Repository navigation
extract_region hangs at np>=2: the annulus submesh never returns (np=1 is fine) #699
Description
Found while trying to build a zero-cell-rank reproducer for #698. A region submesh is the realistic way a rank owns no cells, so this blocks testing that.
Reproduce
parent = uw.meshing.AnnulusInternalBoundary(
radiusInner=0.4, radiusInternal=0.7, radiusOuter=1.0, cellSize=0.1)
sub = parent.extract_region("Inner")
print("cells:", sub.dm.getHeightStratum(0))| np | result |
|---|---|
| 1 | parent 702 cells, submesh 278 cells, returns |
| 2 | hangs — no output at all, so it never gets past construction/extract |
| 4, 8 | hangs the same way |
Under the #678 supervisor the ranks go quiet with rank 0 in <module>(), i.e. still inside the top-level construction, so this is before any of the submesh's own methods are called.
Why it matters beyond the hang
extract_region is only covered in parallel by test_0770. If that passes while this does not, the difference is worth knowing — it may be the internal-boundary annulus specifically, or cellSize=0.1 (702 cells) rather than the coarser fixtures the tests use.
It also means the zero-cell-rank path through cell_size() (#698) cannot currently be exercised the realistic way. The fix there is argued from the code — var.data's first access calls dm.createSubDM and createGlobalVector, both collective, from behind an early return that only starved ranks take — but not demonstrated.
Same family as #691 (isSimplex rank-local on empty ranks), #679/#680 (empty-rank swarm save), #698.
Underworld development team with AI support from Claude Code
Activity
lmoresi commentedon Sep 7, 2026
Withdrawn — this was my measurement error, not a defect.
extract_regionworks at every rank count I tried:It returns cleanly in all of them.
What actually happened: the annulus at
cellSize=0.1takes a while to generate, and I had two runs sharing one time budget. The np=1 run consumed most of it and my cap killed np=2 before gmsh finished. I read 'no output' as 'hang' when it meant 'not started yet'.I should have checked whether the process was making progress before filing. The supervisor prints exactly that — it reports whether ranks have moved between samples — and I did not run it on this until afterwards.
No change needed. #698's zero-cell path still needs a reproducer; the annulus 'Inner' region is a full ring, so every angular partition owns some of it and this route does not starve a rank anyway.