Distributed Molecular Modeling over
Very-Low-Bandwidth Computer Networks
14 Washington Ave.
Arlington MA 02174
email: [email protected]
This is an abstract
for a poster to be presented at the
Fifth
Foresight Conference on Molecular Nanotechnology.
A draft of the full article is available at the author's Web site.
The full paper is also available here.
The Internet offers the possibility of distributing a
molecular modeling task over a number of geographically diverse
computers. If these computations can be arranged not to intrude
on the extant workloads of home and office desktop computers,
vast resources may become available. This scenario offers
extremely limited communication bandwidth, so the problem of
partitioning molecular modeling (normally a fairly
communication-intensive task) in a reasonably efficient manner
becomes interesting. This paper presents a partitioning approach
that will allow individual processors to simulate thousands of
time steps before requiring communication with other processors.
Communication bandwidth is replaced by duplicated
computations, exploiting the spatially localizable influence of
each atom upon its neighbors. Processors are assigned to model
overlapping cubical regions of space. Duplicate modeling results
for overlap regions gradually diverge, due to differences in the
boundary information available to neighboring processors. By
characterizing these divergences and keeping track of them,
processors can update one another locally, each processor acting
as the expert on a small cubical region of space in the center of
its own initially assigned area.
Each processor is ignorant of the world beyond its own
initially assigned region. The unknown effects of that outer
world propogate inward from the known border over time. The
velocity of propogation is characterizable. A processor may
continue computing without communicating with its neighbors until
noticeable external effects reach its area of expertise. With
periodic updates from its neighbors, a processor should be able
to simulate for long periods of time.
The trade-off of communication bandwidth with duplicated
computations is acceptable in the envisioned scenario. This
approach might also be useful in more traditional computing
environments of networked workstations, in cases where
communication bandwidth may be significantly expensive compared
to computation.
*Corresponding Address:
Will Ware,14 Washington Avenue, Arlington MA 02174,
email: [email protected],
http://world.std.com/~wware/
|