On-the-fly clustering for exascale molecular dynamics simulations.

After completing a series of iterations to propagate the minimum local ID within each connected component, we utilize the existing ghost layer communication algorithm, similar to that used in standard Molecular Dynamics simulations. This approach updates the values in the ghost areas based on computations from neighboring MPI processes.

Once updated information is received from neighboring processes, we may encounter additional labels or lower minimum IDs that require further propagation. We alternate between local steps, where the minimum local ID is propagated within the MPI process, and steps that update the ghost areas. This process may reveal lower label IDs in the surrounding area that need to be propagated again.

It is important to note that more iterations may be necessary, especially if a connected component spans multiple subdomains. As propagation occurs, each MPI process updates its ghost areas, and these updates are communicated to neighboring processes. This iterative process continues until a global stable state is achieved, which is identified when no changes are observed in the state across all MPI processes during the local minimum propagation step. At this point, further iterations are unnecessary.