Benchmark Computations of 3D Laminar Flow Around a Cylinder with CFX, OpenFOAM and FeatFlow

E. Bayraktar , O. Mierka and S. Turek

Institute of Applied Mathematics (LS III), TU Dortmund Vogelpothsweg 87, D-44227, Dortmund, Germany. E-mail: ebayrakt@math.uni-dortmund.de

Abstract Numerically challenging, comprehensive benchmark cases are of great importance for researchers in the eld of CFD. Numerical benchmark cases oer researchers frameworks to quantitatively explore limits of the computational tools and to validate them. Therefore, we focus on simulation of numerically challenging benchmark tests, laminar and transient 3D ows around a cylinder, and aim to establish a new comprehensive benchmark case by doing direct numerical simulations with three distinct CFD software packages. Although the underlying benchmark problems have been dened rstly in 1996, the rst case which was a steady simulation of ow around a cylinder at Re = 20 could be accurately solved rst in 2002 by John. Moreover, there is no precisely determined results for non-stationary case, the simulation of transient ow with time varying Reynolds number. The benchmark problems are studied with three CFD software packages, OpenFOAM, Ansys-CFX and FeatFlow which employ dierent numerical approaches to the discretization of the incompressible Navier-Stokes equations, namely nite volume method, element based nite volume method and nite element method respectively. The rst benchmark test is considered as the necessary condition for the software tools, then they are compared according to their accuracy and performance in the second benchmark test. All the software tools successfully pass the rst test and show well agreeing results for the second case such that the benchmark result was precisely determined. As a main result, the CFD software package with high order nite element approximation has been found to be computationally more ecient and accurate than the ones adopting low order space discretization methods.

Corresponding author

1.

Introduction

Despite all the developments in computer technology and in the eld of numerics, the numerical solution of incompressible Navier-Stokes equations is still very challenging. Efcient and accurate simulation of incompressible ows is very important and prerequisite for the simulation of more complex applications, for instance, polymerization, crystallization or mixing phenomena. Numerous open-source or commercial CFD software packages are available to study these complex applications. Nonetheless, it is often observed that some of these software tools which produce colorful pictures as results of the simulations, fail in some benchmark tests subject to laminar ow around obstacles. Therefore, we are motivated to quantitatively compare performances of well known software packages by studying benchmark problems for laminar ow in 3D. A set of benchmark problems had been dened within a DFG High-Priority Research Program by Sch afer & Turek [1] and these test cases have been studied by many researchers [2, 3]. One of the most studied 3D cases from the mentioned study is the ow around a cylinder with Reynolds number being 20. Although it is a low Reynolds number with steady solution problem and it had been formulated many years ago, the exact solutions could have been determined rst by John in 2002 [3] and later by Braack & Richter [2]. Regarding existence of the very precise results for this benchmark test, we expect the employed software tools to accurately reproduce the results, which is considered as necessary condition for the software packages to continue with solving a second benchmark test for higher Reynolds numbers. The second benchmark problem is unsteady and corresponds to a time-varying Reynolds number. There are not many studies on this benchmark test, one of the most recent studies is from 2005 by John [4]. However, in Johns study the benchmark computation is performed to verify the developed methodology and software rather than improving the benchmark results, and his study is focused on the numerical approaches to the solution of incompressible Navier-Stokes equations. In this study, we aim to establish a new comprehensive benchmark case by doing direct numerical simulations with three distinct CFD software packages. The benchmark problems are studied with the open-source software package OpenFOAM, the widely used commercial code ANSYS-CFX (CFX) and our in-house code FeatFlow. OpenFOAM (version 1.6) is a C++ library used primarily to crate executables, known as applications. The applications fall into two categories: solvers that are each designed to solve a specic problem in continuum mechanics; and utilities that are designed to perform tasks that involve data manipulation (see http://www.openfoam.com). From the available solvers, icoFoam which is designed to solve the incompressible NavierStokes equations with a Finite Volume approach, is employed. CFX (version 12.0 Service Pack 1) is a commercial general purpose uid dynamics program that has been applied to solve wide-ranging uid ow problems with the element based Finite Volume Method. 2

The transient solver of CFX is employed for the benchmark computations. FeatFlow is an open source, multipurpose CFD software package which was rstly developed as a part of the FEAT project by Stefan Turek at the University of Heidelberg in beginning of the 1990s based on the Fortran77 nite element packages FEAT2D and FEAT3D (see http://www.featow.de). FeatFlow is both a user oriented as well as general purpose subroutine system which uses the nite element method (FEM) to treat generalized unstructured quadrilateral (in 2-D) and hexahedral (in 3-D) meshes. Studying benchmark problem with these three dierent software packages which employ dierent numerical techniques also give an insight to answer of the following questions: 1. Can one construct an ecient solver for incompressible ow without employing multigrid components, at least for the pressure Poisson equation? 2. What is the best strategy for time stepping: fully coupled iteration or operator splitting (pressure correction scheme)? 3. Does it pay to use higher order discretization in space or time? These questions are considered to be crucial in the construction of ecient and reliable solvers, particularly in 3D. Every researcher who is involved in developing fast, accurate and ecient ow solvers should be interested. The authors had put their honest eort to obtain the most accurate results with the most optimal settings for all the codes; nevertheless, this benchmark study is especially meant to motivate future works by other research groups and the presented results are opened to discussion. The paper continues with the benchmark conguration and the denition of comparison criteria. In Section 3, the software packages and the employed numerical approaches are described. The results are presented within the subsequent section and the paper is concluded with a discussion of the results.

2.

Benchmark Congurations

The solvers are tested in two benchmark congurations with an incompressible Newtonian uid whose kinematic viscosity ( ) is equal to 10-3 m2 /s and for which the conservation equations of mass and momentum are written as follows,

u + u u = p + u, t u=0 3

(1)

The Reynolds number is dened as Re = U D/ where U is the mean velocity of the imposed parabolic prole on the inow boundary and D is the diameter of the cylinder. The benchmark geometry and the corresponding 2D mesh at the coarsest level are shown in Figure 1. The 3D mesh is obtained by extruding the 2D mesh in the z direction with 4 layers of cells, however the rst level mesh for the computations is obtained by two successive renements via connecting opposite midpoints of the the coarsest mesh which yields a mesh of 6144 cells, see Figure 2. Our preliminary studies showed that this mesh oers a good balance between accuracy and computational cost.

Figure 1: The geometry and the coarsest level mesh (in 2D). The software tools employ dierent numerical approaches to the discretization in space which leads to dierent numbers of degrees of freedom (DOF) for the same problem. In the case of OpenFOAM and CFX, the numbers of DOF are comparable while FeatFlow has a greater number of DOF for the same mesh due to a high order nite element approximation. Therefore, while comparing the results, reader should keep in mind that for the same computational mesh, FeatFlow has approximately as many DOF as the others have at one level ner grid. The number of DOF is always proportional to number of cells (equivalent to number of vertices for hexahedral meshes with large number of cells) in all CFD packages. The numbers of DOF are presented with respect to the number of 4

Figure 2: The rst and second level meshes, with 6144 and 49152 cells respectively. cells of the corresponding mesh levels in Table 1. ndof = 3 nvt for velocity and ndof = nvt for pressure in case of CFX. ndof = 3 nel for velocity and ndof = nel for pressure in case of OpenFOAM. ndof = 24nvt for velocity and ndof = 4nel for pressure in case of FeatFlow. where ndof , nvt and nel denote the number of DOF, the number of vertices and the number of cells, respectively. The rst benchmark problem is at Re = 20 and has been studied numerically by many research groups, and very accurate results have been presented. The second benchmark problem is unsteady, with a time-varying inow prole which is very challenging and has not been rigorously studied, consequently the results have not yet been precisely determined. For the benchmark problems, no-slip boundary condition is employed on the walls and natural do-nothing boundary conditions are imposed at the outow plane. The inow conditions are set with Um = 0.45m/s and Um = 2.25m/s for the rst and the second benchmark problem with the parabolic velocity proles from equations (2) and (3), respectively,

U (0, y, z ) = 16Um yz (H y )(H z )/H 4 ,

V =W =0

(2)

Levels L1

# of Cells

Table 1: Number of unknowns. Software # of DOF u # of DOF P 21828 18423 174624 160776 147456 1286208 1232400 1179648 9859200 9647136 9437184 77177104 7276 6144 24576 53592 49152 196608 410800 393216 1572864 3215712 3145728 12582912

total # of DOF 29104 24567 199200 214368 196608 1482816 1643200 1572864 11432064 12862848 12582912 89760016

L2

L3

L4

CFX 6144 OF FeatFlow CFX 49152 OF FeatFlow CFX 393216 OF FeatFlow CFX 3145728 OF FeatFlow

U (0, y, z ) = 16Um yzsin(t/8)(H y )(H z )/H 4 ,

V =W =0

(3)

In the benchmark calculations, the dimensionless drag and lift coecients for the cylinder are regarded as the comparison criteria which are calculated according to (4). 2Fw , U 2 DH 2FL U 2 DH

cD =

cL =

(4)

where FD and FL are dened as, vt ny pnx dS, n vt nx pny dS n

FD =

FL =

(5)

with the following notations: surface of cylinder S , normal vector n on S with x and y component nx and ny , tangential velocity vt on S and tangent vector t = (ny , nx , 0).

3.

Used CFD Software Packages

We studied the benchmark problems with three software packages: open-source software package OpenFOAM (OF), commercial code ANSYS-CFX (CFX), and our in-house code FeatFlow. Each software adopts a dierent approach to the discretization of the 6

Navier-Stokes equations. The rst adopts a conventional Finite Volume Method (FVM), ANSYS-CFX is developed on an element based FVM (ebFVM) and the last adopts a high order Galerkin Finite Element approach. Their approaches to the discretization of Navier-Stokes equations are one of the main dierences between the codes. Moreover, their solution approach to the resulting linear and non-linear systems of equations after discretization is another distinction which inuences the computational performance rather than the accuracy of the results. Nevertheless, in the case of our in-house code, the space discretization method is also eective regarding the performance of the solvers, the nite element spaces were chosen such that the linear solvers would be the most ecient. Table 2: Discretization schemes and solver parameters for OF. Type Time schemes Parameters ddtSchemes Value Description CrankNicholson Euler blended (0.5) Crank0.5 Nicholson, improved stability. linear corrected Gauss linear Linear interpolation (central dierencing). Explicit correction. non-orthogonal Gaussian

Interpolation schemes Surface normal gradientschemes

interpolationSchemes snGradSchemes

Gradient schemes gradSchemes Divergence schemes divSchemes Laplacian schemes solvers p

Second order, integration.

laplacianSchemes Gauss linear corrected solver smoother tolerance Solver preconditioner tolerance GAMG Gauss-Siedel 107 PBICG DILU 106

Unbounded, second order, conservative. Geometric-Algebraic Multigrid with Gauss-Siedel smoother. Convergence criteria.(*) Preconditioned bi-conjugate gradient solver with diagonal based incomplete LU preconditioner. (*)

solvers U

(L2 norm of residual normalized by L2 norm of solution, convergence criterion for the linear solver [6].)

OpenFOAM achieves the spatial discretization by using FVM on block structured meshes with Gaussian integration and linear interpolation. From the available techniques, temporal discretization is obtained with equidistant implicit Euler blended Crank-Nicolson time stepping scheme (blending factor = 0.5). Later, pressure and momentum equations are decoupled by using the PISO algorithm [5]. For the solution of the momentum equation, PBiCG is employed. PBiCG is a preconditioned bi-conjugate gradient solver for asymmetric lduMatrices using a run-time selectable preconditioner. DILU preconditioner 7

which is a simplied diagonal-based incomplete LU preconditioner for asymmetric matrices was chosen. Then, the pressure equation is solved by using a Geometric-Algebraic Multigrid solver with a Gauss-Seidel type smoother. The details of the chosen discretization schemes and solver tolerances are presented in Table 2. CFX employs an element based Finite Volume approach to discretize in space and high resolution scheme is chosen for the stabilization of the convective term. Time discretization is achieved by Second Order Backward Euler scheme. Tri-linear nite element based functions are used as interpolation scheme. ANSYS CFX uses a coupled solver, which solves the hydrodynamic equations (for u, v, w, p) as a single system. First, nonlinear equations are linearized (coecient iteration), then these linear equations are solved by an Algebraic Multigrid (AMG) solver. The chosen discretization schemes and solver parameters are given in Table 3. Table 3: Discretization schemes and solver parameters for CFX. Setting Transient scheme Interpolation scheme Pressure interpolation Velocity interpolation Shape function Advection scheme Convergence criteria Residual type Residual target Value Second Order Backward Euler scheme FE shape functions Linear-Linear Trilinear Parametric High resolution Description Second order, unbounded, implicit, conservative time stepping. True tri-linear interpolation for velocity and linear-linear interpolation for pressure. Numerical advection scheme with a calculated blending factor. Maximum value of normalized residuals. For details, see [7].

MAX 5x10 5

Our in-house developed open source uid dynamics software package FeatFlow (here: module PP3D) is a transient 3D Finite Element based code parallelized on the basis of domain decomposition techniques. Velocity and pressure are discretized with the high order Q2/P1 element pair and their solution is obtained via a discrete projection method. Since this pair of elements is quite stable even in case of moderate Reynolds number ows, no additional stabilization of the advection is required which is very distinctive regarding the other software packages which employ certain stabilization schemes. The discretization in time is achieved by the second order Crank-Nicholson method which provides ecient and accurate marching in time together with the implemented adaptive time-step-control. The adopted discrete projection approach gives rise to a subsequent solution of the so called Burgers equations and the Pressure-Poisson equation, which is solved in order to enforce the incompressibility constraint. Both of these equations are solved with geometric multigrid solvers: here, SSOR/SOR and UMFPACK/SOR are employed as the solver/smoother pairs for velocity and pressure, respectively. 8

4.

Results

The benchmark problems studied with all the codes, and while the problems were being studied, the question was not: Is the software tool capable of solving the problem? but How accurate and ecient is the given software tool?. Therefore, to show the accuracy of the employed CFD softwares, the benchmark problem at Re = 20 is rstly studied. The results for this case obtained for 4 dierent mesh levels are given in Table (4). Table 4: Results for steady test case (Re=20), reference values: cD = 6.18533 and cL = 0.009401 [2]. # of Cells L1 6144 L2 49152 L3 393216 L4 3145728 Software CFX OF FeatFlow CFX OF FeatFlow CFX OF FeatFlow CFX OF FeatFlow cD 6.06750 6.13408 6.13973 6.13453 6.19702 6.17433 6.17481 6.19362 6.18260 6.18287 6.18931 6.18465 cL 0.01255 0.01734 0.00956 0.00817 0.01099 0.009381 0.00928 0.01001 0.009387 0.009387 0.00973 0.009397 %Err cD 1.91 0.83 0.74 0.82 0.19 0.18 0.17 0.13 0.04 0.04 0.06 0.01 %Err cL 33 84 1.8 14 17 0.21 1.3 6.5 0.15 0.15 3.5 0.05

The qualitative results were indistinguishable even for the coarsest level calculations for the Re = 20 test case, thus only a sample snapshot of the results which has been obtained on the nest grid by FeatFlow is given in Figure (3). The expected agreement of the results for the rst benchmark test motivated us to go on with the second benchmark problem which is the challenging part of our benchmarking studies. The second benchmark problem has a xed simulation time, T = 8 s, whereas the rst benchmark problem is simulated towards the steady state solution. There is no precise unique denition of the stopping criteria for all the softwares for the rst case. Therefore, while the software tools are compared with respect to their accuracy in both cases, their computational performance has been tested in the later case, as well. The coarse grid computations were performed sequentially and ne grid ones were done in parallel, based on domain decomposition method. Both sequential and parallel computations were performed on identical compute nodes: Dual-core AMD OpteronTM Processor 250 2.4 GHz with 8 GB total memory. To decrease the latency time due to memory bandwidth limitation each partition is submitted to one node. The nodes were interconnected via 1GHz ethernet connection, regarding the number of nodes it is 9

Figure 3: Snapshot for the Re = 20 test case with 3145728 cells. a suciently fast connection. However, it has to be mentioned that the performance of the software packages will increase for higher number of partitions by using inniband connected nodes. In the second benchmark test, the ow is simulated for 8 seconds, a half period of the imposed inow condition. The simulations start with zero inow and zero initial condition at t = 0 s and nish at t = 8 s with zero inow again, see Equation (3). Due to the transient inow condition, adaptive time stepping technique is a good candidate for this problem. However, in preliminary studies with this technique, numerical oscillations are observed in the results which are obtained by CFX and OpenFOAM. The oscillations were nor visible in qualitative results neither in the drag coecient results. However, when the results of a sensitive variable such as cL are plotted, the numerical oscillations appear, see Figure (4). The results are obtained on a mesh with 393216 cells by OpenFOAM for two dierent values of maximum Courant numbers, maxCo, and with dierent tolerance values of linear solvers. Tolerance values of the velocity solver, uTol, are set to 105 or 106 and tolerance values of the pressure solver, pTol, are set to 106 or 107 . Therefore, a xed time step size is used in the simulations although it leads to excess of computational costs. The benchmark calculations are performed on several levels of rened meshes to obtain mesh independent results and to show convergence of the solvers with respect to the mesh size. The chosen time step sizes are maximum values for which the solution is independent of the chosen time step size. The comparison criteria are 10

Figure 4: Numerical oscillations due to adaptive time stepping algorithm (v1: maxCo=0.25, uTol=105 pTol=106 ; v2: maxCo=0.5, uTol=105 pTol=106 ; v3: maxCo=0.5, uTol=106 pTol=107 ). maximum drag coecient and minimum lift coecient. Since the lift coecient is more sensitive than the drag coecient, the lift coecient results are more representative in accuracy. When all the results are considered, it is clear that FeatFlow has the best convergence behavior, namely showing quadratic convergence, with respect to the mesh size. This result was foreseen by the authors due to the employed quadratic nite element functions. And converged results with respect to mesh size are already obtained on the 3rd level mesh by FeatFlow. In Figure 5, it is clearly shown that level 3 and level 4 results are identical and consequently the nest level results are considered as the reference through this study. Moreover, it is worth to mention that these results are the closest to one reported by John [4].

Figure 5: FeatFlow results for the second benchmark: drag coecient (left), lift coefcient(right). The old results for cDmax and cLmin had been given within the intervals, [3.2000, 3.3000] and [0.0020, 0.0040] respectively, by Turek et al. [1], and cDmax has been determined as 3.2968 by John [4]. These old results are not sucient and accurate enough to establish an inclusive benchmark study. Besides, to evaluate the results of benchmark calculations with respect to only cDmax and cLmin is not much elucidating on the accuracy of the results. Hence, additional to the comparison of cDmax and cLmin values, we also compare the 11

Figure 6: A closer look at the Figure 5: drag coecient (left), lift coecient(right). solution regarding L2 Err and L Err which denotes the L2 norm of the error normalized by L2 norm of the reference solution and the L norm of error, respectively, are obtained through the following calculation steps: 1. An equidistant discrete time step is specied for which L2 Err is independent of the step size. 2. The reference solution and the other solutions are interpolated (linear interpolation) to the specied discrete time step. 3. L2 norm and L norm of the dierences between reference solution and other solutions are calculated. Results of the benchmark calculations are plotted in the Figures (59) and values of the comparison criteria, L2 Err, L Err and relative errors due to the comparison criteria are given in Tables 510. In Figure (11), we present the results of the benchmark calculations for the nest level, and it is obvious that the results are in agreement. An interesting nding is that the FeatFlow results at level 2 are as accurate as results of CFX or OpenFOAM at the nest level, see Table (11, 12). Table 5: FeatFlow results. ( : clock time x number of nodes) Case FFL1 FFL2 FFL3 FFL4 # of Cells 6144 49152 393216 3145728 cDmax 3.2207 3.2877 3.2963 3.2978 cLmax cLmin Tstep (s) 0.010 0.010 0.010 0.005 Time (s) 3220 x 2 17300 x 4 35550 x 24 214473 x 48 0.0027 -0.0095 0.0028 -0.010892 0.0028 -0.010992 0.0028 -0.010999

12

Table 6: Error calculations for FeatFlow results. %Err %L2 Err %L Err Case cDmax cLmin cD cL cD cL FFL1 FFL2 FFL3 2.34 0.31 0.05 13.6 0.91 0.09 2.09 0.29 0.06 13.8 1.05 0.28 7.71 1.01 0.23 0.152 0.013 0.003

Figure 7: OpenFOAM results for the second benchmark: drag coecient (left), lift coefcient (right).

Figure 8: A closer look at the Figure 7: drag coecient (left), lift coecient(right). Table 7: OpenFOAM results. ( : clock time x number of nodes) Case OFL2 OFL3 OFL4 5. # of Cells 49152 393216 3145728 cDmax cLmax cLmin Tstep (s) 0.0025 0.0010 0.0005 Time (s) 4850 76300 x 4 593500 x 24 3.3963 0.0029 -0.0128 3.3233 0.0028 -0.0118 3.3038 0.0028 -0.0112

Conclusions

The results obtained by this benchmark computations denitively replace the existing reference results for the second test case. FeatFlow results at mesh level 3 could be 13

Table 8: Error calculations for the OpenFOAM results. Case OFL2 OFL3 OFL4 %Err cDmax cLmin 3.0 0.8 0.2 16 7.3 1.8 %L2 Err cD cL 2.61 0.67 0.16 14.5 5.91 1.47 %L Err cD cL 10.0 2.56 0.61 0.21 0.08 0.02

Figure 9: CFX results for the second benchmark: drag coecient (left), lift coecient(right).

Figure 10: A closer look at the Figure 9: drag coecient (left), lift coecient(right). Table 9: CFX results. ( : clock time x number of nodes) Case CFXL2 CFXL3 CFXL4 # of Cells 49152 393216 3145728 cDmax cLmax cLmin Tstep (s) 0.010 0.005 0.005 Time (s) 22320 61530 x 4 115300 x 24 3.3336 0.0028 -0.0106 3.3334 0.0028 -0.0118 3.3084 0.0028 -0.0113

already considered as mesh independent results and regarding the results obtained at mesh level 4, we can conclude that fully converged solution of the second benchmark test

14

Table 10: Error calculations for the CFX results. Case CFXL2 CFXL3 CFXL4 %Err cDmax cLmin 1.1 1.1 0.3 3.63 7.27 2.73 %L2 Err cD cL 1.52 0.98 0.29 %L Err cD cL 0.10 0.10 0.03

7.81 5.31 6.31 3.75 2.24 1.20

Figure 11: The nest level results for the second benchmark test: drag coecient (left), lift coecient(right).

Figure 12: A closer look at the Figure 11: drag coecient (left), lift coecient(right). is achieved, which has been the primary goal of the study. Moreover, we got an insight to answers of our questions in the beginning of the study with the obtained results. 1. We have observed that multigrid techniques slightly increase the performance of the segregated solvers in the solution of viscous Burgers equation due to the chosen small time steps. However, in the case of coupled solvers or the solution of the pressure-Poisson problem, there is a drastic dierence between the performance of conventional (single grid) iterative methods and multigrid techniques. Consequently, it seems to be impossible to obtain ecient solvers for laminar incompressible ow problems unless suitable multigrid techniques are employed. 15

Table 11: The nest level results and FeatFlow results at level 2. ( : clock time x number of nodes) Case FFL2 FFL4 OFL4 CFXL4 # of Cells 49152 3145728 3145728 3145728 cDmax cLmax cLmin Tstep (s) 0.0100 0.0050 0.0002 0.0050 Time (s) 17300 214473 593500 115300 x x x x 4 48 24 24 3.2877 0.0028 -0.0109 3.2978 0.0028 -0.0110 3.3038 0.0028 -0.0112 3.3084 0.0028 -0.0113

Table 12: Comparison of FeatFlow results at level 2 with OpenFOAM and CFX at level 4. ( : clock time x number of nodes) Case FFL2 OFL4 CFXL4 %Err cDmax cLmin 0.31 0.18 0.32 0.91 1.82 2.73 %L2 Err cD cL %L Err cD cL Time* (s)

0.29 1.05 1.01 0.01 17300 x 4 0.16 1.47 0.61 0.02 593500 x 24 0.29 2.24 1.20 0.03 115300 x 24

2. Fully coupled implicit solvers (CFX) oer the advantage of using larger time step sizes, however, to be able to reach the desired accuracy, they require more nonlinear iterations. Thus, the overall computational cost has not been changed signicantly. In the light of our calculations, there is not much dierence between these two approaches in the solution of unsteady incompressible laminar ows; however, this question requires further investigation. 3. Using higher order discretization schemes in space leads to denser linear system of equations which can be solved more eciently on state of the art computers. And since, FeatFlow is more accurate and ecient in the test cases (see Table (12)), we can conclude that it pays to use higher order discretization in space and time. Regarding the comparison of the software tools, the most prominent conclusion can be drawn form Table (12): FF calculation at level 2 on 4 nodes has a similar accuracy of other codes at level 4 on 24 nodes. While FF requires 5 hours of computation for these calculations, CFX and OF require much more. As a conclusion, in the Re = 20 case, we succeeded to obtain fully converged results with all three softwares, and although the second benchmark test was particularly challenging, a reliable reference solution has been obtained. Regarding the computational performance of the employed software packages, this benchmark should be considered as still open and a motivation for CFD software developers to join. All data les and the corresponding plots of the results obtained through this study can be downloaded from http://www.featow.de/en/benchmarks.html website. 16

Acknowledgements The authors like to thank the German Research Foundation (DFG) for partially supporting the work under grants Sonderforschungsbereich SFB708 (TP B7) and SPP1423 (Tu102/32-1), and Sulzer Innotec, Sulzer Markets and Technology AG for supporting Evren Bayraktar with a doctoral scholarship.

References [1] M. Sch afer and S.Turek, Benchmark computations of laminar ow around a cylinder (With support by F. Durst, E. Krause and R. Rannacher). In E. Hirschel, editor, Flow Simulation with High-Performance Computers II. DFG priority research program results 1993-1995, number 52 in Notes Numer. Fluid Mech., pp.547566. Vieweg, Weisbaden, 1996. [2] M. Braack and T. Richter, Solutions of 3D Navier-Stokes benchmark problems with Adaptive Finite Elements. Computers and Fluids, 2006, 35(4), pp.372392. [3] V.John, Higher order nite element methods and multigrid solvers in a benchmark problem for 3D Navier-Stokes equations, Int. J. Numer. Math. Fluids, 2002, 40, pp.755798. [4] V.John, On the eciency of linearization schemes and coupled multigrid methods in the simulation of a 3D ow around a cylinder, Int. J. Numer. Math. Fluids, 2006, 50, pp. 845862. [5] J.H Ferziger and M.Peric, Computational Methods for Fluid Dynamics, 3rd ed. Springer, 2002. [6] OpenFOAM Users guide (version 1.6), 2009, http://www.openfoam.com/docs/ [7] ANSYS CFX-Solver, Release 10.0:Theory, http://www.ansys.com/cfx

17