yade-dev team mailing list archive

[Václav Šmilauer <eu@xxxxxxxx>]

Thanks everybody for opinions.

Christian, you were right: I ran a non-trivial simulation (65k steps,
perhaps 100k particles) and indeed *the results are exactly the same
with OpenMP-less or OpenMP/1thread, and differ with 2 threads (and
presumably more).*

For illustration of the scatter, attached is resulting PSD from 5
simulations run (4 sieve curves -- feed is the one ending in 1 (which
also differs slightly, because the timespan is determined from
simulation conditions), the other ones are fractions segregated by the
machine), where the scatter is perhaps 4%. That's not little. But I will
have to live with that, I guess.

What I realize is that in the model I use, there are state changes
(increments in some parameters) in particles, which are done from the
contact law and protected by mutexes, but in that case different
contacts will see unpredictably different states of the particle, which
could make this problem more visible.

I will perhaps experiment with the "deterministic" behavior -- it would
mean accumulate forces in deterministic order (after the contact loop).
I will report if I have some essential findings.

Cheers, Václav

> Yet another example of scattering: crack propagation in mode I+III (attached image). The crack pattern at the surface of the model (the yellow symbols) is not the same for different parallel runs of the same simulation.
> The horizontal distance (along Y directions) of surface crack lines is significantly affected (but not the orientations). I faced probably a lot of skepticism in 3SR lab with that, but it is a fact for me.
>
> Another example all of you may know: the post-peak behaviour (stress-strain curve) of a triaxial simulation, but here the differences between two parallel runs are clearly negligible, and the physical explanation is direct considering the peak stress as a bifurcation point.
>
>
> Jerome
>

Attachment: combi-rep-num.pdf
Description: Adobe PDF document