Computational Molecular Dynamics Software Tools (CMDT)
- O. Yasar, Oak Ridge National Laboratory (ORNL)
- Y. Deng and R. A. McCoy, SUNY-Stony Brook
- R. B. Marr and R. F. Peierls, Brookhaven National Lab (BNL)
Development of generic particle dynamics MD software
to aid materials research for a variety of problems,
including thin-film deposition, solar cells, and
Algorithms and features:
- Short-range, atomistic simulation capability on Intel
Paragons and PVM workstations.
- High degree of memory and CPU efficiency
- General dynamic load balancing algorithm
- classical newtonian mechanics.
- linked-cell method.
- spatial decompostion.
- amenable to parallelization and suitable for large systems of particles.
- the force exerted on a particle comes only from particles spatially
residing within some cutoff distance.
- short-ranged interactions (pair-wise 6-12 Lennard-Jones potential),
but easy to include other forms of interaction including three bodies.
Efficiency and Scalibility:
- preservation of geometric locality to avoid increased communication costs.
- domain can be seen as partitioned into two
grids: a coarse grid corresponding to the processor decomposition, and
a fine grid corresponding to the subdomain decomposition.
- the load on a processor is just the
summation of the loads in its resident subdomains.
- the movement of a coarse grid vertex is determined by
comparison of the loads of the four processors which have
the vertex in common.
- the coarse grid vertices are allowed to move toward more
heavily loaded regions.
- The MD algorithm is highly scalable
with the only limiting factor being the amount of available
memory per node.
- runs up to 200 K particles per 32 MB RAM on OSF/1
without significant paging. 400 M particles on
1024 nodes with 64 MB memory/node.
CCS home page
Osman Yasar email: email@example.com