On Fri, May 11, 2012 at 1:40 PM, David A Case <case.biomaps.rutgers.edu>wrote:
> On Fri, May 11, 2012, Carlos Simmerling wrote:
>
> > I agree with Jason all around, except for the ig=-1. I tend to set the
> seed
> > explicitly just so I can more readily reproduce the simulations. it's not
> > that I think ig=-1 is incorrect, I just feel more comfortable knowing
> that
> > if I run the same input I will get the same output. It does take more
> work.
>
> OK: my $.02 on this question: the program prints out the value of the seed
> that it chose, so you can use this to re-run the job if needed, using the
> printed seed.
>
Just a minor comment here. Reproducibility when ig=-1 is only guaranteed
in serial or if you're using one thread per group in multisander. The same
comment applies to pmemd (except 1 thread per replica is not permitted in
multipmemd).
I realize that the "real" ig taken from the system clock is printed in the
output file (and this value is required to reproduce results in serial or
with 1 thread per replica), but ig==-1 also triggers Langevin
desynchronization now. As part of that, the value of ig is adjusted by
adding "mytaskid" to it.
This behavior is new as of Amber 12 and is not explicit documented (Ross
may correct me here if I'm mistaken). The Amber 12 manual says this under
the ig description:
It is recommended that, unless you specifically desire reproducibility,
that you set
ig=-1 for all runs involving ntt=2 or 3.
HTH,
Jason
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri May 11 2012 - 12:30:04 PDT
