> On Feb 26, 2016, at 12:46 PM, Mohammad Mirzadeh <[email protected]> wrote: > > Mark, > > On Fri, Feb 26, 2016 at 8:12 AM, Mark Adams <[email protected]> wrote: > > 4-4) Along the same lines, I tried a couple of other PCs such as {jacobi, > sor, gamg, ilu} and none of them were able to converge with bcgs as the KSP. > However, with gmres, almost all of them converge with the exception of gamg. > > Note, I'm not sure why you need the null space of A^T, you want the null > space of A. > > > So the idea was to provide nullspace of A^T to make sure the true residual > also converges to zero by projecting the RHS onto the range of A. It however > looks like that GMRES (and sometimes BiCGSTAB) converge in the least-square > sense for which you only need the nullspace of A and not A^T. > > And for singular systems like yours you need to use a pseudo inverse of the > coarse grid because it is singular -- if you represent the null space exactly. > > GAMG is use for AMR problems like this a lot in BISICLES. > > Thanks for the reference. However, a quick look at their paper suggests they > are using a finite volume discretization which should be symmetric and avoid > all the shenanigans I'm going through! I think it would actually be a good > idea for me to swap my solver with a conservative one and see if it makes > things better. > > > You need to use an 'svd' coarse grid solver, or an appropriate iterative > solver. LU is the default. > > > I see. How can I change the GAMG coarse grid solver? Is there an analogue of > "-pc_hypre_boomeramg_relax_type_coarse"?
-mg_coarse_pc_type svd or maybe -mg_coarse_redundant_pc_type svd or maybe -mg_coarse_pc_type redundant -mg_coarse_redundant_pc_type svd run with -help and grep for coarse for the exact syntax. > > Mark > > Thanks, > Mohammad
