SIG HPC Benchmark
 

Hi Forum, Hi Dominic Spreitz,

[/B]I just saw this wiki entry about the special interest group HPC: http://openfoamwiki.net/index.php/Sig_HPC . Unfortunately I couldn't attend OF Workshop 2009 in Montreal, so that I am not aware of Who is member in this SIG? Does their exist a mailing list? What is the current status on a benchmark case? Please discuss!

Some of the basic questions related to this topic are

1. What is the current fastest machine (best architecture, interconnect) for the code in our case OpenFOAM or Am I buying the right machine? (not only questioned by myself but by the "real" HPC guys doing the call for bids for new huge machines) An answer on that is almost imposible because the is no "The code OpenFOAM" - to many solvers for to many physical problems, to many numerical solvers + settings, to many compiler + settings ... changing OpenFOAM versions!! This makes it hard to make a pure hardware evaluation.

2. Is OpenFOAM code fast compared against other codes (comercial and NONCOMERCIAL) or ? - same problems as in 1.

3. Maybe Question 1. + Question 2. should be reformulated as: How much of the computing power (FLOPS) provided is OF (or better icoFoam [or other solvers] from OF-1.6 with certain settings) able to use? This leads to profiling! Has anybody experience with profiling OF (probably Hrv or somebody else of the core developers!)?

4. Not a question but the hope to create some kind of data or knowledge base about HPC and OF which allows to evaluate the improvements of the code and the hardware.

From my experience running through the process of bidding (providing a benchmark + evaluating it) is:

Keep the benchmark simple (limit the number of test cases to one or maybe two)!
I am going so far to suggest a OpenFOAM-bench fork, where the source code is stripped down to one or two solver and the libraries linked?! This will reduce compile time and compile problems for the different architectures.

We used a benchmark based on http://openfoamwiki.net/index.php/Contrib_benchFoam for the last three years with one conclusion too many cases and too much python related problems. (Sorry Bernhard I think you are doing a great job with pyFOAM :-)

To sum things up and getting started:

I suggest 3D test case taken from "Benchmark Computations of Laminar Flow Around a Cylinder" by M. Schäfer and S. Turek. http://www.mathematik.uni-dortmund.d...rTurek1996.pdf
I would prefer the Test case 3D-1Q with ?Mill cells which is a steady flow around a squared cylinder. But if somebody has some serious convincing doubts about it, we might have to change!
I have one which solver are we using for laminar steady flow: simpleFoam with laminar turbulence model (probably not very efficient compared to other codes) or modify simpleFoam to lamSteadyBenchFoam by kicking out turbulence modelling?

Thanks for the discussion!

Jens