yade-users team mailing list archive

[Bruno Chareyre <bruno.chareyre@xxxxxxxxxxxxxxx>]

Dear all,
Another closely related bug has been fixed [1].
In short, it was possible to get different results in terms of overlaps 
for the same pair due to round-off errors. Some parts of the code were 
saying "yes" (the sorting phase based on wrapped coordinates)  while 
others were saying "no" (spatialOverlapPeri() based on absolute 
coordinates), hence breaking the detection logic.
It was happening after billions of particle*iteration, with particles 
traveling away by 1000+ times the period size (definitely increasing the 
round-off errors).  A tolerance is introduced in  [1] to avoid this 
situation.
Again an update is suggested for people using periodic BCs.
Regards
Bruno

[1] 
https://github.com/yade/trunk/commit/9a255743de1dbdad80de265b79aa21950a3decc7

On 04/14/2017 12:53 PM, Bruno Chareyre wrote:
> (forwarded to the intended recipient yade-users)
>
>
> -------- Forwarded Message --------
> Subject: 	Critical bugfix for periodic boundaries
> Date: 	Fri, 14 Apr 2017 12:28:55 +0200
> From: 	Bruno Chareyre <bruno.chareyre@xxxxxxxxxxxxxxx>
> To: 	yade-dev <yade-dev@xxxxxxxxxxxxxxxxxxx>
> : 	
>
>
>
> Hello all,
> A recent commit [1] fixed a critical bug of contact detection in
> periodic boundary conditions. All simulations with periodic conditions
> are potentially affected.
> I strongly suggest an update for everyone using periodic boundaries.
>
> I know personally at least four users who suffered from this bug over
> the last six years, I imagine many more have had the same problem
> without understanding it.
> Typically, a periodic simulation would crash in a (seemingly)
> non-deterministic way, after hours or days of simulations. It was
> actually the consequence of having one or more particle inside another,
> because the contact between them was missed (hence if you have
> unexplained crashes, stop the simulation right before it and check if
> you find particles overlapping too much).
>
> The logic of periodic sorting is so involved (I'm glad Vaçlav invented
> it) that I don't dare trying to explain the problem in details. The
> short story: in contrast with the non-periodic case sorting elements of
> a periodic ring can't be done in N steps (at least with our algorithm, N
> being the number of elements). As it was the case before the fix the
> list of positions was left partially un-ordered.
>
> Cheers.
>
> Bruno
>
> [1]
> https://github.com/yade/trunk/commit/c7c8e6f62d452c81a31415f05a12587a6cc8c452
>
>
>
> --
>
> _______________
> Bruno Chareyre
> Associate Professor
> ENSE³ - Grenoble INP
> Lab. 3SR
> BP 53
> 38041 Grenoble cedex 9
> Tél : +33 4 56 52 86 21
> Fax : +33 4 76 82 70 43
> ________________
>
> Email too brief?
> Here's why!http://emailcharter.org

--
_______________
Bruno Chareyre
Associate Professor
ENSE³ - Grenoble INP
Lab. 3SR
BP 53
38041 Grenoble cedex 9
Tél : +33 4 56 52 86 21
Fax : +33 4 76 82 70 43
________________

Email too brief?
Here's why! http://emailcharter.org