[xiaoyeli]

Summary

• get_perm_c_dist() calls a serial ordering library (MMD, COLAMD, or METIS) independently on every process, with no broadcast or cross-check of the result afterward. Correctness silently depends on that library returning a bit-identical permutation given identical input on every call.
• Real METIS is deterministic, so this has worked historically. It breaks when SuperLU_DIST is linked against ParMETIS built on top of Scotch's METIS-compatibility layer (libscotchmetisv5): with threads enabled, Scotch documents that its orderings are not guaranteed deterministic across calls. When two processes get different perm_c, later collectives in pdgstrf/pddistribute go mismatched and the factorization deadlocks with no error.
• Fix: at every get_perm_c_dist() call site in the 2D/3D/ABglobal drivers (pdgssvx, pdgssvx3d, pdgssvx_ABglobal, and their single- and complex-precision counterparts, including the _d2 variants), compute the ordering only on process 0 and MPI_Bcast it to the rest of the grid — mirroring the existing MPI_Bcast(perm_r, m, MPI_INT, 0, grid->comm) pattern already used for MC64 row permutations in the same files. This makes the result independent of the ordering library's determinism, and saves every other process from redundantly repeating the same serial computation.

Fixes #229.

Test plan

• Built the full static library (libsuperlu_dist.a) with -DXSDK_INDEX_SIZE=64 across all three precisions (single/double/complex16) — zero errors, zero warnings on any of the 11 touched files.
• Would appreciate a maintainer re-run of the reporter's Scotch-based reproducer from Deadlock when METIS_NodeND returns different orderings on different processes (METIS_AT_PLUS_A) #229 to confirm the deadlock no longer occurs, since reproducing that specific environment (ParMETIS built against Scotch with threads enabled) wasn't available here.

---

Generated by Claude Code