[gmx-users] Constraints and efficency

[Mark Abraham]

On 5/01/2011 5:53 AM, Elton Carvalho wrote:
> Hello, fellow GROMACS users!
>
> I noticed many users applying constraints to their simulations, be it
> SETTLE to water molecules or LINCS to some other species. This brings
> about an efficiency concern once that after each Verlet step some
> systems of coupled equations must be solved in order to apply these
> constraints.
>
> My question is: water molecules are the most numerous species in a
> solvated simulation, and the SETTLE algorithm seems to take O(n) time
> for each molecule (Hence, O(3) per molecule), thus yelding O(M) to
> apply SETTLE to all M water molecules in a system, per MD step. Why is
> this beneficial? Does the larger timestep allowed by eliminating
> high-frequency O-H vibrations outweigh the time needed by SETTLE?

Most definitely. Typical GROMACS simulations report up to a few percent 
of the total simulation time is spent in dealing with constraints. If 
the use of constraints permitted a 2fs timestep rather than 0.5fs, then 
it's nearly a four-fold speed-up.

> Any other reason I failed to notice?

I seem to recall that bond constraints are held to model the essential 
QM nature of interatomic bonds better than a Morse or harmonic potential 
(particularly for bonds to hydrogen). I don't recall a reference for 
that offhand, unfortunately. Perhaps the original SHAKE paper.

> Same thing for LINCS in, say, surfactant molecules: Is there any other
> reason to do this other than shave off higher frequencies (C-H for
> instance) and allow a larger timestep?
>
> The only reason I see to constrain movement other than that fraquency
> shaving (and, of course, isolating some behaviour of interest, but
> this is case-specific) would be to reduce the number of dimensions of
> the phase space and save computatuon time, but, as implemented, this
> doesn't seem possible with the code as is.

Various coarse-graining approaches exist to take advantage of that kind 
of saving, but they also gain from the higher maximum time step.

Mark

> My molecular dynamics background is based on Cerius2, so please pardon
> some lack of intimacy with GROMACS behaviour.
>
> Regards,