yade-dev team mailing list archive

[Janek Kozicki <janek_listy@xxxxx>]

Bruno Chareyre said:     (by the date of Thu, 03 Jun 2010 21:48:25 +0200)

> 
> 
> > About 2000 spheres. Each clump is about 4 or 6 spheres. See attached
> > files. The .py files aren't in fact the "just run it" scripts :/
> > Becasue the Triaxial test from Filegenerator works for me best. When
> > run from triax-basic.py - something is different. And I didn't bother
> > to find out what. So I just used bodies generated from
> > triax-basic-with-clups.py and pasted them into .xml file generated
> > from FileGenerator. So the .xml files are "just run it". But remember
> > to update to HEAD.
> >
> >   
> Thanks for explanation!


I'm not an expert with triaxial test! You are. Question that you
can help to answer is whether eigen vs. wm3 provide correct result ;-)
Especially for clumps, because there were some clumps problems with eigen.

I'm worried that one day we will once again see a clump rotating
wildly with no apparent force that could cause this.


> Yes, I experimented this multithread variability in different situations.

I'm not bothered by the fact that different order of performing
calculations leads to different results. That's normal. So I don't
care that curves are different when jobs is not 1. As long is they
are within a certain range, of course. To make sure that this is the
case, we could run 1000 simulations using some python script, plot
all the curves, and see that all of them are in fact generating a
single wider curve that looks coherent. If the eigen - wm3 difference
also fits within this wider curve, then I would assume that it's all
correct.

So I think that running 1000 tests on jobs=2,3,4,5 with both wm3 and
eigen from a SINGLE .xml file, and checking if all those 2000 curves
are similar to each other is the simplest way to verify correctness
for now. I suppose that Vaclav could whip out such python script in
no time ;) For me it would be longer to do :)

-- 
Janek Kozicki                               http://janek.kozicki.pl/  |