What Dave (Cerruti) describes is pretty much the approach taken in
EspressoMD, a code often used for polymers and therefore one where long
molecules are a requirement.
It is a nice structure because it then facilitates things like using
Lees-Edwards boundary conditions to impose shear flow, these days it seems
like bio and soft-matter codes might be converging in the system sizes that
they treat, with the soft matter codes going for more accuracy and the bio
codes going for bigger and longer.
If it ends up being faster then yay, great. Maybe someone could build a
special purpose CPU with one instruction for imaging pbcs....
The detailed approach in Espresso is to build a neighbour list which can
contain either "atoms" (distances which didn't need to be imaged) or "ghost
atoms" (distances which were imaged *or* which came in from a different MPI
process). The ghosts and the non-ghosts are all guaranteed to be in-cutoff
for the given atom but the re-imaging of reaction forces needs to be
handled differently for the two: comms may be involved in the second case.
For a large system there are much more atoms being ghosted due to comms
than due to simple PBC so it makes sense to lump them together and focus on
optimising the imaging for effective comms.
Josh
On 14 November 2017 at 02:36, David A Case <david.case.rutgers.edu> wrote:
> On Wed, Nov 08, 2017, Dave Cerutti wrote:
> >
> > I've got a new setup for computing bonded interactions in pmemd. It
> > improves the throughput somewhat (though I must admit the calculation is
> > surprisingly arithmetic bound). It also affords an interesting
> opportunity
> > to push the envelope of what Amber can do, and I think it would patch a
> > vulnerability we've got right now.
> >
> > WHAT I COULD DO:
> > Permit systems with molecules whose bonded structure wraps in a periodic
> > simulation box. You could do a continuous graphene sheet, a MOF crystal,
> > continuous cellulose fibers. This would all be accomplished by locally
> > re-imaging the groups of atoms taken in by each bond work unit.
> >
> > I THINK IT'S A BUG FIX:
> > The current setup re-images all coordinates, molecule by molecule,
> relative
> > to a single reference frame, then applies the NMR restraints (and any
> > bonded interactions) to those coordinates.
>
> As others have pointed out, this can be a really complex subject, and not
> something to be approached in an ad hoc or seat-of-the-pants approach.
> Here's
> my brief take:
>
> 1. Simplest (present case, I think): atom ID's for interactions are fixed
> in
> the prmtop file, and don't change. The current coordinates are used,
> without
> checking anything: it's up to the user to ensure that this is what is
> really
> wanted. Pretty secure for covalent interactions that don't break, but
> doesn't
> allow bonds "across" periodic boundaries.
>
> 2. Second case, which I think is well-defined, but not implemented: for any
> pair of atoms, the minimum image interaction can be unambiguously
> determined.
> (Tricky but doable in non-orthogonal unit cells.) Then an interaction
> between two atoms could be marked as "use minimum image here". That would
> allow the bonded structure to go across a unit cell boundary. One has to
> take care for angles and dihedrals: first: use minimum image to get the
> atom-atom vectors involved; then construct the angles/dihedrals out of such
> atom-atom vectors. You can probably go back to minimum image for the 1-4
> nonbonded terms.
>
> [I *think* this would only work for multi-atom terms that can be
> constructed
> from atom-atom vectors. You can't use this idea for things like
> center-of-mass restraints, for example.]
>
> But one would have to make sure that is what is really wanted. I think
> it would work for wrapping bonded structures in a periodic simulation
> box. It probably won't work for artificial constraints, like NMR distance
> constraints that may be broken at times, etc.
>
> Note that "minimum image" is not the same as "imaging" (in the sense that
> such a term is used in cpptraj. I think the second ("BUG FIX") idea above,
> of reimaging all coordinates and then applying NMR restraints, is unlikely
> to work.
>
> ...something to think about anyway....dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Nov 13 2017 - 23:30:02 PST
