[petsc-dev] [Fortran] MatLoad() error: MatLoad is not supported for type: mpiaij!

[Jed Brown]

On Fri, 3 Sep 2010 12:53:07 -0500, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> On Sep 3, 2010, at 12:52 PM, Jed Brown wrote:
> 
> > On Fri, 3 Sep 2010 11:57:59 -0500, Barry Smith <bsmith at mcs.anl.gov> wrote:
> >> 
> >>  Jed,
> >> 
> >>    Any sorting algorithm that gives good performance is what we need. The particular algorithm class doesn't matter.
> > 
> > The issue is the possibility/probability of hitting a bad-case for
> > quicksort.  Robust quicksorts usually do something different if they
> > detect that pivot choice is going poorly (like do a lot of work to
> > choose a good pivot or switch to heapsort).  Heapsort usually costs
> > slightly more than quicksort, but I think most users would be willing to
> > pay that in exchange for guaranteed O(n log n) performance (as far as I
> > know, sorting is never an especially performance-sensitive kernel in
> > PETSc, but it's a big deal if 1 second turns into hours occasionally
> > (e.g. when the mesh is changed)).
> 
>    Ok, when we add drop tolerance ILU this requires something like a sort and we found the sort was really making the factorization slow.

Okay, I'll keep that in mind and hopefully get around to putting in a
robust sort with good constants.

Jed