Mark, YOU obtained a matrix with poorly scaled boundary conditions (by a cool factor of 1e13). Some people use penalties which is worse. The preconditioned norm fixes the scaling. We used -ksp_diagonal_scale to actually modify the matrix entries, but that is a bit fragile/restrictive.
I think we would have just as much confusion on petsc-maint/users (from the residual dropping abruptly on the first iteration) if we made unpreconditioned the default. On Nov 10, 2011 7:50 PM, "Barry Smith" <bsmith at mcs.anl.gov> wrote: > > On Nov 10, 2011, at 5:12 PM, Mark F. Adams wrote: > > > > >> > > > > Would this problem also be fixed with -ksp_norm_type unpreconditioned ??? > > > > In that case you could make this the default for debug and not debug, > which is cleaner (I've never figured out why you even offer preconditioned > norm as an option, much less the default) > > > > Someday I'll have to write a paper showing why :-) > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111110/a05489c3/attachment.html>
