'out of range atoms' error with USER-CUDA

LAMMPS LAMMPS Mailing List Mirror
1

Hello,

I’ve been encountering this type of error often (ex. log attached):

Out of range atom: ID: 74906 x: 1.781575e+02 y: 1.781575e+02 z: 1.781575e+02 nx: 2.200000e+01 ny: 2.600000e+01
ERROR: Out of range atoms - cannot compute PPPMCuda!
Outside Atom: 1203 1.006580e+02 8.964741e+01 1.988439e+02 (6 4 26)
… [list of atoms and coords]

This occurs after many time steps have run, often over 100,000, and happens on almost all runs with multiple cpus and gpus. I have seen it once with one cpu + one gpu.

Things we’ve tried:

  • changing to double-precision: not much difference
  • changing neighbor settings (larger skin, removing delay in rebuilding lists): delays the error
  • chopping run commands into shorter sequences (e.g. changing run 50000 to run 10000 run 10000 run 10000 run 10000 run 10000): delays the error, sometimes significantly

Is there some setting I can change to prevent this error?

Bonus question: Is there a min_style available that’s compatible with user-cuda?

Thanks,
Amanda

log01.lammps (30.9 KB)

out01.pim1 (30.5 KB)

2

Hi Amanda

looking at your script my first guess is that it actually is a physics problem. In the run after timestep 100000 you start to use npt. And your volume is collapsing by almost 90% within 6000 steps (from 27e6 to 3e6).

From my point of view thats way to fast for shrinking the volume and can have all kinds of unintended side effects (such as loosing atoms, which will first register with "Out of range messages" coming from pppm). From my experience a pressure relaxation time of 0.1 ps is too small, and if you are far from equilibrium it tends to create very unstable situations.

That said similar problems occured due to earlier (and in the current version fixed) bugs in my communication routines, so it would be nice if you could send me your input and data files so I can test here as well.

Best regards
Christian

-------- Original-Nachricht --------