Choose ib_neighborhood_radius from the thinnest rank, not the widest

[sbryngelson]

Written with Claude Code. Addresses the automatic-choice half of #1851.

The bug

ib_neighborhood_radius counts rank hops. A rank keeps an immersed-boundary patch only while the patch centroid lies inside its own subdomain grown outward by that many hops (s_get_neighbor_bounds → f_neighborhood_ranks_own_location), and s_handoff_ib_ownership deletes it otherwise. So the radius must span the body in every direction, and the hop that needs the most is the one crossing the thinnest rank.

The automatic choice assembled its rank width the wrong way round:

local_rank_width = -1._wp
do each direction:
    local_rank_width = max(local_rank_width, <this rank's extent in that direction>)
call s_mpi_allreduce_min(local_rank_width, min_rank_width)

Each rank reports its widest extent; the minimum is then taken across ranks. Where every rank is long in one direction and thin in another, the width reported is the long one and the radius comes out too small — the reported "picks 1 and faults".

The fix

Reduce over directions with min rather than max, so the result is the narrowest rank extent over every rank and every direction. Two lines.

How the validation is obtained

examples/3D_ibm_neighborhood_radius/ — a thin plate in a long narrow channel, 20 chords by 8 by 6 with 400×50×50 cells, sized so MFC's topology search settles on 16×2×2 at 64 ranks. That is an ordinary shape for a wake, a jet or a channel, and it makes the ranks strongly anisotropic:

direction	ranks	extent per rank
x	16	1.250
y	2	4.000
z	2	3.000

The plate is 1.0 × 2.4 × 0.1, so s_get_ib_bound returns 1.3010. Crossing that at 1.250 per hop needs ceil(1.1 × 1.3010 / 1.250) = 2; the old width of 4.000 gives 1.

./mfc.sh run examples/3D_ibm_neighborhood_radius/case.py -n 64

	Automatic choice of ib_neighborhood_radius selected:
before	1
after	2

Both runs complete (1 M cells, a few minutes on two CPU nodes). SUMMARY=1 python3 case.py prints the half-extent, the rank extents and the arithmetic without running anything. The README carries the same table and the reasoning.

Two real production grids agree, measured from their own lustre_*_cb.dat:

case	topology	old width → radius	new width → radius
gust encounter, 128 ranks	16×2×4	2.051 → 1	1.052 → 2
flapping wing, 128 ranks	8×4×4	1.745 → 1	1.027 → 2

Scope

The new width is never larger than the old, so the chosen radius never decreases — the change can only make the neighbourhood more conservative, at the cost of more hops in the force reduction. Cases that set ib_neighborhood_radius explicitly are untouched, and so is any near-cubic decomposition, where the widest and narrowest extents coincide. That is why a uniform grid with a balanced topology shows no difference, and why no golden file moves.

This does not fix #1863 (a spurious transverse force on multi-rank IB runs); I checked, and raising the radius there changes nothing.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Pull request overview

Fixes the automatic selection of ib_neighborhood_radius for anisotropic MPI decompositions by basing the hop distance on the thinnest rank extent, and adds a reproducer example documenting/validating the behavior.

Changes:

• Change rank-width reduction logic from “widest direction per-rank” to “narrowest direction per-rank” when computing the auto ib_neighborhood_radius.
• Add a new 3D example case that reproduces the issue and prints a quick SUMMARY of the expected arithmetic.
• Document the reasoning, before/after result, and scope in an accompanying README.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
src/simulation/m_start_up.fpp	Fixes the rank-width computation used to auto-select ib_neighborhood_radius on anisotropic decompositions.
examples/3D_ibm_neighborhood_radius/case.py	Adds a minimal reproducer case (and optional SUMMARY output) to demonstrate expected radius selection.
examples/3D_ibm_neighborhood_radius/README.md	Documents the issue, reasoning, and validation results for the new reproducer case.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.