On Fri, Jun 01, 2007, Shawn yy wrote:
> I met a strange problem when I tried to change the sander code. Since I
> want to separate bonded force and nonbonded force in sander, I defined a
> double precision array bfrc and passed this array instead of f (which
> included nonbonded force already) to the subroutine bond, angle and ephi. So
> all the bonded force were stored in bfrc. In order to obtain the total
> force, I just simply add two array bfrc and f. However, this addition of two
> force arrays gives me trouble. I have a testing system: a di-alanine
> peptide solvated with 8 A buffer rectangular TIP3P box. I ran NPT
> simulations using the modified sander and the original sander for 1000
> steps. From the first step, there was already very small difference (1E-16)
> between the total forces on some solute atoms obtained from two runs. The
> differences increased gradually(1E-16 to 9E-16), but they were still found
> on solute atoms only before step 55. After step 55, differences were also
> found on many water molecules. After step 577, small differences on
> nonbonded energies were found in output file (the two run had exactly the
> same energy before step 577). The differences increased with time. At step
> 1000, output files of the two simulations were already very different. The
> potential energy differed by around 9 kcal/mol. What could be the problem?
> Is it induced by overflow/underflow when the bfrc and f array were added
> together? What is the best way if I want to obtain bonded force?
> All my tests were executed on an Intel P4 workstation, RHEL 3.0, kernel
> 2.4.21-32.0.1.ELsmp, Intel fortran compiler 8.0, and Amber 9.
This sounds normal. Please see the discussion on "testing" (section 1.3.4 of
the Amber manual). You would expect to see the same behavior with un-modified
code if you used a different compiler, or different optimization switches,
or different anything.
...dac
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Jun 03 2007 - 06:07:37 PDT
