[petsc-users] Prometheus vs GAMG for elasticity/plasticity problems

[Mark Adams]

>
>
>
> We're seeing a bit lower performance in MatMult with GAMG, perhaps
> because we are not using block formats specialized for elasticity.
>
>
The block info seems to be there.


> Mark, what else is different?


* Prometheus seems to be coarsening slower and taking nearly twice the
iterations.
* GAMG is running slower and the kernels are too.  This would indicate that
the partitions are different (not likely) or the coarse grids are larger,
or something else?



> What does Prometheus do differently in
> setup (not the bottleneck here, but I'm curious).
>

The setup is quite different.  My RAP just does brute force four nested
loops with a lot of unrolling.
My graph setup stuff is more highly optimized in Prometheus in some ways.
 I would think GAMG is doing more work on coarse grids, but it seems to
have fewer of them.  The verbose output should shed some light on this.

These two runs do about the same number of flops and Prometheus is running
a little faster flop rate and is solving a little faster. The GAMG setup is
a lot slower and that has a small effect on the total solve time.

SOR is two iterations is twice as expensive and twice as powerful, roughly.

So GAMG is just running slower.

Prometheus does repartition coarse grids. Perhaps this is a poor network
and/or the initial partitioning is poor so the coarse grid repartitioning
is a big help.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20140130/f3b13039/attachment-0001.html>