espressomd / espresso

The ESPResSo package
https://espressomd.org
GNU General Public License v3.0
225 stars 183 forks source link

Eliminate the need for a ghost force reduction #4556

Open RudolfWeeber opened 2 years ago

RudolfWeeber commented 2 years ago

Currently, Es calculates forces/energies in the RegularDecomposition as follows: \

The drawback of this scheme is the need for an additional communication step, i.e., the ghost force reduction. This is assumed to be one of the main obstacles to Espresso's scaleability, as all nodes have to wait for each other in the communicaiot step of the reduction. The aim of this project is to gradually remove the need for this ghost force reducion.

[ ] Phase 1: Short-range non-bonded forces

At the end of this step, an LJ fluid should run without ghost force reduciotn. Other featuers such as bonds and vitual sites will break.

  1. [ ] Add a flag to the constructor of the RegularDecomposition without_ghost_force_reduction
  2. [ ] Support it in the Python interface. See the additional parameters of the HybridDecomposition for inspiration.
  3. [ ] If the flag is true, the local cells neighboring ghost cells ON ALL SIDES have to be added to the ghost cells red_neighbors. This is done in RegularDecomposition::init_cell_interactions()
  4. [ ] If the flag is true, omit the ghost force reduction by returning an emtyp GhostCommunicator from RegularDecomopsition::collect_ghost_forces_comm()
  5. [ ] Expose the flag via an interface function in the ParticleDecomposition class hierarchy (via a a virtual function returning false in the ParticleDecomposition base class). For the RegularDecomposition override it, returning the actual value of the flag.
  6. [ ] In CellStructure::linked_cell(), query the without_ghost_force_reudciont flag of the decomposition stored in CellStructure::m_decomposition via the interface function added in the previous step. If true, run the algorithm::linked_cell() on the m_ghost_cells in addition to the current call on m_local_cells. You'll have to replicate the construciotn of fisrt/last the way it's done for m_local_cells.
RudolfWeeber commented 2 years ago

@pkreissl can you please read through this and check, whether it aligns with your understanding of the cell system?

pkreissl commented 2 years ago

I concur with your description the cell system, this is also my understanding of how it currently works in ESPResSo.