From: Mark Abraham (mark.abraham_at_anu.edu.au)
Date: Mon Jul 10 2006 - 18:49:50 CDT
> Hi, all.
>
> From your experience, do you think it is better to run many shorter run
or just one longer run (such as four 2ns runs vs. one 8ns run)?
If your runs have equilibrated and are long enough that you have sampled
enough of the equilibrium distribution, then you will see relevant global
averages converging between the runs. It can be instructive to look at
multiple shorter segments from the same run to see if such convergence is
occuring, but the lengths of time required depend on the number of degrees
of freedom and the roughness of the underlying potential energy surface.
> And do you get the exact same result (trajectory, RMSD, RMSF, etc) from
a repeated run (same seed)? Is the MD repeatable at all in NAMD? I
thought I saw it somewhere in the manual that results may differ. Can
someone confirm that, I just want to make sure. :)
Well in the manual quote you give, it's not reproducible if you don't
choose it to be reproducible... :-) However on identical
hardware/OS/library combinations with user-supplied random number seeds
you will see reproducibility (caveat parallel Langevin). You can test this
for yourself easily on some short runs. Note that even slightly different
hardware/OS/library combinations can be enough to generate numerical
differences, because (for example) the order in which long lists of
floating-point numbers are added can differ, causing different round-off
errors.
> If the result is not reproducible, what would you say? How can we claim
the MD simulation experiment is still reliable???? Any ideas please?
My latter point about the hardware means that as soon as your hardware is
retired, your simulations are not reproducible ever, so your goal is not
to make your simulations long-term reproducible, but to make the
conclusions from them reproducible. See the other replies to this thread
for good advice here :-)
Mark
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:42:19 CST