[Dune] How fast is Dune? Benchmark data anyone?

[Robert Kloefkorn]

Dear Christoph,

> Dear Dune developers and users,
> a common argument against Dune by people not using Dune is the
> suspicion that it is slow. It is considered to not be true, but
> has anyone ever measured a real complex simulation and got
> benchmark data? I asked already some participants on the Dune
> user Meeting 2010 in Stuttgart but nobody knew numbers.

> - Are there users who ported their old simulations to Dune and
> where able to compare the same methods with different codes?

there are several benchmarks done in by Martin, Andreas, and myself.
There might be others, I haven't seen them so far.

During the past 5 years I did one test several times. This test is the
solution of the Euler equations, one time on ALUGrid directly and
another time using DUNE with and explicit Finite Volume scheme of first
order. For the grid interface this is a hard test.
I presented this at the DUNE User meeting (as far as I remember you
participated ?!?) The results are published in the second DUNE paper
(http://www.matheon.de/research/show_preprint.asp?action=details&serial=404).
Newer results for the same test are published in my doctoral
dissertation (http://www.freidok.uni-freiburg.de/volltexte/7175/)
chapter 4. The result is, that when using DUNE with a cheep scheme like
an explicit Finite Volume scheme, one looses 10 to 20 % performance. For
more expensive schemes the gap will be smaller.

Martin did tests how much performance is lost, when a grid is wrapped by
another grid implementation adding some features, e.g. GeometryGrid.
These results are available but not published yet. They were also
presented at the DUNE User Meeting. See Martins talk in the wiki.

Andreas does benchmarking all the time when some new implementations are
done, as far as I know there exist no written plots or whatever.

> - So are there data from runs on supercomputers and the reached
> percentage of the theoretical peak performance?

To measure peak performance you can, for example, use the easy-to-use
tool likwid (http://code.google.com/p/likwid/) or other tools like
perfmon. Please tell me what the peak performance of your code is, I'm
really interested.

> - Where there any benchmark problems to compare the results
> speed-wise with results from groups with other codes?

This is not an easy task. Currently we are involved in projects where
this is done. When the results are published a link will be added to the
wiki.

I did compare our DG implementation for the compressible Euler equations
to the very optimized DG implementation of Dennis Diehl (also from our
group). The performance overhead for using DUNE is 	 negligible, using
some fast grid implementation of course.

> If there are some performance proofs out there we should collect
> them in the wiki.

You could please add this to the wiki, thanks in advance.

Regards

Robert

-- 

  Dr. Robert Klöfkorn           <robertk at mathematik.uni-freiburg.de>

  Mathematisches Institut              Tel: +49 (0) 761 203 5631
  Abt. für Angewandte Mathematik       Fax: +49 (0) 761 203 5632
  Universität Freiburg
  Hermann-Herder-Str. 10
  79104 Freiburg

  http://www.mathematik.uni-freiburg.de/IAM/homepages/robertk