Repository navigation
mesh.cell_size() is partition-dependent: the per-cell radius comes from a kd-tree over the rank's centroids, so cells at partition boundaries get a different size on each rank count #687
Description
Mesh.cell_size() returns a P0 field filled from self._radii, which _get_mesh_sizes computes as the RMS distance from each cell's vertices to the NEAREST centroid found by a kd-tree over the centroids this rank holds. Near a partition boundary the nearest centroid is a different cell than in serial (the neighbour set is truncated), so the field, and everything scaled by it (the SUPG tau of the Eulerian solvers, the Nitsche penalty gamma mu / h), depends on the partition.
Measured on the Kovasznay-flow test of the SUPG Navier-Stokes solver (feature/navier-stokes-supg, res 8, three steps from the exact solution, solver tolerance 1e-11 so the linear solve is not the cause):
| tau built on | np 1 error | np 2 error | difference |
|---|---|---|---|
a constant h (get_min_radius) |
0.0023408316563576 | 0.0023408316563546 | 1e-15 |
mesh.cell_size() |
0.0024855792696189 | 0.0024843274043763 | 5e-4 |
The plain Stokes solve and the tau = 0 (Galerkin) step on the same mesh are partition-independent to 1e-12, so the cell-size field is the only rank-dependent input. The scalar SUPG parallel test (test_1077) did not catch it because at its Courant number tau is dominated by the transient term and h barely enters.
The fix is a purely local definition of the size, each cell's vertices against its own centroid, which is identical on any partition and equal to the present value on a regular mesh; on an irregular mesh it is the same quantity without the neighbour's centroid occasionally standing in. Implemented alongside the Navier-Stokes SUPG solver, where it was found.
Underworld development team with AI support from Claude Code
Activity
fix: make stabilization cell sizes local to each cell (underworldcode…
lmoresi commentedon Sep 5, 2026
Duplicate of #569 (15 August), which is still open and carries the consequence this one does not:
add_nitsche_bc(local_h=True)is the default, so Nitsche answers moved 6.6e-03 between rank counts.Same defect, and #692 fixes it — measured across rank counts on that branch: the new own-cell field is bit-identical at np=1/2/4 (max abs diff 0.000e+00), while the legacy kd-tree field moves by 3.3e-03 (np2) and 4.1e-03 (np4).
Suggest closing this one in favour of #569, or closing both from #692.
lmoresi commentedon Sep 6, 2026
See the consolidated proposal on #694 — this, #569, the
get_*_radius()contract and the naming are all one tidy-up rather than four.Short version: three definitions of "cell radius" are in play (kd-tree neighbour spacing, #692's own-centroid RMS, and PETSc's
volume**(1/dim)), the docstrings describe the PETSc one while the code uses the kd-tree one, and the PETSc routine — abandoned years ago as broken — now returns one finite value per cell and is bit-identical across rank counts.fix: make stabilization cell sizes local to each cell (underworldcode…
Make mesh.cell_size() local and partition independent (#692)
Merge origin/development into feature/navier-stokes-supg: take the la…
Parallel tests: a refinement hierarchy for the Navier-Stokes referenc…
Navier-Stokes with Eulerian SUPG momentum transport, and a partition-…
Merge development (#688 Navier-Stokes SUPG, #687 cell size) into feat…
lmoresi commentedon Sep 9, 2026
Fixed by #692 and merged to development via #688 (1d5e8c1): mesh.cell_size() is now cell-local and partition independent.
Guarded by tests/parallel/test_1077 (the field, cell by cell against its own serial answer) and test_1078 (the three radius accessors).