Latent MPI deadlock: spring smoother returns early on rank-local n_free==0 before collectives

[lmoresi]

Found (marker-only, Charter §9) during the Wave D smoothing readability pass, 2026-07-08. A TODO(BUG) marker lands with PR #345 in src/underworld3/meshing/smoothing/spring.py.

The spring-equilibrium mover takes a rank-local early return when a rank has no free (unpinned) nodes, BEFORE reaching the collective reduction calls later in the routine. On a partition where some rank owns only pinned/boundary nodes (small partitions, heavily graded meshes at high rank counts), that rank exits while the others enter the allreduce — a classic collective mismatch deadlock.

Not observed in the campaign's np2 gates (both ranks own free nodes in the test meshes); the hazard is latent and partition-dependent.

Suggested fix: replace the early return with a zero-work path that still participates in every collective (empty contribution), or hoist the collectives above the return. Needs an np>2 test with a deliberately starved rank.

Related: the Wave D smoothing PR #345 (marker placed, behaviour untouched per readability-wave scope discipline).

Underworld development team with AI support from Claude Code

[lmoresi]

Maintainer context (2026-07-08): the spring smoother is one of the retired mover methods — only mmpde really survived testing. This deadlock is therefore in a low-priority legacy path: do not invest in hardening it. The right resolution is likely deletion of the retired movers (spring / Monge-Ampère / OT-step / anisotropic-Winslow) in a future prune, keeping mmpde (and the Taubin surface smoother, which is separate). If a prune is ruled, close this issue with that PR instead of fixing the collective ordering.

Underworld development team with AI support from Claude Code

[lmoresi]

Additional maintainer context: mmpde replaced the Monge-Ampère/OT family — it provides the equivalent equidistribution functionality under an isotropic metric (and handles anisotropic metrics besides). The retirement is supersession, not just test attrition, so a prune of the legacy movers loses no capability.

Underworld development team with AI support from Claude Code

[lmoresi]

Resolved by #371 (4bcee78): the spring mover — and with it this rank-local early-return deadlock — is deleted; mmpde (which superseded the spring/MA/OT family) and the Taubin surface smoother survive.

Underworld development team with AI support from Claude Code