[petsc-dev] defaulting to right preconditioning with GMRES?

[Jed Brown]

Barry Smith <bsmith at mcs.anl.gov> writes:

>    Should we change GMRES to default to right preconditioning? The
>    problem with left is if you get a funky preconditioning you may see
>    “good” convergence in the preconditioned residual norm while the
>    true residual norm goes no where.

For penalty boundary conditions, some interface methods, and similar,
the right-preconditioned norm may drop many orders of magnitude just
producing a compatible solution, before even starting to converge in the
interior.  If unpreconditioned norm is made the default, packages like
libmesh will have to change it in order to run their examples.

Also, when preconditioning works, the preconditioned norm is stronger.

I could go either way.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20140711/2a80163a/attachment.sig>