[gmx-developers] Re: Various integrator questions

[David]

On Wed, 2005-08-10 at 18:02 +0200, Berk Hess wrote:
> Michel Cuendet wrote:
> 
> >
> >
> >> But a more pratical points is that the fast virial calculation 
> >> algorithm is one of
> >> the reasons why Gromacs is so fast. 
> >
> >
> > The molecualar virial is also fast, because all (large) intramolecular 
> > forces cancel out and don't have to be taken ino account. A lot of 
> > noise is avoided like this.
> >
> I thought the molecular might take much more time as much more work might
> be required in the innerloops.
> But now I think of it, it might only require extra bookkeeping and a 
> small performance hit.
I don't think this is the case as now the forces are computed between
one atom and all other atoms in the same energy group pair (pair in the
energy matrix). No track is kept of the molecule pair. One would have to
keep track of the forces on each molecule which may be simple for
molecules that do not have intramolecular non-bonded interactions, but
difficult otherwise. One has to keep track of the aggregate force
between two molecules and multiply that by the vector connecting the
centers of mass. It would basically imply much shorter inner loops, that
are also more complicated. Hence the efficiency would be much lower.

Most important, we can *not* use any of the optimized inner loops that
we have, we'd have to start from scratch.

Michel said:
>There is also a nice review in which they give a Trotter decomposition 
>of an atomic scaling scheme (without constraints).
>
>P. Procacci, M. Marchi
>/Multiple Time Steps Algorithms for the Atomistic Simulations of
>complex 
>molecular systems/
>Proceedings of the NATO-ASI School, Erice 11-21 June 1998
>/ Advances in the Computer Simulations of Liquid Crystals /
>C. Zannoni and P. Pasini Editors, 333-387, Kluwer Academic publishers, 
>Dodrecht the Netherlands (1999)


If we use a proper multiple time step algorithm which still allows 2-4
fs time-steps for the non-bonded interactions we can of course make much
better scaling algorithms. It is trivial to spread the bonded
interactions over parallel computers, if we don't need constraint
algorithms parallellization will be much easier.


-- 
David.
________________________________________________________________________
David van der Spoel, PhD, Assoc. Prof., Molecular Biophysics group,
Dept. of Cell and Molecular Biology, Uppsala University.
Husargatan 3, Box 596,          75124 Uppsala, Sweden
phone:  46 18 471 4205          fax: 46 18 511 755
spoel at xray.bmc.uu.se    spoel at gromacs.org   http://xray.bmc.uu.se/~spoel
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++