Matthew Knepley

Thanks @amneetb ! I will send you a writeup of it. I would also be interested in talking about a solver. Right now, I am fighting it out with implicit...

Isn't the level 0 solve the coarse level? It should be LU.

What kind of boundary conditions are there? Is the coefficient uniform?

Can you give me -ksp_view for the bottom solver? AMG or GMG for constant coefficient Poisson should converge in ~10 iterates, and for variable it should not be more than...

@boyceg is right. The coarse solver should be either - (SuperLU_dist, MUMPS), or - GMRES/(GAMG, Hypre, ML) Since they are still pretty small, I would lean towards a parallel direct...

1) By slow, you must be referring to setup time, since apply a V-cycle looks like a matvec, and thus 20 matvecs beat 100 handily. GAMG/ML setup is much faster...

@amneetb Some replies: Direct Solver: I would love to see times. We have seen the direct beat MG for problems < 50K Tolerances: Can I see the outer convergence for...

I cannot be understanding this right, because nothing makes sense: 1) You have a V-cycle and only solve the coarse grid weakly (Hypre Max iteration = 1), but you get...

@amneetb Good explanation. A few more questions: 1) Can we turn off the coefficient, so that it is just Poisson? Or it integral to the formulation? 2) Can we solve...

I have said before that there is a function for this: https://petsc.org/main/docs/manualpages/DMPLEX/DMPlexMarkBoundaryFaces.html