[gmx-users] quick question about restart

[Mark Abraham]

On 29/02/2012 9:47 PM, Anna Marabotti wrote:
> Dear gmx-users,
> I'm running my simulations on a machine with a queue system that stops 
> the runs longer than 24 hours. Often, for my simulations, I need a 
> longer time to complete the runs. Let's have an example. I have to run 
> a 100ns-long simulation on a machine that produces 10ns/days 
> simulation. I have two possibilities:
> First approach:
> I'm setting within the .mdp file a global duration of 100 ns. Then:
> grompp -f param.mdp -c prot.gro -p topol.top -o input_fullMD.tpr
> mdrun -s input_fullMD.tpr -deffnm output_fullMD
> After 24 h, I see that the run has been stopped after 10 ns. I have 
> the output_fullMD.cpt file. Then, I do:
> mdrun -s input_fullMD.tpr -cpi output_fullMD.cpt -deffnm output_fullMD_2
> and so on for 10 days, until the run has been completed (then I will 
> concatenate the .xtc, .edr, .trr files)

Yep, or use the same -deffnm filename and -append for more simplicity. 
See http://www.gromacs.org/Documentation/How-tos/Doing_Restarts#section_2

> Second approach:
> I'm setting within the .mdp file a global duration of 10 ns. Then (as 
> above):
> grompp -f param.mdp -c prot.gro -p topol.top -o input_fullMD.tpr
> mdrun -s input_fullMD.tpr -deffnm output_fullMD
> After 24 h, the run has been finished. Then, I continue the run for 
> other 10 ns with:
> tpbconv -s input_fullMD.tpr -extend 10000 -o input_fullMD_2.tpr
> mdrun -s input_fullMD_2.tpr -cpi output_fullMD.cpt -deffnm output_fullMD_2
> and so on for 10 days, until the 100-ns run has been completed.
> Are the two procedure producing EXACTLY the same results, or not?

In principle, they are implementing the same algorithm over the 10 runs. 
In practice, the discussion at 
http://www.gromacs.org/Documentation/Terminology/Reproducibility applies 
and you will likely not observe an identical trajectory from each case. 
In theory, mdrun -reprod does a better job in many cases for often 
losing performance, but random number states are not preserved, so 
various algorithms cannot be reproducible across restarts.

> Which is the best preferred procedure?

The first, with appending, is simple and requires no concatenation 
post-processing. Feel free to back up your files between runs if you 
want insurance against loss - but you should do that anyway.

Anything else requires you to do more work later on for no gain.

Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20120229/f9052e32/attachment.html>