HTTPS clone URL
Subversion checkout URL
Experiments with matrix-free Chebyshev collocation for nonlinear elliptic/Stokes problems with fully iterative solutions.
C++ C Shell
Fetching latest commit...
Cannot retrieve the latest commit at this time.
|Failed to load latest commit information.|
Experiments with matrix-free Chebyshev collocation for nonlinear elliptic/Stokes problems with fully iterative solutions. You will need FFTW and a recent PETSc to build. If you are not using petsc-dev, switch the include line in the Makefile. If you do not have CppAD, define WITH_CPPAD to be 0 in stokes.C. The numerical method is spectral collocation on the Chebyshev-Gauss-Lobatto nodes. Derivatives are applied by DCT using FFTW. For preconditioning, a low order method (finite difference or finite element) on the collocation nodes is used to produce a sparse matrix which captures the spectral properties of the high order (dense, thus not actually formed) matrix. If a strong preconditioner (for instance -pc_type lu or -pc_type hypre) is used on this sparse matrix, the number of required iterations is nearly independent of polynomial order. Thus, for a 3D problem with polynomial order P in each cartesion direction, there are P^3 nodes and O(P^3 log P) operations required when using multigrid preconditioning. This is optimal for a spectral method. The scalar elliptic implementation is truly arbitrary dimensional. For instance ./elliptic -dim 12,12,12,12,12 -pc_type hypre -exact 2 -ksp_monitor -ksp_rtol 1e-10 solves the Poisson problem in 5 dimensions with multigrid preconditioning, comparing to an exact solution with inhomogenous Dirichlet data. For the Stokes problem, the we must precondition the saddle point problem. This is done by approximately solving with the Schur complement (a block LU decomposition). There are KSP objects to control. The top level will typically need to be FGMRES since the preconditioner is usually not linear. The Schur complement is controlled by -schur_ksp_xxx. For slow flow problems, the Schur complement is usually preconditioned with the mass matrix. For a collocation method, the mass matrix is the identity, so we do not precondition it. Perhaps we can do something clever for the power law rheology which increases the condition number. Part of the block LU involves a solve with the (viscous) velocity problem. This is controlled by -vel_xx. The preconditioner matrix is the sparse approximation. Typically, we will use AMG and a few iterations (~5) of GMRES on this problem. Applying the Schur complement also involves an (approximate) solve with the velocity system. This solve is controlled by -svel_xx. Here, it is less beneficial to do a good solve, so we may do just one V-cycle on the sparse approximation. Putting this all together, we have something like: ./stokes -exact 2 -cont0 1 -schur_ksp_max_it 3 -vel_ksp_max_it 4 -vel_pc_type hypre -svel_ksp_type preonly -svel_pc_type hypre -ksp_type fgmres -ksp_monitor -dim 20,20,20 -ksp_rtol 1e-10 When the power law rheology has very weak regularization, we must do a continuation for the Newton iteration to converge. The option -cont0 <n> gives the starting value of the continuation (usually 0 to solve the linear problem) and -cont <n> gives the final value. Use -rheology 1 to use power law rheology. Note that the error for exact solutions no longer applies since they are for linear viscosity. ./stokes -exact 2 -cont 4 -rheology 1 -eps 1e-4 -exponent 3 -schur_ksp_max_it 3 -vel_ksp_max_it 4 -vel_pc_type hypre -svel_ksp_type preonly -svel_pc_type hypre -ksp_type fgmres -ksp_monitor -snes_monitor -dim 20,20,20 For the preconditioning matrix, there are a few options, although the simple finite difference poisson-type preconditioner seems pretty good. I suspect the Q1 finite element version will do better with a high order Galerkin method. Interestingly, the finite difference via automatic differentiation does not typically work better than the simple version. Also, there seems to be a bug in CppAD which makes it break on certain problems. Note that just subsampling the spectral matrix with specified stencil provides a very poor preconditioner. Choose velocity preconditioning matrices with -pcvel <n>. The current approach for Neumann and mixed boundary conditions is broken/incomplete. While the Neumann condition works, it destroys the conditioning of the system so convergence is slow. I am not sure that the mixed condition is even correct and convergence is terrible. A suitable outflow boundary is also needed, but this is not implemented.