A couple of notes about the Langevin dynamics codes:
1. version 8.52 of runmd.f fixes a bug that was introduced at version 8.7
(last July). This should have no effect on outputs, but was causing
a major slowdown in parallel Langevin dynamics simulations, at
least on some platforms. RCW reports that Dave Mathews had seen slow
verlet update times for parallel neb runs, and I have seen very slow
verlet update times for regular Langevin dynamics on sgi Altix
machines.
The problem should get worse the more nodes you have, and it may be
worse
on itanium than on pentium. Anyway, I would like to know if this
latest
change to runmd() fixes any problems you might be seeing. Also, you
should be on the lookout for unusually large verlet update times, and
report bad behavior.
[Note: in our implementation, the time spent on procuring random
numbers
for Langevin dynamics stays constant, even as the number of processors
increases. So, this can eventually kill parallelization. If anybody
knows how to *reliably* get around this, it would be good to receive
ideas? Does anyone know how this is handled in CHARMM?]
2. Does anyone understand how amrset() is called in sander? Why was the
argument to amrset() changed from ig to ig+1. It looks to me like the
current code does the initialization twice (once at about linke 747,
then
later at about line 1013). Does anyone understand why we changed from
the behavior that was present in Amber 8?
Received on Wed Apr 05 2006 - 23:49:56 PDT
