On Wed, Oct 9, 2013 at 5:47 PM, Mark F. Adams <[email protected]> wrote:
> > > > (referring to your previous email about the kind of discretization) I'm > > using P_2 finite elements on tetras. The mesh is so small that it might > as > > well be partitioned lexicographically (but in this case it was done by > > Metis) > > > > Well, This seems fine. Higher order elements can confuse AMG sometimes > but the math is decent in hypre and ML and I think it would be about the > same with GAMG with the parameter that I gave you. Something is going > wrong. We are trying to find it but not having any luck. > > There are many parameters, including stupid things like solver parameters > and stupider things like bugs, that can kill an iterative solver and we are > trying to figure out what is wrong here. You should be running with about > 1/10 or 1/100, or even 1/1000, the number of processors that you are using > here. If you eventually want to run in this regime then that is fine but > for debugging it helps to be in a more normal regime, because then you (or > us) can start to see secondary things like super slow flop rates that can > provide a hint of what is going wrong. It is also better to start in > serial so there are less moving parts -- I have to think that these > terrible ML and hypre setup times would go away in serial … not that that > is a solution but it is knowledge. > > Also, keep in mind that MG is better at scaling so for small problems > direct solvers or simple solvers can be faster. Also AMG has significant > setup costs (but they should not be as high as you are seeing) and this > pushes the "cross over" point higher if you are just looking at one solve. > These costs are amortized for multiple solves but these are just the basic > complexity issues with solvers. MG may not always win on Poisson but it > should, just about always. We can run this exact case with SNES ex12. Matt > > Mark > > > Now that at least either GAMG or BoomerAMG are not stuck for small > > matrices I'll try to go bigger, and hopefully won't have to bother you > > again, thanks a lot for all your pointers. > > > > Pierre > > > > PS: about src/ksp/ksp/examples/tutorials/ex56.c, Dirichelet (line 6) > > should read Dirichlet. > > Thanks, > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
