[petsc-users] KSP convergence problem

[Mark F. Adams]

On Mar 20, 2013, at 6:14 PM, John Mousel <john.mousel at gmail.com> wrote:

> Can you comment on a GASM type approach to find a solution for the null space? I notice that the null vectors that successfully make the true residual drop are only complicated in a very thin band around the interface. This band is easy to identify using a level set. Other than that, the null space vector has a low frequency variation. My thought was to break the matrix into two sub-matrices, and somehow apply GAMG as a preconditioner on the far matrix, and ILU on the interface-adjacent matrix. Is this dumb or a complete misunderstanding of GASM?
> 

The null space in GAMG is not a projection, it does not have to be exact.  Do you try using the 6 RBM or giving GAMG coordinates?

Also, you might try not smoothing the (-pc_gamg_nsmooths 0).  Unsymetric matrices can work better this way.


> 
> On Wed, Mar 20, 2013 at 5:08 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> 
> On Wed, Mar 20, 2013 at 5:04 PM, John Mousel <john.mousel at gmail.com> wrote:
> I've wanted to scrap this approach for a long time, but moving away from these GFM-type treatments is not a choice that I've been allowed to follow through on for various reasons which are out of my control.
> 
> Unless there are some clever tricks to characterize the null space or to keep preconditioners compatible with the null space, the folks making the decisions might have to reconsider. It doesn't matter how sexy a method looks if it requires a solve and that solve cannot be done efficiently.
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20130320/f0718809/attachment.html>