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" <[email protected]> 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 :-)