[gmx-developers] regression tests - git failing?

[Ran Friedman]

Berk Hess wrote:
> I don't agree.
>
> Of course people run NPT ensembles, but we should have only three two
> tests for this:
> one NVT and one NPT test.
> We want tests to diagnose a problem, not having all test sets fail with
> virial issues
> because I changed a detail in the Berendsen thermostat.
>   
Are these tests meant only for the development stage?
If so you're right. If all tests fail it can be meaningless.

For production, one has to make sure that simulations do work with the
usual parameters and give the correct results. There were many posts to
the mailing lists from people who compared past and present results and
got pretty large deviations.
> What we do want is complex tests sets that test nearly everything at the
> same time:
> NPT, twin-range, position restaints, free energy, etc, to make sure that
> there are no problems
> when combining algorithms.
>
> Copying test set input files is a very bad idea, these test sets are
> made to test parts of the code,
> not at all to produce physically meaningful output. The test sets are
> surely the worst place
> to copy input files from.
>   
Sure, that's a bad practice. I'm not sure if it's avoidable though
(unless you make the tests difficult to access).
Ran.