Minimization results differ for different numbers of CPUs

LAMMPS LAMMPS Mailing List Mirror
1

Dear Axel and Steve,

Thank you for your quick replies.

I have tried your suggestions, i.e.

  1. increasing the accuracy of the calculations by
  • pair_modify ‘table’
  • ewald criterion up to 10e-8
  • long-range lj potential
  • using the ewald/disp solver
    and 2) eliminating discontinuities at the short-range cutoff in the energetics by
  • pair_modify ‘shift’

Starting from the original structure none of the methods remedied the problem described previously - no matter whether trying 1) and 2) both seperately and combined.

When being close to a local minimum, a difference in the evaluated forces still arises for calculations on different numbers of CPUs.
As illustrated by the attachment, the resulting ‘numerically different’(?) minima do however became very much closer in energy.

As this problem did not occur without long range forces. I still suspect some numerical issues related to parallelism from the KSpace force evaluation.

Would you have any comment on that?
Thanks in advance & all the best
Hendrik

Input+Output.zip (32.5 KB)

2

Minimizing to very high tolerances means the calculation

of successvive steps is subject to round-off issues that
can occur when different numbers of procs are used. Once

you get off a minimization trajectory in a bumpy phase space
and start going down a different minima, you can end up
with a (slightly) different answer. A better Q would be

why does this matter to your calculation, if the energies are

within epsilon of each other?

Steve