Arbitrary composite particles with arbitrary container shape?

[Dan C.]

I just had a quick question.  I noticed that HOOMD is reaching into the realm of larger than molecular scales with composite particles made from many rigid particles...fantastic stuff.  Can hoomd be used with more complex or arbitrary containers?  I would be satisfied with just a tubular domain, but something that is represented with a wall of stationary particles or trangulated faces would be amazing.  Also, I know about LAMMPS and LIGGHTS, Im looking for a faster solution...unless the GPU accelerated versions of LAMMPS and LIGGHTS are scalable.  Thanks for your help.

Dan

[Joshua A. Anderson]

On Jul 3, 2012, at 5:18 PM, Dan C. wrote:

> I just had a quick question. I noticed that HOOMD is reaching into the realm of larger than molecular scales with composite particles made from many rigid particles...fantastic stuff. Can hoomd be used with more complex or arbitrary containers? I would be satisfied with just a tubular domain, but something that is represented with a wall of stationary particles or trangulated faces would be amazing.

In current versions of hoomd, you can make walls out of particles easily, simply leave them out of the group you integrate and they will remain fixed in place. There is some overhead associated with this because those particles must still be accounted for in the neighbor list.

You can also specify planar surfaces that interact with particles via wall.lj. You can represent any convex geometry in this way.

The catch is that wall.lj currently isn't accelerated on the GPU. Sometime this summer/fall, I plan to upgrade it to work on the GPU, offer a wider variety of wall potentials, and to work with a more natural specification of triangular faces. The initial version will likely remain capable of convex geometries only, but a future update may handle concave ones. This method will be faster than using fixed particles, since the number of triangles will be much smaller.

Cylinder/sphere/ellipsoid/etc… boundaries could be added as special cases if there is enough demand for them.

> Also, I know about LAMMPS and LIGGHTS, Im looking for a faster solution...unless the GPU accelerated versions of LAMMPS and LIGGHTS are scalable. Thanks for your help.

Performance is highly dependent on simulation parameters, so I urge everyone to run their own benchmarks for comparison. I'm certain that you will find that HOOMD is much faster than LAMMPS on a single GPU and that scaling up to many GPUs is only possible with huge simulations (>200,000 particles per GPU).

If you find that LAMMPS is faster than hoomd on a single GPU, send me your benchmark scripts and I'll find out why and fix the problem :)

--------
Joshua A. Anderson, Ph.D.
Chemical Engineering Department, University of Michigan