That's correct, as a newly minted MPI programmer I'm using MPI_COMM_WORLD
although I would like to start making my own communicators at some point.
I suppose I could define a communicator for every process during the
direct space decomposition, which would then include processes that owned
one or more of the six nearest neighbors of any one subdomain of the new
communicator's root process.
I'll look into this; for now I'm starting to lay out the foundation for a
robust communication schedule which doesn't rely on MPI_Barrier to keep
communications distinct as I do now. The point-to-point communication
plans are actually pretty well optimized, as far as I can tell--I am
essentially using a simplified version of the GROMACS decomposition, one
of the Neutral Territory methods described by DE Shaw's group (although
this particular decomposition was invented in the late 1980s by someone
else). By the release I want to have a really smooth, valgrind-clean
parallel implementation that scales to 8 CPU. After the release we can
look into much larger parallelism, a proper load balancer, and various
reciprocal space allocations.
Dave
> I'm not sure how you're parallelizing things in mdgx, but I'm going to
> take
> a guess here and say that you're probably doing all communications over
> MPI_COMM_WORLD, yes?
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sat Dec 24 2011 - 08:00:02 PST
