yade-dev team mailing list archive

[Václav Šmilauer <eudoxos@xxxxxxxx>]

Dne 29.9.2009 12:31, Bruno Chareyre napsal(a):
> This is very interesting, thank you.
> Do you have ideas on what should be done in Yade to improve parallel 
> performance (or, say, what is currently limiting the number of 
> processors)?
For what I've seen, there is barely improvement for more than 4 cores.

What ESyS does is that they split the domain in many parts (one per 
core) and communicate particle changes with MPI messages, if the 
particle is in an thin overlap zone. Quite logical, but having this 
would really require rewriting all yade.

Second, but that is more performance thing -- I suspect using virtual 
methods allover has performance impact, but I cannot prove that.
> In my last benchmarks, I had almost 70% time on interactions, and I'm 
> still curious to know why this cannot be improved, as it sounds a 
> easy-parallelizable thing.
I am here to assist you, if you need; openMP was relatively 
non-intrusive way; the rest would be more difficult, I suspect. I don't 
know off the top of my head what else to do on that front --  perhaps 
the grid-based collider, but you said you have 70% on interactions 
already... Hard to say without measuring that more precisely. (ESyS guys 
have spent 10+ years on performance however, maybe it is just easier to 
get familiar with what they've done instead of writing it again)

Cheers, Vaclav