On Mon, Dec 26, 2011 at 12:43, Mark F. Adams <mark.adams at columbia.edu>wrote:
> I've just started using -ksp_rtol 1.e-10 and retested all of these test > and am seeing a very slight improvement going from 1 to 4. I would expect > to see a significant drop, say 25%, in the number of iterations going from > 1 to 2. The < 5% that I am seeing seems too low but I'm not 100% sure -- > maybe there is nothing wrong ... Well, here is a different test, just for giggles. This sets up rapid coarsening so that multigrid convergence with weak smoothers isn't already "too good". $ pwd /home/jed/petsc/src/snes/examples/tutorials $ for n in {1..4}; do echo Cheby $n; ./ex5 -da_grid_x 5 -da_grid_y 5 -da_refine 3 -da_refine_x 4 -da_refine_y 4 -snes_max_it 1 -ksp_monitor -pc_type mg -pc_mg_levels 4 -mg_levels_pc_type jacobi -mg_levels_ksp_type chebychev -mg_levels_ksp_chebychev_estimate_eigenvalues 0,0.1,0,1.1 -mg_levels_ksp_max_it $n; done Cheby 1 0 KSP Residual norm 1.621385151256e+01 1 KSP Residual norm 1.247824515555e+01 2 KSP Residual norm 1.278794763626e+00 3 KSP Residual norm 1.116456717457e-01 4 KSP Residual norm 4.348273343707e-02 5 KSP Residual norm 8.348548286503e-03 6 KSP Residual norm 3.951234403814e-03 7 KSP Residual norm 1.058691144871e-03 8 KSP Residual norm 3.264344380115e-04 9 KSP Residual norm 8.173798438913e-05 Cheby 2 0 KSP Residual norm 1.891099688325e+01 1 KSP Residual norm 9.980070235568e+00 2 KSP Residual norm 8.251956879061e-02 3 KSP Residual norm 8.973973844623e-03 4 KSP Residual norm 6.086189552261e-04 5 KSP Residual norm 8.406518096327e-05 Cheby 3 0 KSP Residual norm 1.990124921919e+01 1 KSP Residual norm 1.111981517647e+01 2 KSP Residual norm 4.888111472342e-02 3 KSP Residual norm 2.109153682849e-03 4 KSP Residual norm 1.680361259338e-04 Cheby 4 0 KSP Residual norm 2.051550954711e+01 1 KSP Residual norm 1.049577156306e+01 2 KSP Residual norm 3.445878799134e-02 3 KSP Residual norm 9.302079492533e-04 4 KSP Residual norm 3.995254526577e-05 I don't think the Cheby implementation is buggy and I think we are using it correctly. I agree that it is somewhat disappointing if increasing the order above 1 doesn't make a more noticeable improvement. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111226/ab3f860c/attachment.html>