yade-dev team mailing list archive

[Bruno Chareyre <bruno.chareyre@xxxxxxxxxxx>]

Hi Anton,
Two problems I suspect:
1/ It needs speed benchmarking. This little change may have a huge
impact on parallel performance, because - if I don't miss something -
each addForce/addTorque trigger the synchronization of the whole
force/torque containers [1], which is not small thing as you can see.
If this is done for each clump it makes 3 nested loop O(N² x numThreads).
If it occures as I think, it should be visible in O.forces.syncCount.
Normally this value is equal to the number of timesteps.
2/ It is not thread-safe. addForce/addTorque are ok in interaction loop
because it is write-only. Here different threads will try to
write-synchronize-read at the same time.

Bruno


[1] https://github.com/yade/trunk/blob/master/core/ForceContainer.hpp#L128


On 16/05/14 11:48, Anton Gladky wrote:
> Dear users/devs, who is working with clumps,
>
> I have done a tiny modification in NewtonIntegrator [1], which
> should fix numerical instability with clumps in a rare cases.
>
> ============================================
> ============================================
> Subject: [PATCH] Do not use getTorqueUnsynced and forces.getForceUnsynced in
>  clumps
>
> It seems both of these commands are unsafe. It causes NaN velocities
> in a very specific cases. From my point of view it is OK, because
> addForceTorqueFromMembers does not modify data itself.
> ---
>  pkg/dem/NewtonIntegrator.cpp | 6 ------
>  1 file changed, 6 deletions(-)
>
> diff --git a/pkg/dem/NewtonIntegrator.cpp b/pkg/dem/NewtonIntegrator.cpp
> index e961bf2..082add8 100644
> --- a/pkg/dem/NewtonIntegrator.cpp
> +++ b/pkg/dem/NewtonIntegrator.cpp
> @@ -146,14 +146,8 @@ void NewtonIntegrator::action()
>   // clumps forces
>   if(b->isClump()) {
>   b->shape->cast<Clump>().addForceTorqueFromMembers(state,scene,f,m);
> - #ifdef YADE_OPENMP
> - //it is safe here, since only one thread will read/write
> - scene->forces.getTorqueUnsynced(id)+=m;
> - scene->forces.getForceUnsynced(id)+=f;
> - #else
>   scene->forces.addTorque(id,m);
>   scene->forces.addForce(id,f);
> - #endif
>   }
>   //in most cases, the initial force on clumps will be zero and next
> line is not changing f and m, but make sure we don't miss something
> (e.g. user defined forces on clumps)
>   f=scene->forces.getForce(id); m=scene->forces.getTorque(id);


-- 
_______________
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
________________