[petsc-dev] design of KSPSPECEST

[Jed Brown]

I think this would work fine. The current design is less intrusive, avoids
code duplication, and would immediately work with a "new" KSP that required
estimates. But I think you are correct that the user interface with
-ksp_chebychev_estimate_ is simpler and thus preferable, even if it has no
less typing.

Jed

On Dec 31, 2010 10:48 AM, "Barry Smith" <bsmith at mcs.anl.gov> wrote:


  It seems to me that KSPSPECEST is a pre-processor for Richardson and
Chebychev that computes "good" parameters using either GMRES or CG during
the "first" solve.

  So why not make the user interface more reflect this, for example:

  -ksp_type chebychev -ksp_chebychev_precompute_parameters
-ksp_chebychev_precompute_ksp_type cg -ksp_chebychev_precompute_minfactor
.95


  Is there something this approach is missing? I think what I propose is
much clearer and straightforward then this business with a new KSP type

  Barry
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20110101/0bf33512/attachment.html>