Revisions №60396

branch: master 「№60396」
Commited by: Vikram K. Mulligan
GitHub commit link: 「89f026f6774b0d60」「№653」
Difference from previous tested commit: code diff
Commit date: 2018-09-12 12:14:12

Merge pull request #653 from RosettaCommons/vmullig/npro_not_a_donor Removes DONOR property from Npro atom type The Npro atom type was incorrectly listed as a hydrogen bond donor. This results in its Lennard-Jones radius being shrunk, and bad things happening in protocols that iterate over donors and whatnot. We've had the pull request to correct this open for three years, and have agreed at more than one meeting that it should be merged. Frank confirmed in 2016 that while the change does have a small effect on numerical values of scores, it doesn't alter his scientific benchmarks. It seems that everyone's on board with this being merged, so here we go. ***SCORING TESTS WILL SHOW SMALL NUMERICAL CHANGES*** Virtually all integration test trajectories are likely to change. Prepare for lots of observer e-mails as integration tests change. This is expected.

Rocco Moretti 6 years
I don't know if this commit is directly responsible, but it looks like this commit is the first one which causes the mac.clang.integration hotspot_hashing test to consistently hard-fail with a timeout error. It's probably worth looking into if this change substantially alters the runtime of this test (and if so, why and if there's an easy way to correct it.)

Vikram K. Mulligan 6 years
Yes, this change likely resulted in small trajectory changes that are resulting in a big change in runtime (possibly because some filter isn't being satisfied as soon or something like that). The best solution might just be to play with the random seed for the test. [list]

Rocco Moretti 6 years
It also killed the linux.clang.performance.protocols_optimization_Minimizer_dfpmin_armijo_nonmonotone performance test -- it's running about a third of what it once was (consistently in future releases, too). Is that just because of the particular structure being tested, or is that more of a reflection in a global slowdown of minimization?

Vikram K. Mulligan 6 years
Since the only thing that has changed is a database change (removing DONOR from the Npro atom type, with all other changes being corrections to the expected outputs in unit tests), I don't see how this could have fundamentally changed the minimizer. It must be the particular structure being tested. [list]

Rocco Moretti 6 years
Over the weekend I ran a runtime length test across various proteins. On the whole there doesn't seem to be any change pre/post commit, so it does look like it may be just a quirk of the particular structure being tested.

Summary

...