[gmx-users] comparison of reaction field and PME

[Erik Lindahl]

David van der Spoel wrote:

> On Mon, 4 Mar 2002, Yuguang Mu wrote:
>
> >I have a little experience of AMBER package, there using PME
> >a little slower that CUT-OFF.
>
> But in both cases, gromacs is considerably faster. The difference in speed
> is due to the more complex innerloop. Also, you can reduce the cut-off for
> Coulomb when using PME (e.g. to 0.9) while you keep the LJ at 1.2. This
> will give you the same Coulomb forces as when using 1.2, but much faster.
> The LJ forces between 0.9 and 1.2 are only updated when you do
> neighboursearching (recommend nstlist = 5).
>

I think this is the most important part - when you use PME, we introduce a
lot of new computations that are performed in the reciprocal space. The whole
idea with this is that you are able to decrease you real space cut-off
significantly!

The default settings with a spacing of 0.12 and electrostatic cut-off at 0.9
should give you HIGHER accuracy than a reaction-field.

PME will always be slower than reaction-field it the cut-off is the same,
even on a single node: First, we add a significant amount of computations in
reciprocal space, and second we have to use tabulated interactions instead of
analytical in real space (This is 30-50% slower, depending on your cpu).

Cheers,

Erik