John, try that. Will this be smart enough to filter out very small values?...
On Mar 30, 2012, at 11:24 AM, Jed Brown wrote: > -mg_coarse_pc_type svd? > > (Use redundant for parallel.) > > On Mar 30, 2012 9:21 AM, "Mark F. Adams" <mark.adams at columbia.edu> wrote: > > On Mar 30, 2012, at 10:52 AM, John Mousel wrote: > >> Mark, >> >> I've run GAMG twice with different coarse grid sizes of 2 and 8 with 1 sweep >> of SOR on the coarse grid. For a size of 8 it converges nicely, but for a >> size of 2, I think the null space is causing too many problems. > > YEs, the iterative method is seeing the null space because of floating point > error. > >> If GAMG were to coarsen to a size of 1, then there would be no hope because >> only the null space would remain, right? This doesn't ever seem to occur >> with ML because there are at least as many rows as processors. > > Yes that seems like a good assumption. The right thing to do here would > probably be to do and SVD and filter out the very low modes explicitly. For > now I guess tweaking -pc_gamg_coarse_eq_limit n is al that can be done. Not > very satisfying. We will think about this ... any thoughts anyone? > > Mark > >> >> John >> >> On Fri, Mar 30, 2012 at 8:42 AM, Mark F. Adams <mark.adams at columbia.edu> >> wrote: >> >> On Mar 29, 2012, at 2:40 PM, John Mousel wrote: >> >>> I'm attempting to solve a non-symmetric discretization of a 3D Poisson >>> problem. The problem is singular. I've attached the results of KSPView from >>> runs with ML and GAMG. When I run ML, I get convergence in 30 iterations. >>> When I attempt to use the same settings with GAMG, I'm not getting >>> convergence at all. The two things I notice are: >>> >>> 1. GAMG is using KSPType preonly, even though I've set it to be Richardson >>> in my command line options. >> >> PETSc seems to switch the coarse grid solver to GMRES in Setup. This seems >> to be a bug and I unwisely decide to override this manually. I will undo >> this in the next checkin. This should not be the problem however. >> >>> 2. ML only coarsens down to 4 rows while GAMG coarsens to 2. My problem is >>> singular, and whenever I try to use LU, I get zero pivot problems. To >>> mitigate this, I've been using Richardson with SOR on the coarse matrix. >>> Could the smaller coarse grid size of GAMG be causing problems with SOR. If >>> so, is there a way to put a lower limit on the coarse grid size? >>> >> >> I'm thinking that with a 2x2 coarse grid 8 iterations of SOR is picking up >> the null space. Maybe try just one SOR iteration on the coarse grid. >> >> Also, can you run with options_left so that I can see your arguments. One >> known bug is the mat_diagaonal_scale breaks GAMG, but it should also break >> ML. >> >> Mark >> >>> John >>> >>> On Thu, Mar 29, 2012 at 11:03 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote: >>> On Thu, Mar 29, 2012 at 09:18, John Mousel <john.mousel at gmail.com> wrote: >>> [0]PETSC ERROR: Error in external library! >>> [0]PETSC ERROR: Cannot disable floating point exceptions! >>> >>> Looks like something is strange with your environment because fesetenv() is >>> returning an error. I have disabled the call if the trap mode is not >>> changing. >>> >>> http://petsc.cs.iit.edu/petsc/petsc-dev/rev/352b4c19e451 >>> >>> <KSPView_GAMG.txt><KSPView_ML.txt> >> >> >> <KSPView_ML.txt><KSPView_GAMG.txt> > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120330/3e509600/attachment.htm>
