AMBER Archive (2005)

Subject: AMBER: Critical PMEMD 8 Bugfix

From: Robert Duke (rduke_at_email.unc.edu)
Date: Mon Jul 11 2005 - 07:38:59 CDT


Folks -
If you use PMEMD 8, please pick up bugfix 46 from the amber.scripps.edu
website (under the link Bugfixes-->8.0 bugfixes), patch pmemd, and rebuild.
This bugfix corrects potential errors in the nonbonded pairlist code caused
by a combination of floating point rounding errors as well as the fact that
the standard amber fractional coordinates algorithm can produce fractional
coordinates in the range of slightly < 0.d0 through 1.d0 (thus there is
redundancy on both ends, which can mess up gridding based on coordinates).
The probability that anyone has actually been, or will be, impacted by this
bug is extremely low, but if you are, you may get anything ranging from a
segment violation to an infinite electrostatic energy to a small error in
the electrostatic energy. The low probability of this event is driven by
the fact that an atom must be exactly on a box boundary for the error to
occur. One might think that is likely, but it is actually extremely
unlikely due to the density of floating point numbers (ie., "exactly on"
means matching completely to something like 15-16 decimal digits). Petr
Kulhanek reported this error and observed it because he was using
nonstandard procedures (not leap) to generate prmtop/inpcrd inputs. Thus,
his models start with at least some atoms directly on the box boundary.
Leap always leaves a void around the box edge to avoid close contact in the
images, and then the probabilty that you land an atom on the box edge is low
in dynamics. Much thanks to Petr for finding this though; he did a very
nice job of analyzing the problem and proposed a pretty good fix. My fix is
a bit different, but essentially the same except that there should not be
ewald bomb scenarios. The performance impact should be negligible. I have
not done a fix for pmemd 3.1 which also has the problem, but if I get
complaints I'll consider it. The same problem also exists in the entire
sander code base, though the symptoms are most probably small energy
differences rather than some sort of disaster due to the use of different
pairlist algorithms. Other folks on the amber team more familiar with the
sander pairlist code are working on a fix.
Best Regards - Bob Duke

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu