[berkgeveci]

A reader that gives some ranks nothing leaves those ranks without pieces. Every rank must make the same communicating calls, so no rank may decide from its own pieces alone what the others decide from theirs. Algorithms that did either refused on such a rank or hung the job.

Fixed

• Refused or failed on a rank without pieces, leaving the others waiting: probe of pieces or an AMR hierarchy, streamlines and advect over pieces, voxelize, and Advection (with the path filters). Such a rank now takes part through a piece of nothing (_common._stand_in): its points and seeds go to the pieces that hold them, and the answers come back.
• Chose their fields, or a shortcut, from this rank's pieces: update_ghosts, extract_surface of pieces, AMR ghost_cells and average_down, integrate over pieces, and the VTK Statistics, Entropy and Integrate algorithms. A rank without any now gets the others' choice. The VTK Integrate also takes one origin for the centroid over every rank.
• merge_pieces onto a root rank without pieces refused. The merge is now made on the first rank that has pieces and moved to the root.

Test

test_empty_ranks.py runs _empty_ranks_mpi.py under mpiexec -n 3. It has 36 cases: pieces, AMR hierarchies, octrees, advection and the VTK algorithms over pieces. They run with four pieces on rank 0 only, on ranks 1 and 2, on ranks 0 and 1, and on none. Each answer is held to the same pieces in one process.

• Hangs are stopped by a deadline that names where each rank was.
• No pieces anywhere may refuse, with the same error on every rank.
• CI: the MPI job runs the test as a step of its own. Its limit goes from 15 to 30 minutes; the test takes about 1-1.5 minutes a layout here.

Not fixed: an error only the ranks with data can detect (a scalar field given as a velocity) still leaves the others waiting (#364, 0.8).

Tests: test_empty_ranks.py (4 layouts), the touched modules' tests with their MPI ones (672 passed), the quick tier (3,124), lint.

🤖 Generated with Claude Code