DistributedMovingTessellation — the library-level distributed dynamic-Voronoi driver. More...
#include <mpi.h>#include <array>#include <cmath>#include <vector>#include "../repair.hpp"#include "voronoi_halo.hpp"
Go to the source code of this file.
Classes | |
| struct | peclet::voro::mpi::DistributedMovingTessellation< Real, MAXP, MAXT > |
| struct | peclet::voro::mpi::DistributedMovingTessellation< Real, MAXP, MAXT >::StepStats |
Namespaces | |
| namespace | peclet |
| namespace | peclet::voro |
| namespace | peclet::voro::mpi |
Detailed Description
DistributedMovingTessellation — the library-level distributed dynamic-Voronoi driver.
Promotes the validated orchestration that previously lived only in tests/kokkos_mpi/bench_repair_mpi.cpp into a reusable header: VoronoiHalo (host ghost gather/refresh over core's decomposition + NBX halo) composed with the device MovingTessellation (two-pass repair) under the distributed Verlet-skin invariant.
Per step, with this rank's owned positions:
- trip test: if ANY rank's max owned displacement since the last (re)gather exceeds skin/2, ALL ranks re-gather (halo.gather rebuilds the ghost topology via NBX) and cold- rebuild the tessellation. The trip is a LOCAL test but the gather and refresh paths are both collective, so it is reduced to a GLOBAL decision (MPI_Allreduce MAX) — without this the run deadlocks once ranks diverge on the trip (observed at np>=4).
- fast path: refresh ghost positions on the established topology (halo.refreshPositions) and run the local device two-pass repair (MovingTessellation::step).
Owned cells are [0, nOwned) of the combined view; their geometry equals a cold rebuild of the same combined positions to the repair tolerance (gated by the repair_mpi ctests at np=1,2,4).
