On Thu, Jan 18, 2018, Jason Swails wrote:
>
> Everybody breaks stuff ....
[Moving to the developers list:] I agree with Jason's comments. But let's
use this episode as an incentive to be more careful, and to develop better
habits.
In this particular case, there *was* a test in pmemd.cuda with the intent
of quitting with an error message if 12-6-4 prmtops were loaded. But this was
not at the correct place in the code -- I don't know if that was true at the
time of writing, or whether it became the case following subsequent code
drift.
We should work harder to add test cases for erroneous behavior. Right now, we
need a test that 12-6-4 prmtops will indeed by rejected by pmemd.cuda. Then
we will know that the desired behavior is maintained as the code evolves.
The recent update.10 appears to have fixed the problem, but no test was added
to ensure that the fix won't stop working in the future. (And, update.10 has
misleading error messages, which should get fixed for the Amber18 release.)
Related: there are lots of places where we need better input verification.
And, places where we are making progress:
* Scott wrote code to automatically call Junmei's "acdoctor" program
whenever antechamber is run, catching many errors right at the beginning
* cpptraj has mostly evolved (thanks Dan!) to exit when it can't parse
input (and is not in interactive mode)
* Scott is doing the same thing with tleap, which is notorious for
soldiering on even when it doesn't understand the input.
Use your common sense here: there are pages of code in sander and pmemd
looking for inconsistent input. I don't think we need a regression test
case for every one of those, but periodic reviews of that section
of the code would be worthwhile. [I think the 12-6-4 cuda problem arose
because the lj1264 variable was set *after* the corresponding test had been
run.] Finding errors early is important: it gives users much better
error messages than generic "prmtop was not created"-type messages later on.
....dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Jan 19 2018 - 07:00:04 PST
