MPI overdecomposition

[joaander]

Original report by Jens Glaser (Bitbucket: jens_glaser, GitHub: jglaser).

---

It would be a nice performance feature to have to be able to over-decompose a system into small domains. This can be important for very compute-intensive simulations (e.g. HPMC depletants), which are currently limited by the domain fine-graining size. The feature would require having more than one (n) halo-layers around each domain. I haven't investigated if LAMMPS can already do that.

Required changes

• Communicator::migrateParticles can send particles out to more than one neighbor shell (some care need to be taken with periodic boundary conditions)

• Communicator::exchangeGhosts/updateGhosts can pull in ghosts from more than one neighbor shell

• CellList update the number of ghost layers (trivial)

[joaander]

Interesting idea for MD. But I don't think it gains you much for HPMC. In HPMC the inactive region is 1 particle diameter wide in each domain. As the domain size approaches 1 particle diameter, very few trial moves will be performed. And none at all when the domain is exactly 1 wide.

[joaander]

Original comment by Jens Glaser (Bitbucket: jens_glaser, GitHub: jglaser).

---

You're right, with HPMC the active layer gets thinner and thinner, so the number of valid trial moves will go down the way you described, and yes, this holds for hard particles only. This suddenly makes the proposal less attractive, as MD is typically less compute intense.

[joaander]

Probably doesn't make sense to implement.