mpp & nonbonded cutoffs

[rec %! at !% snll-arpagw.llnl.GOV (Ray Cline)]

 Bill Ross mentions:
 > But you could further imagine marching in waves of forces from nested
 > few hundred angstrom shells from previous steps: a freight train
 > leapfrog algorithm (think of the memory & communication bandwidth
 required).
 When using a short-range cutoff it is possible to develop a systolic
 type algorithm that pipelines timesteps as well as parallelizing the
 force/potential calculations. I originally thought about this a couple
 of years ago to increase sample sizes by doing out-of-core stuff. The
 interesting thing about this technique is that you can replace the
 disk with another computer and therefore would be useful for
 distributed computing and/or multiple mpp machines. You can even do
 this with periodic boundary conditions! (Though personally I believe
 that one of the major advantages to larger scale simulations would be
 to remove the need for periodic boundaries) Again, I would be
 interested in discussing these ideas with all others interested
 (perhaps point-topoint e-mail would be better, though if the general
 net is interested I'll continue to respond).
 Raymond E. Cline, Jr.
 Organization 8300-A
 Combustion Research Facility
 Massively Parallel Computer Research Laboratory
 Sandia National Laboratories
 Box 969
 Livermore, CA 94551
 phone: (510) 294-1395
 email: rec %! at !% sandia.llnl.gov
 FAX  : (510) 294-2276