Repository navigation
Replies: 2 comments 4 replies
|
HOOMD-blue stores explicit ghost particles just outside the boundaries that separate ranks in domain decomposition. When a rank shares a periodic boundary with itself, HOOMD-blue stores only the primary copy of the particle and any You cannot solve this with a It is true that wrapping hoomd-rs stores all ghost particles explicitly, both to make implementing custom interactions less complex and to avoid the expensive call to wrap in the innermost loop (hoomd-rs is a CPU-only code). |
|
I forgot to mention: To detect which boundaries are periodic (and need ghosts generated), check Lines 156 to 162 in 44ea3f1 You will also need to generate ghosts at the edges and vertices when more than one direction is periodic. |
Uh oh!
There was an error while loading. Please reload this page.
I am porting KIM-API to HOOMD-Blue (very early alpha stage: https://github.com/ipcamit/hoomd-kim). This should enable users to use any of the OpenKIM models on HOOMD-Blue.
As of now it all works fine for non-periodic systems. Problem is that KIM-API expects neighbors to have trivial position vectors, and models being able to compute relative vector as r_i - r_j , without any additional step(as I understand it only has MIC convention based nl).
Is there a way to do that in HOOMD-Blue natively? Currently I am computing the padding atoms manually before each step. Or perhaps I can export my own neighbor list code, extending the hoomd::md::nlist::NeighborList? Also how does it do domain decomposition on distributed memory systems? I always assumed LAMMPS like ghost atom approach was more efficient.
All reactions