[gmx-developers] (ir)reproducibility of tests

[Mark Abraham]

Hi,

Two successive failures of https://gerrit.gromacs.org/#/c/9381/ last night
gave me a lead, because only the second post-submit failure showed
differences in velocities, and those were all near zero.

I found that some of the regressiontest cases have zero initial velocities
and no gen-vel. Some of the cases I've found so far are our usual suspects
for instability (various flex, various ljpme-lb, complex/urea). I can see
how that would lead to poorly reproducible velocity updates of small
numbers, depending exactly how we compare them. But sometimes we only get
the failure at the final step and only for potential energy (at the given
tolerance) - see the first post-submit failure above.

We should probably fix those tests to have sensible velocities. But can we
think of why zero initial velocities would make a trajectory less
reproducible? It does not seem related to the way GPU forces are less
reproducible, for some cpu-only mdruns have failed apparently similarly.

Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20190402/073fa353/attachment.html>