[petsc-users] Reducing cost of MatSetValues

[John Fettig]

Hi Jed,

On Wed, Jun 1, 2011 at 4:11 PM, Jed Brown <jed at 59a2.org> wrote:
> 2D or 3D? What kind of elements?

This is 3D with linear tets.


> The overhead of the profiling can be nontrivial for small elements. How does
> the run-time change if you deactivate that event?
> How much of the overall time is spent in matrix assembly (your code plus
> MatSetValues)?

I actually started off by profiling the application with HPCToolkit,
and only turned on the PETSc profiling as a sanity check and to send
to this list.  The results are pretty consistent with the profiling
from HPCToolkit.  I ran without -log_summary and it was about 8%
faster.

> It looks like the linear solve is around 18 iterations. For first order
> elements, that is probably too many for matrix-free with a lagged Jacobian
> to pay off. The trade-off is much more favorable for higher order elements,
> or if the preconditioner is better.

There are several linear solves, and some of them do pretty well with
just cg/jacobi.  Perhaps we could move a few of the solves to
matrix-free.

Thanks,
John