[petsc-dev] PetscHashIJ scaling problem?

[Matthew Knepley]

On Mon, Nov 11, 2013 at 4:20 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:

> Matthew Knepley <knepley at gmail.com> writes:
> > I did not check, but it would be interesting to check 3D since that it 4
> > part keys.
>
> That should just change the cost of the hash function, so I wouldn't
> expect a big difference.
>
> > It by far the leading cost in mesh setup, but matrix preallocation is
> > 3x more expensive right now.
>
> Okay, the brain-dead way to make preallocation easy is to use a hash to
> store (row,column).  Then the user does a dummy assembly (MatSetValues)
> that adds to the hash (and could discard the values).  I used to want a
> more elegant solution, but the perfect is the enemy of the good and this
> would be perfectly sufficient and users would love it because they
> wouldn't have to think about preallocation.  It uses a bit more memory,
> but preallocation is usually before preconditioner setup and if we leave
> the values out, the overhead is manageable.
>

That would certainly be simpler, but I don't think it would be any cheaper
than what
I am doing now.

We should have the Mat have this mode, however. It would make it easier on
people.

    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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20131111/6d8b0350/attachment.html>