Performance degradation after upgrade to 3.0.0

[Matthew Knepley]

On Tue, Feb 3, 2009 at 4:35 PM, Billy Araujo <billyaraujo at gmail.com> wrote:
> "Can you send the matrix and rhs as PETSc binary output, and I will find out
> what the difference is. It looks to me like the default ICC ordering
> has changed.

Okay, I have the explanation. The default factorization options
changed in 3.0.0.
No, by default, we use a diagonal shift to remedy zero diagonal elements, like

  -pc_factor_shift_positive_definite TRUE

You do not have a zero per se, but you do have a singular matrix, which you can
check using -ksp_type preonly -pc_type lu, so the last row has a zero diagonal.
Since nothing is ever divided by it, it causes no problems if you do
nothing. However,
shifting generates a much poorer preconditioner. You can recover the original
behavior using

  -pc_factor_shift_positive_definite FALSE

However, I would also advise uncovering the source of the singularity.

  Thanks,

     Matt

-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which
their experiments lead.
-- Norbert Wiener