Different simulations depending on number of CPU?

Submitted by Baric on
Daniel Queteschiner

Particle insertion depends on the number of processors used, for example, due to optimization reasons. This is expected behaviour.

Thu, 30 Mar 2017 - 09:01 Permalink

Hi everybody,
I could not find anything about this in the forum yet, so maybe I missed it. But I have experienced something weird lately.
When I run the same simulation on 1 or 2 CPU I get different results. And by different I mean slightly different in the trajectory and the forces.
You can check this with the tutorial movingMeshGran. It does not matter how often you run it on a single core, the result is always the same (checked with diff on the dump files, no difference). As soon as I start a parallel run they differ. And not only this, 2 CPU differ from 4 CPU usw..

Example:
In the movingMeshGran, I get 1355 particles left at timestep 30000, while I have 1326 particles left on 2 CPU. So it is not just some last few digits but real differences in the trajectory.

I have also tried using mpirun -np 1 and no mpirun, which does not make any difference.

What can be the reason for this? Does anyone know?

Regards
Valentin