Matthew Knepley via petsc-users <petsc-users@mcs.anl.gov> writes: >> 2) I tried all the suggestions mentioned before: setting >> -pc_gamg_agg_nsmooths 0 -pc_gamg_square_graph 10 did not improve my >> convergence. Neither did explicitly setting -mg_coarse_pc_type lu or more >> iterations of richardson/sor. >> > > 1) Can we see some convergence for these? LU HAS to be better than the > alternative, but you are saying that convergence still > stalls above that I think. Okay, but I do not understand how SOR was > working on the smaller problem, and cannot solve the coarse > problem. However, it could be if the interpolation is crap, which it sounds > like is true here. > > Jed, does that makes sense? If so, I would think about doing some bootstrap > stuff.
Justin, can you compute and plot the spectrum of a small problem (<1000 dofs) as well as the eigenvectors of the extremal eigenvalues?