[gmx-users] About the binary identical results by restarting from the checkpoint file

[Mark Abraham]

On Sun, Jun 2, 2013 at 10:37 PM, Cuiying Jian <cuiying_jian at hotmail.com>wrote:

> Hi GROMACS Users,
>
> These days, I am testing restarting simulaitions with .cpt files. I
> already set nlist=1 in the .mdp file. I can restart my simulations (which
> are stopped manually) with the following commands (version 4.0.7):
> mpiexec mdrun_s_mpi -v -s md.tpr -cpt 0 -cpi md.cpt -deffnm md -reprod
> -reprod is used to force binary identical simulaitons.
>
> During the restarted simulations, same number of processors are used as
> that in the simulation interrupted. The only case, in which I can get
> binary identical results with those from the continuous simulations (which
> are not stopped manually), is for SPC water molecules. Any other molecules
> (like -heptane), I can never get binary identical results with those from
> the continuous simulations.
>
> I also try to get new .tpr files by:
> tpbconv_s -s md.tpr -f md.trr -e md.edr -c md_c.tpr -cont
> and then:
> mpiexec mdrun_s_mpi -v -s md_c.tpr -cpt 0 -cpi md.cpt -deffnm md_c -reprod
> But I still cannot get binary identical results.
>
> I also test the simulations with only one processor and binary identical
> results are still not obtained. Using double precision does not solve the
> problems.
>
> I think that the above problems are caused by some information may not be
> stored during the running of the simulations.
>

That seems likely. The leading candidate would be a random number generator
you're using for a stochastic integrator. Your .mdp file would have been
useful.

On the other hand, if I run two independent simulations using the exactly
> same number of processors, the same commands and the same input files, i.e.
> mpiexec mdrun_s_mpi -v -s md.tpr -deffnm md -reprod
> I can always get binary identical results from these two independent
> simulations.
>
> I understand that MD is chaotic and if we run simulation for enough long
> time, simulation results should converge. Also, there are factors which may
> affect the reproducibility as described in the GROMACS website. But for my
> purpose, I am curious about whether there are certain methods through which
> I can get binary identical results from restarted simulations and
> continuous simulations. Thanks a lot.
>

There are ways to be fully reproducible, but probably not every combination
of algorithms has that property. 4.0.7 is so old no problem will be fixed,
unless it can also be shown in 4.6 ;-)

Mark