Molecular Dynamics Simulation on Commodity Shared-Memory Multiprocessor Systems
with Lightweight Multithreading
John Thornley, Maria Hui, Hao Li, Tahir Cagin, and William Goddard III
High Performance Computing Symposium (HPC '99),
1999 Advanced Simulation Technologies Conference (ASTC 1999)
Abstract
The simulation of N-body particle systems has many important applications in
science and engineering, ranging in scale from molecular dynamics to
astrophysics. Because of the time required to simulate realistic numbers of
particles, a great deal of research has been devoted to parallel N-body
simulation. Much of this research has focused on the development of
sophisticated algorithms for data distribution and workload balancing on
message-passing systems. In this paper, we investigate the implications on
parallel N-body simulation of the recent advent of powerful, low-cost,
commodity shared-memory multiprocessor systems with support for lightweight
multithreading. As the basis of our investigation, we have ported a large
parallel molecular dynamics simulation program (MPSIM), developed at the
Materials and Process Simulation Center at Caltech, to a lightweight
multithreaded implementation for Intel Pentium family multiprocessors
running Windows NT. We compare the performance of our version of MPSIM
running on commodity multiprocessors with that of previous versions running
on expensive scientific supercomputers. In particular, we test the
hypothesis that data distribution and workload balancing are significantly
simpler on commodity shared-memory multiprocessors than on traditional
supercomputers. Our experiments indicate that (i) the performance of
commodity multiprocessors is competitive with that of expensive high-end
scientific platforms, and (ii) parallel N-body simulation algorithms can be
made considerably simpler and more flexible by taking advantage of shared
memory and lightweight multithreading. With the trend in commodity
multiprocessor systems toward lower prices, faster chips, and larger numbers
of processors, simulations that now require supercomputers will soon be
possible on inexpensive multiprocessor PCs. It is important that we learn
how to de-velop the most effective simulation algorithms for these exciting
new platforms.