[gmx-users] question about continuation using tpbconv and mdrun -cpi

[Mark Abraham]

Baofu Qiao wrote:
> Hi Berk,
> 
> Thanks for pointing out my mistake!
> What I worry about is how big the deviation is after several continuations? If 
> it is unpredictable, it is a sad news.

The objective is to avoid perturbing the system *with the restart*. Any 
perturbation will make the trajectory sufficiently far in the future 
unrecognizable from the unperturbed one, i.e. chaotic. However such a 
perturbation that is small enough that it does not disturb the ensemble 
will lead to results that are statistically identical to the unperturbed 
system. For example, you could add up the forces on the atoms in the 
opposite order, which perturbs the system through rounding error, but 
(ought) not shift the ensemble. Dynamic load-balancing achieves a 
similar outcome.

Non-reproducibility in the integrator can be acceptable if the above 
condition is true. Hence all my previous advice - by all means check 
that you can achieve reproducibility over restarts, but ignore it for 
production simulation, unless you actually need it.

Mark