigrid Method for Cell-centered Finite Difference on Triangular Meshes

THE EFFECT OF MULTIGRID PARAMETERS IN A 3D HEAT DIFFUSION EQUATION

Efficient implementation of box-relaxation multigrid methods for the poroelasticity problem on semi-structured grids

Geometric Multigrid Methods

INTRODUCTION TO MULTIGRID METHODS

Research Article Evaluation of the Capability of the Multigrid Method in Speeding Up the Convergence of Iterative Methods

Efficient smoothers for all-at-once multigrid methods for Poisson and Stokes control problems

INTERGRID OPERATORS FOR THE CELL CENTERED FINITE DIFFERENCE MULTIGRID ALGORITHM ON RECTANGULAR GRIDS. 1. Introduction

Multigrid absolute value preconditioning

Discrete Projection Methods for Incompressible Fluid Flow Problems and Application to a Fluid-Structure Interaction

Algebraic Multigrid as Solvers and as Preconditioner

Geometric Multigrid Methods for the Helmholtz equations

Compact High Order Finite Difference Stencils for Elliptic Variable Coefficient and Interface Problems

The Effect of the Schedule on the CPU Time for 2D Poisson Equation

New Multigrid Solver Advances in TOPS

Boundary value problems on triangular domains and MKSOR methods

Journal of Computational and Applied Mathematics. Multigrid method for solving convection-diffusion problems with dominant convection

Computational Linear Algebra

arxiv: v2 [math.na] 2 Apr 2018

EFFICIENT MULTIGRID BASED SOLVERS FOR ISOGEOMETRIC ANALYSIS

Stabilization and Acceleration of Algebraic Multigrid Method

MULTIGRID METHODS FOR NONLINEAR PROBLEMS: AN OVERVIEW

Constrained Minimization and Multigrid

Solving PDEs with Multigrid Methods p.1

Numerical Analysis of the Double Porosity Consolidation Model

Algorithms for Scientific Computing

arxiv: v1 [math.na] 6 Nov 2017

arxiv: v1 [math.na] 11 Jul 2011

Chapter Two: Numerical Methods for Elliptic PDEs. 1 Finite Difference Methods for Elliptic PDEs

Solving Symmetric Indefinite Systems with Symmetric Positive Definite Preconditioners

arxiv: v1 [math.na] 15 Mar 2015

Iterative Methods and Multigrid

Adaptive Time Space Discretization for Combustion Problems

Multigrid Acceleration of the Horn-Schunck Algorithm for the Optical Flow Problem

Mathematics Research Report No. MRR 003{96, HIGH RESOLUTION POTENTIAL FLOW METHODS IN OIL EXPLORATION Stephen Roberts 1 and Stephan Matthai 2 3rd Febr

Introduction to Scientific Computing II Multigrid

A greedy strategy for coarse-grid selection

Scientific Computing I

Elliptic Problems / Multigrid. PHY 604: Computational Methods for Physics and Astrophysics II

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C.

Lehrstuhl Informatik V. Lehrstuhl Informatik V. 1. solve weak form of PDE to reduce regularity properties. Lehrstuhl Informatik V

Robust solution of Poisson-like problems with aggregation-based AMG

3D Space Charge Routines: The Software Package MOEVE and FFT Compared

An Introduction of Multigrid Methods for Large-Scale Computation

Chapter 5. Methods for Solving Elliptic Equations

Adaptive algebraic multigrid methods in lattice computations

Spectral element agglomerate AMGe

Multigrid Method for 2D Helmholtz Equation using Higher Order Finite Difference Scheme Accelerated by Krylov Subspace

Introduction to Multigrid Methods

A fast method for the solution of the Helmholtz equation

DELFT UNIVERSITY OF TECHNOLOGY

AMG for a Peta-scale Navier Stokes Code

Multigrid Solution of the Debye-Hückel Equation

Scientific Computing II

Multigrid solvers for equations arising in implicit MHD simulations

Finite Element Methods

Scientific Computing: An Introductory Survey

Poisson Equation in 2D

Partial Differential Equations

FRIEDRICH-ALEXANDER-UNIVERSITÄT ERLANGEN-NÜRNBERG. Lehrstuhl für Informatik 10 (Systemsimulation)

Robust multigrid methods for nonsmooth coecient elliptic linear systems

ANALYSIS AND COMPARISON OF GEOMETRIC AND ALGEBRAIC MULTIGRID FOR CONVECTION-DIFFUSION EQUATIONS

Remarks on the analysis of finite element methods on a Shishkin mesh: are Scott-Zhang interpolants applicable?

Numerical Programming I (for CSE)

INTRODUCTION TO FINITE ELEMENT METHODS

An Accurate Fourier-Spectral Solver for Variable Coefficient Elliptic Equations

In this chapter we study elliptical PDEs. That is, PDEs of the form. 2 u = lots,

Simulating Solid Tumor Growth Using Multigrid Algorithms

arxiv: v1 [math.na] 3 Nov 2018

A PRECONDITIONER FOR THE HELMHOLTZ EQUATION WITH PERFECTLY MATCHED LAYER

A MULTIGRID-BASED SHIFTED-LAPLACIAN PRECONDITIONER FOR A FOURTH-ORDER HELMHOLTZ DISCRETIZATION.

Multilevel Preconditioning of Graph-Laplacians: Polynomial Approximation of the Pivot Blocks Inverses

A Hybrid Method for the Wave Equation. beilina

Efficient numerical solution of the Biot poroelasticity system in multilayered domains

Multigrid and Domain Decomposition Methods for Electrostatics Problems

NUMERICAL ALGORITHMS FOR A SECOND ORDER ELLIPTIC BVP

A SHORT NOTE COMPARING MULTIGRID AND DOMAIN DECOMPOSITION FOR PROTEIN MODELING EQUATIONS

Block-Structured Adaptive Mesh Refinement

Basic Aspects of Discretization

SOLVING MESH EIGENPROBLEMS WITH MULTIGRID EFFICIENCY

Preface to the Second Edition. Preface to the First Edition

Multigrid Method ZHONG-CI SHI. Institute of Computational Mathematics Chinese Academy of Sciences, Beijing, China. Joint work: Xuejun Xu

On Multigrid for Phase Field

c 2004 Society for Industrial and Applied Mathematics

ME Computational Fluid Mechanics Lecture 5

Bootstrap AMG. Kailai Xu. July 12, Stanford University

Optimal multilevel preconditioning of strongly anisotropic problems.part II: non-conforming FEM. p. 1/36

Solving the stochastic steady-state diffusion problem using multigrid

High order, finite volume method, flux conservation, finite element method

FLEXIBLE MULTIPLE SEMICOARSENING FOR THREE-DIMENSIONAL SINGULARLY PERTURBED PROBLEMS

SUPERCONVERGENCE PROPERTIES FOR OPTIMAL CONTROL PROBLEMS DISCRETIZED BY PIECEWISE LINEAR AND DISCONTINUOUS FUNCTIONS

Numerical Solution Techniques in Mechanical and Aerospace Engineering

Transcription:

XXI Congreso de Ecuaciones Diferenciales y Aplicaciones XI Congreso de Matemática Aplicada Ciudad Real, -5 septiembre 009 (pp. 8) Multigrid finite element methods on semi-structured triangular grids F.J. Gaspar, J.L. Gracia, F.J. Lisbona, C. Rodrigo Applied Mathematics Department, University of Zaragoza, Pedro Cerbuna, 50009 Zaragoza, Spain. E-mails: fjgaspar@unizar.es, jlgracia@unizar.es, lisbona@unizar.es, carmenr@unizar.es. Keywords: Geometric multigrid, Fourier analysis, triangular grids Abstract We are interested in the design of efficient geometric multigrid methods on hierarchical triangular grids for problems in two dimensions. Fourier analysis is a well-known useful tool in multigrid for the prediction of two-grid convergence rates which has been used mainly for rectangular grids. This analysis can be extended straightforwardly to triangular grids by using an appropriate expression of the Fourier transform in a new coordinate systems, both in space and frequency variables. With the help of the Fourier Analysis, efficient geometric multigrid methods for the Laplace problem on hierarchical triangular grids are designed. Numerical results show that the Local Fourier Analysis (LFA) predicts with high accuracy the multigrid convergence rates for different geometries.. Introduction Multigrid methods [3, 7, 8] are among the most efficient numerical algorithms for solving the large algebraic linear equation systems arising from discretizations of partial differential equations. In geometric multigrid, a hierarchy of grids must be proposed. For an irregular domain, it is very common to apply a refinement process to an unstructured input grid, such as Bank s algorithm, used in the codes PLTMG [] and KASKADE [5], obtaining a particular hierarchy of globally unstructured grids suitable for use with geometric multigrid. A simpler approach to generating the nested grids consists in carrying out several steps of repeated regular refinement, for example by dividing each triangle into four congruent triangles [4]. An important step in the analysis of PDE problems using finite element methods (FEM) is the construction of the large sparse matrix corresponding to the system of equations to be solved. For discretizations of problems defined on structured grids with constant

F.J. Gaspar, J.L. Gracia, F.J. Lisbona, C. Rodrigo coefficients, explicit assembly of the global matrix for the finite element method is not necessary, and the discrete operator can be implemented using stencil-based operations. For the previously described hierarchical grid, one stencil suffices to represent the discrete operator at nodes inside a triangle of the coarsest grid, and standard assembly process is only used on the coarsest grid. Therefore, this technique is used in this paper since it can be very efficient and is not subject to the same memory limitations as unstructured grid representation. LFA (also called local mode analysis []) is a powerful tool for the quantitative analysis and design of efficient multigrid methods for general problems on rectangular grids. Recently, a generalization to structured triangular grids, which is based on an expression of the Fourier transform in new coordinate systems in space and frequency variables, has been proposed in [6]. In that paper some smoothers (Jacobi, Gauss Seidel, three color and block line) have been analyzed and compared by LFA, the three color smoother turning out to be the best choice for almost equilateral triangles. The organization of the paper is as follows. In Section the way in which Fourier analysis can be extended to multigrid methods for discretizations on regular non-rectangular grids is explained. The proposed relaxation methods are introduced in Section 3. Some Fourier analysis results for Poisson problem are presented in Section 4 in order to justify the choice of the smoothers, and finally, in Section 5 a numerical experiment is performed to show the efficiency of the proposed algorithm.. Fourier analysis on non-orthogonal grids This analysis is based on the multi-dimensional Fourier transform using coordinates in an orthonormal basis of R. We aim for discretizations on triangular grids in the twodimensional case, so the key to accomplishing this is to introduce the two-dimensional Fourier transform using coordinates in non-orthogonal bases fitting the new structure of the grid. Let {e,e } be a unitary basis of R, 0 < γ < π the angle between the vectors of the basis and {e,e } its reciprocal basis, i.e., e i e j = δ ij, i,j, where δ ij is Kronecker s delta. If {e,e } is the canonical basis, we will denote by y = (y,y ), y = (y,y ) and y = (y,y ) the coordinates of a point in the bases {e,e }, {e,e } and {e,e }, respectively. By applying variable changes x = F(x ) and θ = G(θ ) to the usual Fourier transform formula, the Fourier transform with coordinates in a non-orthogonal basis results in û(g(θ )) = sin γ e ig(θ ) F(x ) u(f(x )) dx. π R In a similar way, the back transformation formula is given by u(f(x )) = e ig(θ ) F(x )û(g(θ )) dθ. π sin γ R Since the new bases are reciprocal bases, the inner product G(θ ) F(x ) is given by θ x + θ x. Now using the previous expressions, a discrete Fourier transform for nonrectangular grids can be introduced. Let h = (h,h ) be a grid spacing and G h = {x = (x,x ) x i = k ih i, k i Z, i =,} a uniform infinite grid oriented in the directions e and e.

Multigrid finite element methods on semi-structured triangular grids Now, for a grid function u h, the discrete Fourier transform can be defined by û h (θ ) = h h sin γ π x G h e i(θ x +θ x ) u h (x ), where θ = (θ,θ ) Θ h = ( π/h,π/h ] ( π/h,π/h ] are the coordinates of the point θ e + θ e in the frequency space. Its back Fourier transformation is given by u h (x ) = e i(θ x +θ x ) û h (θ )dθ. () π sin γ Θ h From (), each discrete function u h (x ) with x G h, can be written as a formal linear combination of the discrete exponential functions ϕ h (θ,x ) = e iθ x e iθ x, called Fourier modes, which give rise to the Fourier space, F(G h ) = span{ϕ h (θ, ) θ Θ h }. Due to the fact that the grid and the frequency space are referred to as reciprocal bases, the Fourier modes have a formal expression, in terms of θ and x, similar to those in Cartesian coordinates. Therefore, the Local Fourier analysis on non rectangular grids can be performed straightforwardly. Figure : A regular triangular grid on a fixed coarse triangle T and its extension to an infinite grid. Let T h be a regular triangular grid on a fixed coarse triangle T ; see left picture of Figure. T h is extended to the infinite grid given before, where e and e are unit vectors indicating the direction of two of the edges of T, and such that T h = G h T, see right picture of Figure. Neglecting boundary conditions and/or connections with other neighboring triangles on the coarsest grid, the discrete problem L h u h = f h can be extended to the whole grid G h. It is straightforward to see that the grid functions ϕ h (θ,x ) are formal eigenfunctions of any discrete operator L h which can be given in stencil notation. Using standard coarsening, (H = (H,H ) = (h,h )), high and low frequency components on G h are distinguished in the way that the subset of low frequencies is Θ H = ( π/h,π/h ] ( π/h,π/h ], and the subset of high frequencies is Θ h \ Θ H. From these definitions, LFA smoothing and two-grid analysis can be performed as in rectangular grids, and smoothing factors for the relaxing methods µ, and two-grid convergence factors ρ, which give the asymptotic convergence behavior of the method, can be well defined. 3

F.J. Gaspar, J.L. Gracia, F.J. Lisbona, C. Rodrigo 3. Relaxing methods It is well known that multicolor relaxation procedures are very efficient smoothers for multigrid methods. Besides, they are well suited for parallel computation. Nevertheless, they have the disadvantage that the Fourier components of L h are not eigenfunctions of the relaxation operator, what makes their analysis more complicated. Three-color smoother and some line-wise smoothers are proposed as relaxing methods. These smoothers appear as a natural extension to triangular grids of some smoothers widely used on rectangular grids, as red-black Gauss-Seidel and line-wise relaxations of zebra type. To apply three-color smoother, the grid associated with a fixed refinement level η of a triangle T of the coarsest triangulation, G = {x = (x,x ) x j = k jh j, k j Z, j =,, k = 0,..., η, k = 0,...,k }, () is split into three disjoint subgrids, G i = {x = (x,x ) G x j = k jh j, j =,, k + k = i(mod 3)}, i = 0,,, each of them associated with a different color, as shown in Figure a), so that the unknowns of the same color have no direct connection with each other. The complete three-color smoothing operator is given by the product of three partial operators, S h = Sh S h S0 h. In each partial relaxation step, only the grid points of G i are processed, whereas the remaining points are not treated. For triangular grids, three different zebra smoothers can be defined on a triangle. We will denote them as zebra-red, zebra-black and zebra-green smoothers, since they correspond to each of the vertices of the triangle which can be associated with these colors. In Figure b) the zebra-type smoother corresponding to the red vertex is shown. They consist of two half steps. In the first half-step, odd lines parallel to the edges of the triangle are processed, whereas even lines are relaxed in the second step, in which the updated approximations on the odd lines are used. RED POINTS R BLACK POINTS GREEN POINTS B G a) b) Figure : a) Three-color smoother. b) Zebra-red smoother: approximations at points marked by are updated in the first half-step of the relaxation, those marked by in the second. In order to perform these smoothers, a splitting of the grid G into two different subsets G even and Godd is necessary. For each of the zebra smoothers these subgrids are 4

Multigrid finite element methods on semi-structured triangular grids defined in a different way, and the corresponding distinction between them is specified in Table, where k, and k are the indices of the grid points given in (). Thus, these three Relaxation Zebra-red k even k odd Zebra-black k even k odd Zebra-green k + k even k + k odd G even Table : Characterization of subgrids G even G odd and Godd for different zebra smoothers. smoothers Sh zr, SzB h and Sh zg are defined by the product of two partial operators. For example, if zebra-red smoother is considered, Sh zr = S zr even h S zr odd h where S zr even h is in charge of relaxing the points in G even and SzR odd h is responsible for the points in G odd. These smoothers are preferred to the lexicographic line-wise Gauss-Seidel because in spite of having the same computational cost, their smoothing factors are better than those of lexicographic line-wise relaxations as we will see further on. 4. Fourier analysis results for Laplace operator To analyze the influence of grid-geometry on the properties of multigrid methods, in this section we apply the LFA to discretizations of the Laplace operator by linear finite elements on a regular triangulation of a general triangle. The geometric parameters that we consider are two angles of the triangle, denoted here by α and β and the distance h between them. The components of the coarse grid correction are the standard coarsening, the natural coarse grid discretization, and as transfer operators we consider linear interpolation and its adjoint (I H h = 4 (Ih H) ) as the restriction. In Tables, 3 and 4 we show the smoothing factors µ ν +ν and the two grid convergence factors ρ for triangles with angles α = β = 60 0, α = 90 0, β = 45 0 and α = β = 75 0 respectively, and for different pre smoothing (ν ) and post smoothing (ν ) steps. We also display in these tables the experimentally measured W cycle convergence factors, ρ h, using nine levels of refinement, obtained with a right hand side zero and a random initial guess to avoid round-off errors. We have chosen W cycles to verify two grid convergence factors since they run at similar rates to the two grid rates but at less cost. In the case of Jacobi relaxation, we have used the optimal parameters w = /(5 ) for equilateral triangles, which can be analytically calculated, and the well known w = 0,8 for the five point discretization of the Laplace operator on rectangular grids. For the last triangle, we have performed an LFA showing that the optimal value is w =. From these tables, we can observe that the convergence factors are very well predicted by LFA in all cases. These convergence factors are improved when an equilateral triangle is considered, and we point out the exceptional convergence factors in Table associated with the three-color smoother. On the other hand, the highly satisfactory factors obtained for equilateral triangles worsen in Table 4,where an isosceles triangle with common angle 75 0 is considered, showing that none of these point-wise smoothers is robust over all angles. It can be seen that the 5

F.J. Gaspar, J.L. Gracia, F.J. Lisbona, C. Rodrigo Damped Jacobi Gauss-Seidel Three-colors ν,ν µ ν +ν ρ ρ h µ ν +ν ρ ρ h µ ν +ν ρ ρ h, 0 0.478 0.437 0.447 0.46 0.38 0.35 0.30 0.34 0.36, 0.30 0.7 0.7 0.73 0.4 0.3 0.053 0.039 0.039, 0. 0.0 0.0 0.07 0.070 0.069 0.09 0.05 0.05 Table : LFA results and measured W cycle convergence rates ρ h for equilateral triangles. Damped Jacobi Gauss-Seidel Three-colors ν,ν µ ν +ν ρ ρ h µ ν +ν ρ ρ h µ ν +ν ρ ρ h, 0 0.600 0.600 0.600 0.500 0.483 0.480 0.98 0.85 0.83, 0.360 0.360 0.359 0.50 0.73 0.74 0.089 0.65 0.60, 0.6 0.60 0.59 0.5 0.8 0.8 0.03 0.5 0. Table 3: LFA results and measured W cycle convergence rates ρ h for α = 90 0, β = 45 0. Jacobi Gauss-Seidel Three-colors ν,ν µ ν +ν ρ ρ h µ ν +ν ρ ρ h µ ν +ν ρ ρ h, 0 0.773 0.766 0.766 0.637 0.63 0.67 0.597 0.588 0.587, 0.597 0.600 0.599 0.405 0.393 0.39 0.356 0.346 0.345, 0.46 0.454 0.453 0.58 0.4 0.40 0. 0.03 0.0 Table 4: LFA results and measured W cycle convergence rates ρ h for α = β = 75 0. 6

Multigrid finite element methods on semi-structured triangular grids Equilateral Isosceles (75 o ) Isosceles (85 o ) ν,ν µ ν +ν ρ µ ν +ν ρ µ ν +ν ρ, 0 0.39 0. 0.5 0.097 0.5 0., 0.053 0.04 0.05 0.043 0.05 0.05, 0.033 0.07 0.033 0.07 0.033 0.03 Table 5: LFA results obtained with zebra-type smoothers for different triangles. optimal values correspond to the case of almost equilateral triangles and the convergence factors deteriorates when any of the three angles becomes smaller. To overcome this difficulty we have designed three block line (or coupled) Gauss Seidel smoothers of zebra-type, introduced in Section 3. For almost equilateral triangles the smoothing and two grid factors for the point-wise Gauss Seidel and the zebra-type smoothers are quite similar, and we have obtained different results for non equilateral triangles. Therefore, these smoother perform very well for small values of its corresponding angle. Then, for each acute triangle we can always activate a zebra type smoother to have an optimal two grid convergence factor. This behavior can be seen in Table 5. 5. Numerical experiment We consider the model problem u = f. The right-hand side and the Dirichlet boundary conditions are such that the exact solution is u(x, y) = sin(πx) sin(πy). This problem is solved in an A-shaped domain, as it is shown in Figure 3a), and the coarsest mesh is composed of sixty triangles with different geometries, which are also depicted in the same figure. Nested meshes are constructed by regular refinement and the grid resulting after refining each triangle twice is shown in Figure 3b). Y 0.8 0.6 0.4 0. 0-0. -0.4-0.6 Y 0.8 0.7 0.6 0.5 0.4 0.3 0. 0. 0-0. -0. -0.3-0.4-0.5-0.6 Y 0.8 0.6 0.4 0. 0-0. -0.4-0.6 Three-colorsmoother Block-line smoother -0.4-0. 0 0. 0.4 0.6 0.8 X -0.4-0. 0 0. 0.4 0.6 0.8 X -0.5 0 0.5 X a) b) c) Figure 3: a) Computational domain and coarsest grid, b) Hierarchical grid obtained after two refinement levels, c) Different smoothers used in each triangle of the coarsest grid The considered problem has been discretized with linear finite elements, and the corresponding algebraic linear system has been solved with the geometric multigrid method proposed in previous sections, that is, we choose the more suitable smoother for each 7

F.J. Gaspar, J.L. Gracia, F.J. Lisbona, C. Rodrigo triangle of the coarsest triangulation, as we can see in Figure 3c), depending on the results predicted by LFA. From the local convergence factors predicted by LFA on each triangle, a global convergence factor of 0,44 is predicted by taking into account the worst of them. In order to see the robustness of the multigrid method with respect to the space discretization parameter h, in Figure 4 we show the convergence obtained, with an F(,)- cycle and the multigrid method proposed, for different numbers of refinement levels. The initial guess is taken as u(x, y) = and the stopping criterion is chosen as the maximum residual to be less than 0 6. Besides an asymptotic convergence factor about 0,43 has been obtained with a right hand side zero and a random initial guess. Note that Fourier two grid analysis predicts the convergence factors with a high degree of accuracy. e+008 e+006 6 levels 7 levels 8 levels 9 levels 0 levels 0000 maximum residual 00 0.0 0.000 e-006 e-008 0 5 0 5 0 5 30 35 cycles Figure 4: Multigrid convergence F(,)-cycle for the model problem. Acknowledgements This research was partially supported by the project MEC/FEDER MTM007-6304 and the Diputación General de Aragón. References [] R. Bank, PLTMG: A software package for solving elliptic partial differential equations. Users Guide Version 0.0, Department of Mathematics, University of California, 007. [] A. Brandt, Multi-level adaptive solutions to boundary-value problems, Math. Comput., 3 (977), 333 390. [3] A. Brandt, Multigrid techniques: 984 guide with applications to fluid dynamics, GMD-Studie Nr. 85, Sankt Augustin, Germany, 984. [4] B. Bergen, T. Gradl, F. Hülsemann, U. Ruede, A massively parallel multigrid method for finite elements, Comput. Sci. Eng., 8 (006), 56 6. [5] P. Deuflhard, P. Leinenand, H. Yserentant, Concepts of an adaptive hierarchical finite element code, Impact Comput. Sci. Engrg., (989), 3 35. [6] F.J. Gaspar, J.L. Gracia and F.J. Lisbona, Fourier analysis for multigrid methods on triangular grids, SIAM J. Sci. Comput., 3 (009), 08 0. [7] W. Hackbusch, Multi-grid methods and applications, Springer, Berlin, 985. [8] U. Trottenberg, C.W. Oosterlee, A. Schüller. Multigrid Academic Press, New York, 00. 8