Mark Abraham wrote:
mrshi...@gmail.com wrote:
What other input might you need for a test set? As a minor developer
and a stickler for accuracy, I would be very much interested in the
sorts of inputs your looking for, and have some ideas as well.
I'd be willing to help, as well.
There's a number of issues listed in the posts in the URLs below, chief
among them the absence of a bug-free reference version of GROMACS. The
other issues mostly arise because if there's no documentation of what is
being tested *for*, it's hard to do maintenance on the test.
Can we guarantee any version will ever be bug-free? Or should there just be a
test set for every version such that the tests succeed given the inherent
limitations or potential bugs in the software? This could get a bit laborious,
re-creating reference data for every version, but might be the most thorough way
to proceed.
There's currently no tests designed for GROMACS in parallel. It's far
from clear that there's a suitable reference GROMACS version anyway.
Would it even be possible to design a meaningful parallel set, given the
inherent potential for deviations due to, i.e. dynamic load balancing? Even
mdrun -reprod doesn't completely guarantee reproducibility, does it? Might some
of the more advanced features also depend on the FFT implementation, as well?
One clear need is a mechanism to permit features to be tested in
combination in an automated manner. The set of "complex" tests that
already exist are a good start, but they're far from complete. It should
be possible to ask a script to test thermostats in (X,Y) with barostats
in (W,Z), using -sum/-nosum with the constraint of -npme 0. (This is not
at all silly - I spent several weeks this year proving that I'd found a
GROMACS bug. It transpired that the problem was with the V-rescale
thermostat under -nosum, and I only noticed that because I was using
-rerun!) To avoid combinatorial explosion of the reference data, that
data would have to be generated at the same time as the test data. Thus
the user would need to have installed some known good GROMACS version,
and done some "bootstrap" correctness tests of that against supplied
reference data, before moving on to more complex cases with
user-generated reference data. This requires that the script "know" how
to test each feature, so that it can correctly construct reference and
test runs. The above example is easy - the script knows that to test a
thermostat or barostat, the reference and test .mdp files need to have a
certain form, and testing a command line flag is easier still. The
script would also need to know how to reject tests of mutually-exclusive
features.
That does sound relatively simple to do, but would probably also require a bit
of re-organization in the test set. For example, instead of the four or so
directories now, we'd probably have to expand to substantially more depending on
the features being tested (which also helps in determining what the tests do).
README files are also a must in each directory, similar to the AMBER test set.
In principle, each new feature implemented should be regarded as
incomplete until there's a test that functions correctly. This means
that the author of the feature needs to designate a GROMACS version that
is a suitable reference case (e.g. you can't test V-rescale against a
3.x reference version because it wasn't implemented back then!) That
becomes rapidly untenable for a user of the test suite, since they would
have to have access to multiple different versions - there'd have to be
a web server for providing reference data. There's further complications
if testing feature A (whose reference version is 4.0.2) in combination
with feature B (whose reference version is 4.0.4). Clearly you'd have to
use at least 4.0.4 to generate a reference case for A & B together, and
then have to test that A alone in 4.0.4 is correct with respect to 4.0.2.
I don't know how to bring order to this chaos! I do know that the lack
of a solution will continue to cost everyone time and money doing broken
simulations and chasing bugs.
Would it be useful to start a wiki page on the topic, perhaps somewhere within
the development section, sort of like what was once done for features to be
implemented in the main software? That way, there's a central site for listing
ideas, comments, and progress.
-Justin
--
========================================
Justin A. Lemkul
Ph.D. Candidate
ICTAS Doctoral Scholar
MILES-IGERT Trainee
Department of Biochemistry
Virginia Tech
Blacksburg, VA
jalemkul[at]vt.edu | (540) 231-9080
http://www.bevanlab.biochem.vt.edu/Pages/Personal/justin
========================================
--
gmx-users mailing list gmx-users@gromacs.org
http://lists.gromacs.org/mailman/listinfo/gmx-users
Please search the archive at http://www.gromacs.org/search before posting!
Please don't post (un)subscribe requests to the list. Use the
www interface or send it to gmx-users-requ...@gromacs.org.
Can't post? Read http://www.gromacs.org/mailing_lists/users.php