Max,
In petsc-dev it will scale both the mat and pmat by the diagonal of the
pmat if they are different so it appears the documentation is simply out of
date. Note that -ksp_monitor and the various convergence tests are run in the
scaled system so there is no certainty that convergence in this scaled system
results in the same amount of convergence as in the original system, though for
reasonably conditioned systems it should be fine.
In other words if the condition number of your systems is 10^14 this will
not help you get a better convergence, you had better use quad precision
available with that latest gcc compilers and petsc-dev
Barry
On Mar 19, 2012, at 5:06 PM, Max Rudolph wrote:
> I was looking at the documentation for KSPSetDiagonalScale and it
> contains this comment:
>
> "This routine is only used if the matrix and preconditioner matrix are
> the same thing."
>
> However, I have found that even if the matrix and preconditioner are
> different, using the -ksp_diagonal_scale command line argument changes
> the convergence behavior of the solvers. Is the man page correct?
> Also, can you point me towards documentation of the type of scaling
> that is applied?
>
> Thanks for your help,
> Max Rudolph
