Dear all,
I performed a quick test on that topic. Here is what I did.
Setup:
First, I created modified top files of FactorIX in that way that the
five negative values in the section %FLAG DIHEDRAL_PERIODICITY were made
positive. Both simulation setups, NPT and NVE, from the Benchmark Suite
were used, but I extended the runtime (nstlim) by a factor of 10. Two
simulations for each system were run on two different GPUs (RTX3080,
A100) using Amber20 (with patches 1-8).
Afterwards, I processed the out files with process_mdout.perl to obtain
a list of ETOT values (summary.ETOT).
Results:
1, Binary trajectory and restart files were identical for the two
consecutive runs of the identical system on the same GPU type, as were
the summary.ETOT files. -> Simulations ran successfully, as expected.
2, Binary trajectory and restart files were not identical for the
systems with original and changed top files, as were the summary.ETOT
files. -> Modification of top files affected simulation.
3, Difference
The mean absolute differences between the runs were as follows (400 data
points, orig: original top file, corr: modified top file):
NPT orig vs. corr (A100): ca. 9700 kcal/mol
NPT orig va. corr (RTX3080):ca. 9800 kcal/mol
NVE orig vs. corr (A100): ca. 9900 kcal/mol
NVE orig va. corr (RTX3080):ca. 9900 kcal/mol
NPT-orig (A100 vs RTX3080): ca. 80 kcal/mol
NPT-corr (A100 vs RTX3080): ca. 70 kcal/mol
NVE-orig (A100 vs RTX3080): ca. 20 kcal/mol
NVE-corr (A100 vs RTX3080): ca. 10 kcal/mol
-> Modification of the top files yielded results in ETOT at least ca.
two orders of magnitude larger than differences in MD runs on different
hardware.
=> Changing the sign in the section %FLAG DIHEDRAL_PERIODICITY seems to
be significant.
But I encourage others to do similar checks, or course!
Best regards,
Anselm
Bioinformatik | NHR.FAU
Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)
Germany
Am 25.01.2022 um 23:06 schrieb John Chodera:
> David et al.,
>
> Thanks so much for the quick response!
>
> On Mon, Jan 24, 2022 at 1:43 PM David A Case <david.case.rutgers.edu> wrote:
>
> I think what John has found is one example of the twenty-year-old prmtop
>> file where this rule is not followed.
>>
>
> I suspected this might be case, but wanted to make sure we weren't
> accidentally causing problems for Amber users with valid use cases!
>
> About the factor_IX benchmark: the first simple task (volunteers, please!)
>> would be to manually change the five negative entries to be positive, and
>> see
>> if the results change. If we are lucky, we can just update the prmtop
>> file.
>
>
> I'll give this a try!
>
> Thanks!
>
> Best,
>
> John
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jan 26 2022 - 03:30:02 PST
