Repository navigation
Issues Snapshot 2026 10 08
Tomislav Janjusic edited this page Oct 8, 2026
·
4 revisions
Snapshot date: 2026-10-08
Total open issues: 261 (261 analyzed)
Analysis by: Claude Opus 5.5 (claude-opus-5-5, effort: medium). Verdicts are recommendations; issues are closed only after human review.
Closed issues since the last snapshot (2026-10-07): 31, plus 1 new. Jump to the list.
| Field | Meaning |
|---|---|
| Year (section header) | Year the issue was opened |
| Type | bug, feature, question or documentation (our call, else from GitHub labels; - = unknown) |
| Category | Triage group: Held for review, Fixed (waiting on backport or decision), Close candidate (fixed, answered, another project, usage, unsupported, ROMIO), Decision needed, Tracking, Keep (with a reason), Needs check. Rows are grouped in that order |
| Note | Our triage note; blank = not investigated yet |
| Year | Open | Analyzed |
|---|---|---|
| 2026 | 58 | 58 |
| 2025 | 30 | 30 |
| 2024 | 33 | 33 |
| 2023 | 43 | 43 |
| 2022 | 33 | 33 |
| 2021 | 20 | 20 |
| 2020 | 14 | 14 |
| 2019 | 6 | 6 |
| 2018 | 3 | 3 |
| 2017 | 3 | 3 |
| 2016 | 7 | 7 |
| 2015 | 2 | 2 |
| 2014 | 9 | 9 |
By category: Fixed, waiting on backport or decision 3; Close candidate: another project 5; Close candidate: answered 3; Close candidate: ROMIO 1; Close candidate: unsupported 5; Close candidate: usage 1; Tracking 7; Keep: fix PR open 6; Keep: live defect or request 10; Keep: MPI-5 conformance 7; Keep: partly fixed 4; Needs check 6.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #14337 | Performance regression with cuda drivers 580+ | angainor | bug | Fixed, waiting on backport or decision | (rough pass; reported on v5.0.10): VMM/mpool probes skipped in check_addr by #13773/#14382 on main and v6.0.x; v5.0.11 (#14227) covers host buffers only |
| #13988 | MPI-5.0 Chapter 1 - Introduction to MPI: conformance gaps | jsquyres | feature | Fixed, waiting on backport or decision | (rough pass; reported on main): sole gap #14001 closed; VERSION reports MPI 5.0 on main, v6.0.x and v6.0.0rc2; maintainer-filed so confirm with jsquyres before closing. |
| #13662 | Stack-use-after-return in ompi_info | devreal | bug | Fixed, waiting on backport or decision | (rough pass; reported on main): fixed on main by #13663 (09a2d300a0); v6.0.x/v5.0.x still have the &value leak, so a backport is worth asking before closing |
| #14385 | UCX Memory mismatch in ompi_coll_base_gather_intra_binomial | DreamerLuu | question | Close candidate: another project | (rough pass; reported on unspecified): binomial gather intentionally uses host temp buffers; the out-of-tree IPCL UCX transport rejecting host memory is a transport issue; workaround basic_linear gather |
| #14305 | CUDA-aware MPI: inter-node MPI_Irecv into a CUDA IPC-imported device buffer fails | hppritcha | question | Close candidate: another project | (rough pass; reported on CUDA-aware OMPI+UCX): question answered (legal buffer, zero-copy impossible by driver); remaining failure is UCX, tracked in openucx/ucx#12013, with a verified workaround |
| #13886 | pmix2/prrte3 help files not found | jngrad | bug | Close candidate: another project | (rough pass; reported on v5.0.10): Ubuntu 26.04 packaging puts PRRTE help files in libprrte-dev; filed as Launchpad bug 2155666. |
| #13803 | btl/ofi: double free if fi_domain fails | devreal | bug | Close candidate: another project | (rough pass; reported on main): CXI provider returned a freed cached fabric; fixed in libfabric by ofiwg/libfabric#12891, OMPI workaround #14332 was dropped |
| #13688 | UCX_IB_MLX5_DEVX mystery | ashterenli | question | Close candidate: another project | (rough pass; reported on main + UCX): UCX_IB_MLX5_DEVX perf/hang is UCX transport behavior; redirect to openucx/ucx, possibly related to openucx/ucx#11558 |
| #14223 | What is the expected timeline for MPI 5.0 / OpenMPI 6 ? | amckinstry | question | Close candidate: answered | (rough pass; reported on v5.0.10): ABI ships in v6.0.0 (#13280 merged, in v6.0.0rc1); timeline and compatibility questions answered by jsquyres. |
| #13836 | btl/smcuda: Message completion not properly waited if there was a cuda context switch after MPI_Init ? | AntonDaumen | question | Close candidate: answered | (rough pass; reported on v5.0.10): maintainer explained switching CUDA context after accelerator init is unsupported; redesign not planned. |
| #13778 | Fabric domain sharing between MTL and BTL component | sadlochr | question | Close candidate: answered | (rough pass; reported on v5.0.10): sharing is done via fi_getinfo hints and works only when the provider returns the hinted domain; maintainer explained it is provider-dependent, not an Open MPI bug |
| #14342 | Unable to build ROMIO in Open MPI v5.0.10 with GCC 15.2.0 | vterzi | bug | Close candidate: ROMIO | (rough pass; reported on v5.0.10): ROMIO build failure with newer Level Zero headers; ROMIO removed in v6.0.0, workaround --disable-io-romio (default io is OMPIO). |
| #14378 | Problems installing OpenMPI (from ubuntu packages) in Ubuntu 24.04 | JOSECCZ | question | Close candidate: unsupported | (rough pass; reported on Ubuntu 24.04 package): distro Open MPI is an unsupported series; launch hang looks like network/ssh configuration, no bug evidence. |
| #14173 | 4.x bug: MPI_Get_accumulate / MPI_Rget_accumulate with MPI_REPLACE hang across nodes on IB |
jeffhammond | bug | Close candidate: unsupported | (rough pass; reported on 4.1.9a1): 4.x-only RMA livelock report; reporter never reproduced on 5.x and the 5.x counterpart #14174 was an rdma-core issue |
| #14153 | A compilation error occurs when Making all in mca/base | Jaden-Anderson | bug | Close candidate: unsupported | (rough pass; reported on v4.1.8): build error is in v4.1.x embedded pmix3x atomics; v5.0.11 bundled OpenPMIx has no such code |
| #13843 | OpenMPI 4.1.8 with --map-by slot:PE=N N > 1 fails when > 1 node | bernstei | bug | Close candidate: unsupported | (rough pass; reported on v4.1.8): v4.1.x/ORTE-only mapping failure; reporter confirms v5.0.10 handles --map-by slot:PE=N across nodes. |
| #13725 | OPAL ERROR: Unreachable in file ext3x_client.c at line 112 | JOHN194872 | question | Close candidate: unsupported | (rough pass; reported on 4.0.3 distro build): error comes from the v4.x ext3x PMIx glue of an Ubuntu 4.0.3 install; v5.0.x does not contain this code |
| #13722 | Setting OMPI_MCA_btl_tcp_if_include=lo as default | amckinstry | question | Close candidate: usage | (rough pass; reported on Debian v5.0.x): config question; package-wide if_include=lo would break multi-node TCP; set it only in the test environment or use btl=^tcp |
| #14000 | MPI-5.0 Chapter 20 - Application Binary Interface (ABI) | jsquyres | feature | Tracking | (rough pass; reported on main): C ABI done via #13280; Fortran ABI deliberately deferred per docs (mpi-forum-abi.rst:425-449). Maintainers decide whether to close with the deferral or keep open. |
| #13997 | MPI-5.0 Chapter 14 - I/O: conformance gaps | jsquyres | feature | Tracking | (rough pass; reported on main): Chapter 14 I/O tracker; all four gap sub-issues (#14012-#14015) still open |
| #13996 | MPI-5.0 Chapter 12 - One-Sided Communications: conformance gaps | jsquyres | feature | Tracking | (rough pass; reported on main): Chapter 12 epic child; its only gap #14011 is still open. |
| #13994 | MPI-5.0 Chapter 9 - MPI Environmental Management: conformance gaps | jsquyres | feature | Tracking | (rough pass; reported on main): Chapter 9 epic child; both gap sub-issues now done or fixed; close after #14008 is closed (maintainer-filed) |
| #13993 | MPI-5.0 Chapter 7 - Groups, Contexts, Communicators, and Caching: conformance gaps | jsquyres | feature | Tracking | (rough pass; reported on main): Chapter 7 MPI-5.0 conformance tracker; its only gap #14007 is still open, keep until that lands |
| #13991 | MPI-5.0 Chapter 5 - Datatypes: conformance gaps | jsquyres | feature | Tracking | (rough pass; reported on main): Chapter 5 tracker; #14004 closed, #14005 still open. |
| #13987 | [EPIC] MPI-5.0 conformance: close the remaining gaps | jsquyres | feature | Tracking | (rough pass; reported on main): MPI-5.0 conformance epic; 15 chapter/gap sub-issues still open (e.g. #13997, #14000, #14012) |
| #14469 | opal_os_dirpath_create() still follows a symlink when the path has a trailing separator | tonyskjellum | bug | Keep: fix PR open | (rough pass; reported on main): fix PR #14470 open; hole still at os_dirpath.c:149 main / :148 v5.0.x |
| #14348 | Install MPI ABI into separate prefix | eschnett | feature | Keep: fix PR open | (rough pass; reported on main): agreed fix is #14426 (--disable-ompi-abi), still open; wrapper naming (mpicc vs mpicc_abi) unresolved. |
| #14291 | ABI: need to add in some plumbing for MPI T event callback object types | hppritcha | feature | Keep: fix PR open | (rough pass; reported on main): MPI_T event ABI handle plumbing pending in open PR #14306. |
| #14175 | osc/ucx: MPI_Accumulate with derived datatypes SIGSEGVs (NULL deref in the reduction) on the short-message path |
jeffhammond | bug | Keep: fix PR open | (rough pass; reported on v5.0.10rc2): fix PR #14207 (osc/ucx lower-bound adjust) is open on main awaiting review; needs merge plus v5.0.x backport. |
| #14086 | Bogus UCX error message | rhc54 | bug | Keep: fix PR open | (rough pass; reported on main): bogus osc/ucx error still emitted on main/v6.0.x/v5.0.x; fix PR #14521 open. |
| #13762 | opal_lifo tests hangs/crashes on specific riscv64 hardware (P550) | andreabolognani | bug | Keep: fix PR open | (rough pass; reported on main 9479fd182f): fix PR #14495 still open awaiting hardware test; earlier PRs #13789/#13969 closed. |
| #14588 | flang: make check dies using flang-23 | hppritcha | bug | Keep: live defect or request | (rough pass; reported on main): maintainer-filed macOS flang-23 make check link failure from bare -framework flags; flang libtool patch (config/ltmain_flang_darwin.diff) does not cover this path; no fix yet |
| #14546 | security policy doc | hppritcha | documentation | Keep: live defect or request | (rough pass; reported on n/a): owned request for a security policy doc; nothing on main yet. |
| #14410 | MPI_T Events: What is considered dereferencing a handle? | Matthew-Whitlock | question | Keep: live defect or request | (rough pass; reported on v6.0.x pre-release): unanswered question exposes ambiguous docs (MPI_T_Events.3.rst:213-217); needs a maintainer answer and doc clarification. |
| #14299 | Stale plugins allowed to load | rhc54 | feature | Keep: live defect or request | (rough pass; reported on main): valid design request by rhc54; TODO still at mca_base_component_repository.c:492; related #8439 open |
| #14298 | A fresh window has a name ("rdma window 3") | eschnett | question | Keep: live defect or request | (rough pass; reported on main): default window names remain; behavior pending MPI Forum ruling in mpi-forum/mpi-issues#1117 (PR #14406 withdrawn). |
| #14160 | Open MPI program stops to work due to ptrace linux kernel vulnerability CVE-2026-46333 | jfikar | bug | Keep: live defect or request | (rough pass; reported on v5.0.5): smsc/cma query (smsc_cma_component.c:115-124) treats a successful PR_SET_PTRACER as enough even at ptrace_scope 2/3; workaround --mca smsc ^cma. BTL-open failure without explicit btl list unexplained. |
| #14142 | ABI: Performance Concern: ABI Datatype Converter Uses Linear Search in Hot Path | hppritcha | feature | Keep: live defect or request | (rough pass; reported on main): maintainer-filed perf/design question on ABI converters; linear chain still in c.py on main, no PR or profiling yet. |
| #14029 | opal/mca/if: posix_ipv4 ignores if_base_retain_loopback (#if 0 since 2010); loopback enumeration differs across Linux/macOS vs the BSDs | jsquyres | bug | Keep: live defect or request | (rough pass; reported on main): loopback filter still #if 0 at if_posix.c:198 (main) and :196 (v5.0.x); no fix PR |
| #13740 | OMPI OFI BTL accumulate buggy/not optimized | aingerson | bug | Keep: live defect or request | (rough pass; reported on v5.0.9): osc/rdma accumulate takes per-target exclusive lock with get/op/put (osc_rdma_accumulate.c:985) giving ms-scale latency over btl/ofi; no response or fix from OFI owners |
| #13648 | Make better configury check for all the python modules required for Sphinx | rhc54 | feature | Keep: live defect or request | (rough pass; reported on main): configure (oac_setup_sphinx.m4) still only probes sphinx_rtd_theme; sphobjinv/jsonschema from docs/requirements.txt unchecked. Valid owned request. |
| #14015 | Collective-read status ucount reports the requested, not transferred, byte count (breaks EOF-by-short-read) | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): live OMPIO defect on main and v5.0.x, collective read status reports requested bytes (common_ompio_file_read_all.c:142, io_ompio_file_read.c:153); no PR yet |
| #14014 | MPI_File_get_type_extent ignores the file's data representation (returns native extent even for external32 views) | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): OMPIO get_type_extent still returns native extent (io_ompio_file_open.c:358 main); no PR. |
| #14013 | MPI_Register_datarep is non-functional (dead version guard; converters never invoked) | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): verified dead version guard io_base_register_datarep.c:52 and OMPIO stub io_ompio_component.c:364 on main and v5.0.x; no fix PR. |
| #14012 | The required "internal" data representation is rejected by MPI_File_set_view | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): 'internal' datarep still rejected at io_ompio_file_set_view.c:76 on main (and v5.0.x:69); no PR |
| #14011 | MPI_Win_get_info does not synthesize default values for honored-but-unset hints (same_size / same_disp_unit / accumulate_ops) | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): MPI_Win_get_info default-hint synthesis still missing; TODO at osc_rdma_component.c:1548 on main. |
| #14007 | MPI_Comm_split_type(MPI_COMM_TYPE_RESOURCE_GUIDED) ignores mpi_pset_name and yields MPI_COMM_NULL | jsquyres | bug | Keep: MPI-5 conformance | (rough pass; reported on main): live MPI-5.0 gap, comm.c:1618-1628 on main never consults mpi_pset_name so pset-only RESOURCE_GUIDED splits return MPI_COMM_NULL; no PR yet |
| #14005 | MPI_Type_get_value_index returns MPI_DATATYPE_NULL for unnamed value-index pairs; MPI_COMBINER_VALUE_INDEX never produced/decoded | jsquyres | feature | Keep: MPI-5 conformance | (rough pass; reported on main): MPI-5.0 conformance gap, COMBINER_VALUE_INDEX absent from ompi/datatype decode on main; no PR. |
| #14336 | ABI: Comments after running libabigail against the MPI ABI stubs | dalcinl | bug | Keep: partly fixed | (rough pass; reported on main): internal-symbol leak fixed by #14367 (#14387 v6.0.x); missing deprecated MPI_Attr_/MPI_Keyval_ in libmpi_abi still unaddressed. |
| #14125 | MPI Sessions: repeated MPI_Session_init after full finalize deadlocks on macOS (shmem backing-file ENOENT + PMIx fence hang) | jsquyres | bug | Keep: partly fixed | (rough pass; reported on main): PMIx fence part fixed in PMIx master, but btl/sm per-session segment-name reuse and unlink race (btl_sm_component.c:388, :303 on main) unaddressed per rhc54 analysis |
| #13915 | [ULFM] coll_ftagree_agreement MCA parameter cannot be set, and MPIX_Comm_agree deadlocks cross-node (v5.0.8) | chzyao | bug | Keep: partly fixed | (rough pass; reported on v5.0.8): ftagree stack-pointer fix in v5.0.11 (#13952) and main (#13663) but missing on v6.0.x (v6.0.0rc2 coll_ftagree_component.c:96); HAN/cuda under ULFM fixed only main/v6.0.x |
| #13708 | accelerator_cuda_check_addr expensive in CUDA-aware MPI | NamjaeChoi | bug | Keep: partly fixed | (rough pass; reported on v5.0.5/v5.0.8): probe-deferral fixes landed (#14227 in v5.0.11, #13773/#14382 on main/v6.0.x) but the opt-out tunables PR #13943 is still open; reporter has not retested. |
| #14413 | Unwanted debug/verbose output with nVidia card | amckinstry | bug | Needs check | (rough pass; reported on v5.0.9 Debian): string not in Open MPI; test whether lstopo prints it or HWLOC_COMPONENTS=-gl,-opencl silences it (likely Mesa via hwloc plugin). |
| #14383 | uct BTL busy-spins forever in singleton MPI_Init() on a system with no RDMA hardware (v5.0.10) | GitPaean | bug | Needs check | (rough pass; reported on v5.0.10): singleton MPI_Init hang blamed on btl/uct, but default uct memory_domains (mlx5_0,mlx4_0) should exclude it; check Ubuntu /etc/openmpi mca-params, retry with --mca btl ^uct and an upstream v5.0.11 tarball, get gdb backtrace |
| #13956 | macOS libmpi_abi absolute install name complicates MPI Forum ABI runtime selection | jsquyres | bug | Needs check | (rough pass; reported on main): no @rpath change landed; docs claim DYLD_LIBRARY_PATH works on macOS. Test: build hello with Open MPI mpicc_abi, run with DYLD_LIBRARY_PATH pointing at MPICH ABI lib, check which runtime loads. |
| #13755 | Parallelization issue with Monte Carlo codes on Apple M3 Ultra | MATHISB15 | bug | Needs check | (rough pass; reported on v5.0.9): reporter never answered bosilca's request for verbose output and lldb stacks (7 months stale); needs a run of hwloc lstopo and --display-map on M3 Ultra or close as stale. |
| #13680 | hdf5 mpi tests fail/hang | ashterenli | bug | Needs check | (rough pass; reported on v5.0.9): t_mpi GB file-size test hang not reproducible by maintainer; test on v5.0.11 with --mca io ompio and --mca fbtl/fs verbose, or ask reporter for a stack trace of the hang |
| #13661 | Leaks reported by LSAN on mpi4py tests | devreal | bug | Needs check | (rough pass; reported on main): leak cleanup is partially done; re-run the mpi4py ASAN workflow on current main with LSAN detect_leaks=1 to see what remains before closing (maintainer-filed) |
By category: Held for review 1; Tracking 5; Keep: assigned to janjust 2; Keep: fix PR open 3; Keep: live defect or request 7; Keep: maintainer-owned 6; Keep: partly fixed 2; Keep: small fix, no PR 2; Needs check 2.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #13048 | Device to Device transfers don't work with OpenMPI + LinkX provider on AMD GPUs | angainor | bug | Held for review | (2026-10-07; reported on v5.0.6): LNX (shm+cxi) GPU buffers failed to register (FI_ENOKEY) because the OFI MTL CXI bypass matched only names starting with cxi; fixed by #12290 (#13090 for v5.0.x) in v5.0.8, reporter confirmed (ofiwg/libfabric#10712). v5.0.9 (#13332) switched to an FI_MR_HMEM check and libfabric 2.2 refactored LNX (name lnx, mr_mode union with shm FI_MR_HMEM), so GPU buffers are now registered explicitly via LNX; that path passed D-D OSU tests on AMD and NVIDIA Slingshot systems in #13156 (closed). Untested: v5.0.9+ with refactored LNX on AMD. |
| #13485 | v6.0.0 checklist | hppritcha | documentation | Tracking | (rough pass; reported on v6.0.x): v6.0.0 release checklist owned by hppritcha/edgargabriel; rc1 and rc2 tagged, final not yet released. Keep open until v6.0.0 ships. |
| #13482 | Connect-accept between multiple mpirun instances | rhc54 | feature | Tracking | (rough pass; reported on v5.0.x/main): rhc54 RFC on whether cross-allocation connect/accept between mpirun instances is needed or only docs; no replies; docs already cover the shared-DVM procedure (launching-apps/unusual.rst). |
| #13300 | Need to update release notes for 6.0.x | hppritcha | documentation | Tracking | (rough pass; reported on main/v6.0.x): v6.0.x changelog is now populated (32 entries on v6.0.x) but v6.0.0 is still in rc; owner (hppritcha) can close it at the v6.0.0 release. |
| #13078 | v4.1.9 release checklist | bwbarrett | feature | Tracking | (rough pass; reported on v4.1.x): v4.1.9 release checklist; milestone says the release is not expected, but v4.1.x still gets backports (last merge #14366, 2026-09-10). Release manager (bwbarrett) should decide whether to close as not planned. |
| #13074 | RFC: Provide equivalence of MPICH_ASYNC_PROGRESS | hominhquan | feature | Tracking | (rough pass; reported on main): RFC for an MPICH_ASYNC_PROGRESS-style OPAL progress thread; implementation PR #13088 is open with changes requested and still active (Oct 2026). Keep open until #13088 lands or is dropped. |
| #13561 | mpi4py: osc/ucx dynamic window heap-buffer-overflow | devreal | bug | Keep: assigned to janjust | (rough pass; reported on main): osc/ucx ompi_osc_find_attached_region_position reads dynamic_wins[-1] before its bounds check (ASAN overflow on Win_attach); unchanged on v5.0.x (:961), v6.0.x and main (:1025); no PR. Assigned janjust. |
| #13064 | UCX nonblocking accumulate leads to program freeze | mts42 | bug | Keep: assigned to janjust | (rough pass; reported on v5.0.2/v5.0.6): osc/ucx nonblocking accumulate fences and progresses UCX inside the completion callback (v5.0.x osc_ucx_comm.c:1910); draft #14263 addresses it, unmerged. Assigned bosilca/janjust. |
| #13525 | Feature: Provide network hardware counter summary at completion time | sadlochr | feature | Keep: fix PR open | (rough pass; reported on main): feature request for CXI NIC hardware-counter reports at finalize; owned by the reporter, whose PR #13830 (ompi/mca/hook/hwpc_cxi) is open with changes requested. |
| #13428 | segfault due to thread-unsafe getenv() call | zsalvet | bug | Keep: fix PR open | (rough pass; reported on v4.1.3/v4.1.8, same code on v5.0.x): getenv() in the PMIx model callback can race with setenv in app threads (ORCA segfaults); fix PR #13433 open without review; maintainers disagree on whether to change it. |
| #13181 | Linking fails with AOCC-5.0.0 compilers | bourdin | bug | Keep: fix PR open | (rough pass; reported on v5.0.7 with AOCC 5.0): configure short-float probe gives a false positive with clang, so libmpi needs __extendhfsf2 at link time; fix PR #13188 is open with changes requested; workaround --disable-alt-short-float. |
| #13383 | Allow use of OFI BTL with RXM | jakemoroni | feature | Keep: live defect or request | (rough pass; reported on unstated version): feature request to allow ofi_rxm in btl/ofi; exclusion still in v5.0.x/main btl_ofi_component.c, no maintainer response or owner yet; needs an OFI maintainer to evaluate FI_DELIVERY_COMPLETE semantics. |
| #13376 | ompio produces adverse, conflicting access patterns for some collective writes | carns | bug | Keep: live defect or request | (rough pass; reported on main): ompio vulcan partitions aligned to absolute offsets, so an unaligned collective write gives an aggregator split extents that fbtl/posix data-sieves into conflicting read/modify/write; code unchanged on v5.0.x and main, no reply or fix. |
| #13361 | mprotect failure on Fedora 41 during memory hook installation when clang asan is used. | etiennemlb | bug | Keep: live defect or request | (rough pass; reported on v5.0.5): memory/patcher binary-patches mmap and breaks ASAN interceptors; same as closed #12819; workaround OMPI_MCA_memory=^patcher (also used by OMPI ASAN CI). No guard on v5.0.x/main. Alternative: close as dup of #12819 with the workaround. |
| #13325 | ULFM cannot handle faults in MPI_Comm_spawn-ed processes | Robyroc | bug | Keep: live defect or request | (rough pass; reported on main 2025-07 with --with-ft ulfm): a failure in an MPI_Comm_spawn-ed job aborts the run instead of raising ULFM errors; no reply and no fix; spawn does not mark the child job recoverable. |
| #13215 | How is PRTE's failure detector supposed to work? | arielscht | documentation | Keep: live defect or request | (rough pass; reported on v5.0.7): question answered and confirmed, but rhc54 said the ULFM docs list PRTE errmgr_detector_* options that do not exist; still in docs/features/ulfm.rst (v5.0.x :208-220, main :206-218). Keep for the docs fix, or close and file a docs issue. |
| #13134 | opal_lifo test fails on FreeBSD amd64 | LaurentChardon | bug | Keep: live defect or request | (rough pass; reported on v5.0.7): opal_lifo fails on FreeBSD amd64 with clang unless CFLAGS has -mcx16; maintainers saw a configure gap and a broken non-128-bit LIFO path, nothing changed since. FreeBSD port adds -mcx16. Labelled question but is a bug. |
| #13068 | Segfault when choosing ofi btl in One-Sided mode | angainor | bug | Keep: live defect or request | (rough pass; reported on v5.0.6): ob1 segfaults when the only BTL to a peer is RDMA-only (btl/ofi one-sided mode); v5.0.x pml_ob1_isend.c:92-93. #13331 changed the ofi default on main/v6.0.x only; the missing unreachable check remains. |
| #13523 | srun with OpenMPI 5.0.3 unexpectedly launches MPI jobs as singletons without ERROR |
GayatriManda | feature | Keep: maintainer-owned | (rough pass; reported on v5.0.3): plain srun without PMIx silently starts N singletons; a notice appears only when PMI_FD/PMI_CONTROL_PORT is set (ompi_rte.c:521-576); valid enhancement assigned to hppritcha (milestone Future), no PR. |
| #13453 | Efficiency improvements for error code/class management | naughtont3 | feature | Keep: maintainer-owned | (rough pass; reported on main): maintainer-owned enhancement (naughtont3) to reuse freed error code/class slots; not implemented on main (errcode.c:386), with an open design concern from bosilca. |
| #13341 | mpirun man page needs to be audited for the 6.0.0 release | hppritcha | documentation | Keep: maintainer-owned | (rough pass; reported on main): v6.0.0 docs task; mpirun.1.rst still has 23 'old, hard-coded content' admonitions on main, v6.0.x and v6.0.0rc2. No PR yet. |
| #13241 | ompi_info needs to be more accurate about fortran f08 support | hppritcha | bug | Keep: maintainer-owned | (rough pass; reported on main): maintainer-owned v6.0.0 item; on main MPI_SUBARRAYS_SUPPORTED is still hardwired .false. (mpif-config.h.in:85) and ompi_info hardcodes f08 subarrays to 0 (param.c:164) despite TS 29113 bindings; #13239 only added configure output. |
| #13153 | ompi and opal datatype infrastructure should support MPI big count API | hppritcha | feature | Keep: maintainer-owned | (rough pass; reported on main): ompi-layer big count support landed in #13460 (main) and #14096 (v6.0.x), unlinked here, but opal descriptor counts/loops are still uint32_t (opal_datatype_internal.h:133,151), so the opal half looks unfinished. Maintainer-filed. |
| #13127 | spawn: the final one to fix in ibm test suite | hppritcha | bug | Keep: maintainer-owned | (rough pass; reported on main): maintainer-filed hang in ibm dynamic/no-disconnect; ompi_dpm_dyn_finalize still gates on 1 < ompi_comm_num_dyncomm (dpm.c:1764 v5.0.x, :2017 main), no fix or PR found. |
| #13427 | Failing to MPI_Comm_spawn proc on different node with running PRTED but no running MPI rank | renatobhf | bug | Keep: partly fixed | (rough pass; reported on v5.0.8): ULFM spawn onto a node whose last proc died fails; runtime resource recovery fixed only in PRRTE master/v4.1 (openpmix/prrte#2400, main/v6.0.x pin), not PRRTE v3.0.14; MPI-side failed-spawn path still open, no owner. |
| #13397 | cm/psm2 fails to get key from PMIx | schiotz | bug | Keep: partly fixed | (rough pass; reported on v5.0.7/v5.0.8): mpirun case fixed by PMIx v5.0.11 restoring pnet/opa (bundled in Open MPI v5.0.11, unlinked); singletons still fail because mtl/psm2 has no fallback key (mtl_psm2.c:116-131). |
| #13431 | with clang 21.1.6 as wrapper compiler for cuda-aware openmpi-5.0.9 and compile options -fopenmp and -fopenmp-targets=nvptx64-nvidia-cuda one gets the message "The call to cuMemHostRegister(0x7fb38c000000, 134217728, 0) failed." for every code with OpenMPI [Patch is proposed as temporary solution in the last comment] | bschulz81 | bug | Keep: small fix, no PR | (rough pass; reported on v5.0.8/v5.0.9): smcuda's cuMemHostRegister of its sm segment fails under clang OpenMP offload (root cause open, llvm/llvm-project#162586); verified OMPI bug: the help call lacks an argument and prints garbage. |
| #13120 | Valgrind suppression writev(vector[...]) does not suppress for recent valgrind | khanc | bug | Keep: small fix, no PR | (rough pass; reported on v5.0.7): tcp_send valgrind suppression uses writev(vector[...]), which valgrind 3.23+ no longer emits; entry unchanged on v5.0.x and main (contrib/openmpi-valgrind.supp). Easy fix; milestone v5.0.10 has passed, retarget. |
| #13340 | Performance regression in atmospheric model when using OpenMPI – traced to slow memset | yuhao0102 | bug | Needs check | (rough pass; reported on v5.0.1): 2-4x slowdown vs Intel MPI attributed to memset; maintainers suggested comparing mapping/binding, no reply. Test: same --map-by core --bind-to core as IMPI, perf call graph for memset callers, on v5.0.11. |
| #13130 | CUDA-aware MPI build on fresh Ubuntu 24.04 LTS, MPIX_Query_cuda_support() returns zero |
niklebedenko | question | Needs check | (rough pass; reported on v5.0.7): cuda accelerator DSO not found under mpirun but found as singleton; unreproduced, Docker build works, request for config.log (2025-04-03) unanswered. Test: rebuild with internal PMIx; check mpirun env for OPAL_/OMPI_MCA_ vars. |
By category: Held for review 1; Close candidate: fixed 1; Tracking 2; Keep 29.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #12931 | I am getting a 'Profiling Timer Alarm' signal after upgrading openmpi versions | David12811 | - | Held for review | (2026-10-07): SIGPROF kills mpirun; nothing in Open MPI/PRRTE/PMIx arms ITIMER_PROF; identical #13496 was a profiling (-pg) prterun installed over the release because --disable-binaries does not gate mpirun, PRRTE or PMIx tools on v5.0.x/main. Plan: open an issue (maybe a PR) for the --disable-binaries limitation, then close this pointing to it. |
| #12547 | dead coll tuned alltoall mca parameters | burlen | - | Close candidate: fixed | (rough pass; reported on v5.0.3): unused coll tuned alltoall params marked deprecated by #12551 (main/v6.0.x only); reopened by wenduwan, likely for a v5.0.x backport that never happened |
| #12625 | Remove script wrappers in v6.0.x | jsquyres | question | Tracking | (rough pass; reported on main/v6.0.x planning): open design question on dropping script wrapper compilers; conda-forge may use them for cross compilation and they are still in v6.0.x and main. |
| #12204 | MPI 4.1 - opaque data handles and state preservation across sessions | hppritcha | - | Tracking | (rough pass; reported on main): MPI 4.1 design question whether committed datatypes must survive session/WPM finalize; maintainers disagree (document behavior vs datatypes retaining the instance); no PR. |
| #13002 | Singletons do not capture ctrl-C and therefore do not cleanup | rhc54 | bug | Keep | (rough pass; reported on main/v5.0.x): singletons do not handle ctrl-C, so session dirs and shm files are left behind; #13261/#14584 fixed only normal-exit cleanup and leave ctrl-C open; no handler on v5.0.x/main. |
| #12971 | Memory leak with CUDA-aware OpenMPI without UCX | kvoronin | question | Keep | (rough pass; reported on v4.1.7a1): GPU memory grows in a cudaMalloc/Bcast/cudaFree loop because smcuda caches remote CUDA IPC mappings (rgpusm leave_pinned=1, unlimited by default, same on v5.0.x); see #12849; workaround btl_smcuda_use_cuda_ipc=0 |
| #12969 | Partitioned communications hang | ggouaillardet | - | Keep | (rough pass; reported on main): MPI_Startall of partitioned sends hangs because part_persist_start reuses the outer loop index (part_persist.h:521/537 main, :520/:536 v5.0.x); fix PR #12972 was closed unmerged; live v5 bug. |
| #12875 | Open MPI fortran configury has at least one bad test code | hppritcha | - | Keep | (rough pass; reported on main): the Fortran STORAGE_SIZE configure test uses storage_size on an assumed-size array (invalid; Cray FTN rejects it); unchanged on v5.0.x/main; maintainer-filed audit request, no PR. |
| #12812 | MPI_Comm_split_type not creating disjoint subgroups for certain cases | mshanthagit | bug | Keep | (rough pass; reported on v5.0.x and v4.1.6): MPI_Comm_split_type(L3CACHE) builds overlapping groups when ranks span L3 domains (intersect-based locality, v5.0.x hwloc_base_util.c:638); strict-match fix discussed, never merged. |
| #12773 | m4 cleanup on main since we will not be supporting external PRRTE | hppritcha | - | Keep | (rough pass; reported on main): maintainer cleanup to drop external-PRRTE configure logic on main; ompi_setup_prrte.m4 and v6.0.x docs still support external PRRTE, so confirm with hppritcha whether the plan stands. |
| #12657 | Bad persistent collective communication performance | mikesoehner | - | Keep | (in-depth 2026-10-07; reported on v5.0.3): persistent allreduce ~2x slower than blocking because persistent collectives only come from coll/libnbc on v5.0.x (coll_libnbc_component.c:379). #13374 added UCC persistent collectives on main/v6.0.x only, and they are opt-in (coll_ucc_enable defaults to 0; default coll_ucc_cts has no persistent ops), so by default libnbc is still used on every branch. |
| #12646 | Reduce collectives latency by reusing temporary buffers | wenduwan | feature | Keep | (rough pass; reported on main): enhancement to pool and reuse temporary collective buffers. #13782 and #13824 partly implement it for coll/han (main/v6.0.x). Owned by AWS (yinliaws). |
| #12637 | lower coll accelerator priority | bosilca | bug | Keep | (rough pass; reported on main/v5.0.x): coll/cuda (v5.0.x) and coll/accelerator (main) still default to priority 78 over coll/ucc at 10, so GPU buffers get staged through host; assigned to bosilca, no fix yet. |
| #12602 | coll tuned alltoall algorithm ignored after initialization | burlen | - | Keep | (rough pass; reported on v5.0.3): settable coll_tuned_*_algorithm cvars are cached per communicator at enable time, so MPI_T changes are ignored on existing comms; unowned enhancement (rules file is the existing workaround). |
| #12585 | Error when using MPI_Comm_spawn with ULFM enabled | goutham-kuncham | - | Keep | (rough pass; reported on main): ULFM + MPI_Comm_spawn fails coll selection because ftagree installs a NULL iagree on intercomms (main coll_ftagree_module.c:46, coll_base_comm_select.c:334); bosilca's patch never merged; also hits v6.0.x. |
| #12584 | when i run mpi program using ASAN, asan reports some memory leaks | helin1995 | - | Keep | (rough pass; reported on v5.0.3): MPI_Init/Finalize leaks under ASAN. #12615/#13868 and #13846/#13867 fixed some in v5.0.11, but the avx op double retain (#13306, main only) is still in v5.0.x. Backport #13306 and rerun LSan. |
| #12548 | Discrepancy between 'oshcc' compiler wrapper and corresponding pkg-config files "oshmem-c.pc" and "oshmem.pc" | scienceplease | - | Keep | (rough pass; reported on v5.0.2): live configure typo, opal_setup_wrappers.m4:938 on v5.0.x (main :986) overwrites OSHMEM_PC_CFLAGS with the Fortran flags, adding -I<moduledir> to oshmem-c.pc/oshmem.pc and leaving oshmem-fort.pc Cflags empty. |
| #12536 | Support zero-copy non-contiguous send | pascal-boeschoten-hapteon | feature | Keep | (rough pass; reported on general): feature request for zero-copy non-contiguous sends; needs OMPI datatype/XPMEM work and a UCX datatype API (yosefe); no owner or PR, could also be closed as not planned. |
| #12517 | openmpi osc_ucx_component error | SLgramps | bug | Keep | (rough pass; reported on v5.0.2): osc/ucx prints a misleading error on every window creation when UCX is unusable (no network). Still in v5.0.x osc_ucx_component.c:375 and main :386. Workaround: --mca osc ^ucx. Assigned to bosilca and janjust. |
| #12479 | Revisit the request management login for the nonblocking resilience support | bosilca | bug | Keep | (rough pass; reported on main): maintainer to-do (owner abouteiller) to rework request handling for ULFM ishrink and other nonblocking FT/comm-creation calls; #12480 only fixed a use-after-free. |
| #12452 | Q: preferred way to implement GPU communication | angainor | question | Keep | (rough pass; reported on unspecified, Apr 2024): unanswered GPU indexed-datatype performance question assigned to janjust; answer it (docs say non-contiguous GPU datatypes are slow due to per-block cuMemcpy), then close. |
| #12437 | Deprecated MCA variable warning mistakenly quotes config file as source of issue instead of environment | lrbison | bug | Keep | (rough pass; reported on v5.0.2): deprecated-MCA-variable warning blames the params file for an env setting; mca_base_var.c:1686-1721 still mixes original/var source; fix PR #13958 open |
| #12419 | Internal error using shmem_reduce in example/oshmem_max_reduction.c | smguzik | bug | Keep | (rough pass; reported on v5.0.2): OpenSHMEM team reductions are unimplemented in spml/ucx (stub at v5.0.x spml_ucx.c:1914, main :2023); assigned to MamziB; workaround: the older *_to_all calls. |
| #12413 | NAG 7.2 does not build 5.0.2, but does 4.1.6: -pthread passed to ld | mathomp4 | - | Keep | (rough pass; reported on v5.0.2): NAG Fortran + gcc build fails (ld: unrecognized option -pthread) when linking libmpi_mpifh.la; fix PR #12458 was closed unmerged in Aug 2026, and the reporter confirmed it was still broken in 2025. |
| #12349 | Singleton init mode, MPI_Comm_spawn, and OPAL_PREFIX | dalcinl | bug | Keep | (rough pass; reported on v5.0.2): singleton comm_spawn finds prte via PATH only, ignoring OPAL_PREFIX; fixed on main by #12390 (v6.0.0rc1) but never backported, v5.0.x dpm.c:1994 unchanged; backport or close as fixed in v6.0.x. |
| #12328 | MPI_Test and MPI_Request get status return wrong state with partitioned communication | cniethammer | bug | Keep | (rough pass; reported on v5.0.2): restarted MPI_Psend_init request completes before MPI_Pready because flags are reset to 0 and inactive sub-requests test as complete (part_persist.h:527); assignee mdosanjh |
| #12323 | Help needed with understanding ompi_display_comm output | tlivolsi | question | Keep | (rough pass; reported on v5.0.1): ompi_display_comm shows generic ucx for peers with no UCX endpoint yet; the init warm-up ring only sends one way, so the report is incomplete under pml/ucx. Display-only defect on v5.0.x/main; assigned janjust. |
| #12286 | Slow MPI_Group_difference | k202077 | feature | Keep | (rough pass; reported on v4.1.3): MPI_Group_difference/union use an O(n*m) overlap scan (group_plist.c:38 on v5.0.x, unchanged on main); valid scalability enhancement, unowned. |
| #12282 | OpenMPI 5.0.1-3 Installation Failure | vorlkets | - | Keep | (rough pass; reported on v5.0.1): make check loads v4.1 components (pmix/ext3x) from the same install prefix and crashes at finalize; owner ggouaillardet, fix PR #12285 still open |
| #12257 | mpifort does not pass -I/usr/include to gfortran | lahwaacz | - | Keep | (rough pass; reported on v5.0.1): with prefix=/usr, mpifort drops -I/usr/include because opal_wrapper.c filters it from compiler_flags; gfortran then cannot find mpif.h. Still in v5.0.x/main; distros patch around it. |
| #12255 | osc/rdma does not select btl/sm for intra-node communications | wenduwan | - | Keep | (rough pass; reported on main/v5.0.x): osc/rdma still prefers btl/ofi over btl/sm for intra-node windows (osc_rdma_component.c:249/1036 on v5.0.x); #12256 only reverted FI_REMOTE_COMM and defers the real fix to this issue. |
| #12205 | MPI 4.1 - error handler argument to MPI_COMM_CREATE_FROM_GROUP | hppritcha | - | Keep | (rough pass; reported on main): MPI 4.1 errhandler-argument errata owned by hppritcha (self-assigned 2025-07-31); validity checks landed via #12960/#12975 and MPI_ERR_ERRHANDLER on main, so the assignee may be able to close it once the #588 fallback part is confirmed |
| #12202 | CUDA + MPI_TYPE_INDEXED + MPI_SEND/RECV super slow | chhu | feature | Keep | (rough pass; reported on v4.1.x and main): GPU pack/unpack of indexed datatypes does one cudaMemcpy per block, still so on v5.0.x and main; perf enhancement assigned to bosilca. |
By category: Tracking 6; Keep 25; Needs check 12.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #12186 | MPI 4.1 - need to check if open mpi T is handling counter overflow correctly | hppritcha | - | Tracking | (rough pass; reported on main, MPI 4.1 item): to-do to check MPI_T overflow semantics (category_changed stamp, dropped-events count) against MPI 4.1; stamp is an unchecked int counter and events just landed on main/v6.0.x. |
| #12088 | MPI 4.1 - MPI_Win_free is synchronizing, except when it isn't | hppritcha | - | Tracking | (rough pass; reported on main, MPI 4.1 item): verify MPI_Win_free synchronization per mpi-forum/mpi-issues#551; osc rdma/ucx/sm free always barrier, so likely no change needed; owner can confirm and close. |
| #11680 | Clean up MCA var de-registration of common components | jsquyres | - | Tracking | (rough pass; reported on main/v5.0.x): code-cleanup item to make common-component MCA var de-registration safe for static string vars; workaround still in common_ucx.c, unowned. |
| #11340 | Documentation beautification on the rst pages | zerothi | documentation | Tracking | (rough pass; reported on v5.0.x docs): docs beautification checklist; most items done via #11367 and follow-ups, three items still open (constants backticks, constants page links blocked on open #11359, man-page link macro). |
| #11290 | v6.x: Remove internal hwloc and libevent | jsquyres | - | Tracking | (rough pass; reported on main/v6 planning): RFC to drop bundled hwloc/libevent; v6.0.x still bundles both and hwloc was just updated to 2.15.0 (#14498), so retarget or decide to close. |
| #11288 | CI for runtime testing (starter) | naughtont3 | - | Tracking | (rough pass; reported on main/CI 2023): runtime-testing CI task; harness merged in open-mpi/ompi-tests-public#25, last step (enable on ompi PRs) never reported; owner can close or update. |
| #12194 | openmpi-4.1.6: ldflags in pkg-config incompatible with CMake's pkg_check_modules(... IMPORTED_TARGET ...) | SomeoneSerge | - | Keep | (in-depth 2026-10-06; reported on v4.1.x/v5.0.x): jsquyres asked to leave it open (2024-01-09). CMake issue 25587 (pkg-config import drops repeated -Wl,-rpath tokens) is still open; Open MPI still emits the split -Wl,-rpath -Wl,<dir> pairs. Milestone/label still say v4.1.x. |
| #12187 | MPI 4.1 - need to check MPI_WIN_SYNC for correct behavior for windows using shared memory | hppritcha | - | Keep | (rough pass; reported on main, MPI 4.1): check MPI_WIN_SYNC is a memory barrier for shared memory; osc/sm and osc/ucx do opal_atomic_mb, but osc/rdma sync (main) does not although #13330 enabled shared_query on it; owner devreal. |
| #12111 | Cannot include mpi-ext.h without mpi.h |
eschnett | - | Keep | (rough pass; reported on v5.0.0): mpi-ext.h is not self-contained (needs mpi.h first); fix #12122 open since 2023; still unfixed in config/ompi_ext.m4 on v5.0.x and main. |
| #12087 | MPI 4.1 - may want to recognize new RMA info key mpi_accumulate_granularity | hppritcha | - | Keep | (rough pass; reported on main, MPI 4.1 item): RMA info hint mpi_accumulate_granularity is still not recognized on main/v6.0.x/v5.0.x; assigned to devreal as part of the MPI 4.1 set. |
| #12078 | MPI 4.1 - recognizing mpi_assert_strict_persistent_collective_ordering would be nice | hppritcha | - | Keep | (rough pass; reported on main): MPI 4.1 optional info assertion mpi_assert_strict_persistent_collective_ordering is not recognized anywhere (grep on main/v6.0.x/v5.0.x); maintainer-filed enhancement, unassigned. |
| #12051 | v5.0.0: Default configure on MacOS with OneAPI compilers fails | jsquyres | bug | Keep | (in-depth 2026-10-06; reported on v5.0.0, icc on macOS): not really an icc issue. With an external libevent found via pkg-config, opal_config_pmix.m4:86-90 passes nothing to the bundled PMIx, which skips pkg-config, so PMIx fails to find libevent; Apple clang users hit the same error (#12168, #13129, closed with --with-libevent=internal). Consider retitling. |
| #12048 | MPI CUDA USAGE | yunusemredemirdag | question | Keep | (in-depth 2026-10-06; reported on v5.0.0): bare --with-cuda finds /usr/local/cuda/include/cuda.h (opal_check_cuda.m4:79-86) but OAC_CHECK_PACKAGE then searches only default paths and silently sets CUDA support to no, with no error although CUDA was requested. Present on v5.0.x and main (regression from #10857). |
| #12045 | coll collectives segfaults on CUDA buffers | wenduwan | - | Keep | (rough pass; reported on v5.0.0): nonblocking reductions on CUDA buffers segfault because ompi_op_reduce is host-only; libnbc nbc.c:522-539 (v5.0.x) still has no device check and coll/cuda or coll/accelerator cover no i* ops. Owned by wenduwan. |
| #12038 | mtl/ofi: call to fi_domain fails on Crusher/Frontier | devreal | documentation | Keep | (rough pass; reported on v5.0.0): CXI fi_domain failure needs PRTE_MCA_ras_base_launch_orted_on_hn=1; documented via #12162/#12163 (v5.0.2), reopened 2026-05 by devreal asking for automatic detection; nothing merged for that. |
| #12019 | MPI_Intercomm_create leaks memory even with symmetric MPI_Comm_free calls | oj-lappi | - | Keep | (rough pass; reported on v4.1.6, repro on v5.0.5): MPI_Intercomm_create leaks per call; v5.0.x comm.c:1692-1719 keeps an extra ref on the remote group (no OBJ_RELEASE after ompi_comm_set); owner bosilca, milestone v5.1.0. |
| #12012 | Performance issue in part/persist | devreal | - | Keep | (rough pass; reported on main 2023): part/persist progress callback is always registered and takes two atomics plus a lock per opal_progress even when unused (v5.0.x part_persist.h:170-185, part_base_select.c:225-227). Assigned bosilca. |
| #12004 | The flow hangs when using MPI_Comm_Spawn for large number of processes | YarShev | - | Keep | (rough pass; reported on v4.1.5/v5.0.0rc/main): the spawn+merge hang went away, but MPI_Comm_disconnect on a dynamic intracomm fences every proc twice and can hang; #12119 closed unmerged; a9f356d4da on main does not appear to dedup. |
| #11845 | MPMD MPI_APPNUM equivalent with srun | b-fg | bug | Keep | (in-depth 2026-10-06; reported on v4.1.5/v5.0.x with Slurm): bosilca reopened it as a bug after jsquyres closed it. v5.0.x/main read PMIX_APPNUM only as PMIX_UINT32 (ompi_rte.c:693-698, else app_num=0); Slurm 25.05 heterogeneous steps reportedly send PMIX_INT, so MPI_APPNUM likely stays 0. Fix: accept any integer type. Untested. |
| #11794 | Han collective do not return errors from internal colls | abouteiller | bug | Keep | (rough pass; reported on main/v5.0.x 2023): HAN collectives drop sub-collective error codes (coll_han_barrier.c:53-61 on v5.0.x/main still returns OMPI_SUCCESS), breaking ULFM error reporting; workaround --mca coll ^han. |
| #11692 | Document new "soft breakpoint' functionality | qkoziol | documentation | Keep | (rough pass; reported on main/v5.0.x docs): docs request for the OMPI_BREAKPOINT / --stop-in-app soft breakpoint; code is in ompi_rte.c but no docs mention it; unowned, maintainers may instead decide it is developer-only. |
| #11691 | Missing BEGIN_C_DECLS (and END_C_DECLS) in various headers | julian-klode | - | Keep | (rough pass; reported on v4.1.4): common_ompio_print_queue.h:65 still has an unmatched END_C_DECLS and autogen's generated frameworks.h lacks C_DECLS on v5.0.x and main; mca_base_var.h already fixed. Easy fix. |
| #11681 | Error propagation in free_fn callback for generalized requests | dalcinl | bug | Keep | (rough pass; reported on v5.0.x): grequest free_fn error handling; first part fixed on main/v6.0.x by #11683, v5.0.x backport #12253 and follow-up #13654 (bosilca) both still open. |
| #11634 | Deprecated info key: map_by | siserte | - | Keep | (rough pass; reported on v5.0.0rc11): dpm_convert in ompi/dpm/dpm.c compares the colon-truncated existing mapby ('ppr') to the full new directive ('ppr:1:node'), so identical map_by keys in MPI_Comm_spawn_multiple fail; unchanged on v5.0.x/main. |
| #11560 | Inconsistent implementation of compare and swap | sunkuamzn | - | Keep | (rough pass; reported on main): plain opal CAS is relaxed in atomic_stdc.h but acquire in gcc_builtin/atomic.h; still inconsistent on v5.0.x/main, needs an owner to decide the intended ordering. |
| #11556 | MPI_2COMPLEX and MPI_2DOUBLE_COMPLEX should be removed | jeffhammond | - | Keep | (rough pass; reported on v4.1.5/main): non-standard MPI_2COMPLEX/MPI_2DOUBLE_COMPLEX still exported in mpi.h (main mpi.h.in:1355) with no usable reductions; request to remove or deprecate is unowned. |
| #11501 | High latency sending from device memory on Summit | devreal | - | Keep | (rough pass; reported on main 2023-03): latency is UCX (openucx/ucx#8761, reloc workaround confirmed), but the ob1+btl/uct segfault on device buffers looks live: btl_uct_am.c:84 has no opal_convertor_on_device check. Test osu_latency -d cuda D D with pml ob1, btl uct. |
| #11492 | MPI_T_pvar_get_num() returns extraneous number of variables and causes MPI_T_pvar_get_info() to return with error later | kingshuk00 | - | Keep | (rough pass; reported on v4.1.4): MPI_T_pvar_get_info returns MPI_T_ERR_INVALID (not INVALID_INDEX) for pvars of unloaded components; error mapping in pvar_get_info.c/mca_base_pvar.c unchanged on v5.0.x and main. |
| #11444 | MPI_Graph_map with no local nodes did not return MPI_UNDEFINED | jeffhammond | - | Keep | (rough pass; reported on v4.1.2): MPI_Graph_map on a non-graph comm returns rank instead of MPI_UNDEFINED; still in v5.0.x graph_map.c:75; fixed on main by #14530, backports to v6.0.x/v5.0.x pending. |
| #11410 | osc/rdma: regression of acc_single_intrinsic for fetch-and-op AMOs | s417-lama | - | Keep | (rough pass; reported on main pre-v5.0.0): osc/rdma fetch_and_op always takes an exclusive lock, so acc_single_intrinsic is ignored; the redundant lock block is still at osc_rdma_accumulate.c:1092 on v5.0.x and :1094 on main. |
| #11359 | Documentation of global variables | zerothi | documentation | Keep | (rough pass; reported on main/v5.0.x docs): valid docs request for a constants/datatypes man page, still missing on main; volunteer stalled, so maintainers should decide whether to do it or point to the MPI standard appendix. |
| #12124 | [5.0.0] MPI communication peer process has unexpectedly disconnected | SimZhou | - | Needs check | (rough pass; reported on v5.0.0): TCP peer disconnect mid-training in k8s; the message points at a dead peer and the reporter never answered whether older versions work; needs peer-side logs or a v5.0.11 rerun. |
| #12123 | OpenMPI hangs during allocation of shared memory if done after allgather | arunjose696 | - | Needs check | (rough pass; reported on v5.0.0 and 4.1.6): hang in Win_allocate_shared after spawn of 128 procs, merge and allgather on Xeon 8468; no root cause, likely tied to open #12115 and #12004. Rerun the mpi4py reproducer on v5.0.11 with 128+ cores. |
| #12115 | Open MPI hangs on MPI_Comm_split_type | YarShev | - | Needs check | (rough pass; reported on v5.0.0): MPI_Comm_split_type hangs on an intracomm merged from a spawn intercomm; main showed btl/tcp 'unexpected process identifier'; rerun the C reproducer on v5.0.11 and main (ob1 with tcp/sm, and ucx). |
| #12053 | No output from mpirun parallel execution | pietro-sillano | - | Needs check | (rough pass; reported on v4.1.5): LAMMPS hangs with -np 4 on a laptop, no diagnosis; needs a plain MPI test (examples/ring_c -np 4) on v5.0.x and a non-GPU LAMMPS run to tell if Open MPI is involved. |
| #11874 | MTT: onesided/c_lock_illegal failures | awlauria | - | Needs check | (rough pass; reported on v5.0.x nightly Aug 2023, ppc64le XL MTT): MPI_Init segfault in mca_pml_base_select with pml ucx; no analysis; rerun the ibm onesided test on current v5.0.x and close if clean. |
| #11740 | Need to increase the timeout value for PMIx_Get when using async_modex | gpaulsen | bug | Needs check | (rough pass; reported on 2023-05 main with PMIx v4.2.3rc1): async-modex MPI_Init failure blamed on PMIx_Get timeout; v5.0.x passes no timeout; main/v6.0.x now default to async modex (#14316). Test v5.0.11 multi-node with pmix_base_async_modex 1. |
| #11719 | mpirun interprets program's command line arguments as its own | bernstei | - | Needs check | (in-depth 2026-10-06; reported on v5.0.0rc-era main): the lone ':' cause was answered (reporter confirmed), but the '--' handling that openpmix/prrte#1746 added was removed by PRRTE eabc861f2f (in v5.0.1 and the current v5.0.x pin), so ':' after '--' still splits apps by code reading. Test: mpirun -n 1 -- echo a : b on v5.0.11. |
| #11618 | osu one-sided failure with vader and rocm | naughtont3 | - | Needs check | (rough pass; reported on v5.0.0rc11/main): osc/rdma PSCW put with ROCm device buffers over btl/sm + smsc/xpmem SIGBUS at 32 KiB; re-run osu_put_bibw -d rocm D D on v5.0.x with xpmem to confirm. |
| #11582 | Possibly incorrect F08 test in ./configure script | cparrott73 | - | Needs check | (rough pass; reported on v3.1.5/v4.1.5, nvfortran 23.3): IGNORE_TKR probe hardened by #12681/#12706 (v5.0.6), but a 2025 user said that patch did not help with NVHPC 24/25. Test v5.0.11 configure with nvfortran 24+/25+ and a use-mpi program. |
| #11547 | --with-pmix=internal should be on by default | jeffhammond | - | Needs check | (rough pass; reported on main 2023 and v5.0.8): configure fails instead of falling back to internal PMIx when a header-less pmix.pc (from an OMPI 4.1.x install) is on PKG_CONFIG_PATH. Reproduce on v5.0.x. |
| #11428 | Deadlock in multi-threaded communication (lock-order inversion in opal?) | jprotze | - | Needs check | (rough pass; reported on v4.1.2/v3.1.3): MT hang in ompi_sync_wait_mt with a small OpenMP reproducer; v5 wait_sync code has the same lock nesting. Run the reproducer on v5.0.x (ob1/sm and ucx, loop, TSan). |
| #11404 | Getting stuck on MPI_FInalize() when using ULFM | rcoacci | - | Needs check | (rough pass; reported on main pre-v5.0.0 and v5.0.2): ULFM survivors hang in MPI_Finalize (or the DVM never exits) on multi-node runs; rerun 02.err_handler.c across 2 nodes on v5.0.11. |
By category: Tracking 2; Keep 22; Needs check 9.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #10459 | List of issues with OMPI master+ULT+PartCom | janciesko | - | Tracking | (rough pass; reported on main / v5.0.0 pre-release): ULT tracking list: 4 of 6 items checked done (configury, mutex backend in v5.0.0); open: partcomm unequal-partition deadlock (unverified) and ULT CI coverage. |
| #10244 | MTT ibm one-sided test failures for ompi v5.0.x | shijin-aws | - | Tracking | (rough pass; reported on v5.0.x branch (2022-04)): MTT tracking list; most items fixed in v5.0.0 (#10462, #10463, #10473, libfabric #7829); 3 residual osc/rdma items over tcp/ofi unverified since 2022-07. |
| #11239 | ompi_mpi_instance_init_common needs to be local | hppritcha | - | Keep | (rough pass; reported on main (Dec 2022, hppritcha design issue)): ompi_mpi_instance_init_common still runs a global PMIx fence and add_procs (v5.0.x instance.c:602,762), so MPI_Session_init is not local. smsc cma/xpmem still use OPAL_MODEX_RECV_IMMEDIATE. |
| #11193 | Wrong received values when using 2 nodes and MPI_indexed datatype | Iasonaspg | - | Keep | (likely datatype-engine bug on v5.0.x/v6.0.x/main; triage note posted 2026-10-05): errlog shows rank 1's data one element (8 bytes) low from ~993 KiB of 4000 KiB onward (all 12,027 wrong values are the last element of a 64-element block). opal_convertor_generic_simple_position() completes a partially unpacked element by count (count_desc--) without advancing base_pointer (opal_datatype_position.c:216-231 v5.0.x, :236-252 main), unlike opal_unpack_partial_predefined(). Trigger: ob1 set_position forward to an out-of-order fragment while an element is split (e.g. striping over two TCP interfaces); pml/ucx unaffected. Not yet reproduced on a real build; fix + datatype unit test proposed. |
| #11184 | osc/ucx: optimizations | janjust | feature | Keep | (rough pass; reported on main/v5.0.x (enhancement by janjust)): Neither optimization is done. Ranges store only base/tail with no target rank, overlaps trigger a WORKER-scope flush, and the array is only reset as a whole (v5.0.x osc_ucx_comm.c:972,999,1028). |
| #11168 | Hang at coll/adapt reduce with ucx | gkatev | - | Keep | (rough pass; reported on v5.0.0rc8 + UCX 1.13.1 (also main)): coll/adapt reduce deadlocks when sends need matching receives (UCX rendezvous) and the outstanding-request limits are small. Unfixed in v5.0.x. main only raised the ireduce segment size (4a72960050). |
| #11166 | Launch scaling issue: calling PMIx_Fence five times during MPI_Init/Finalize | rhc54 | - | Keep | (rough pass; reported on main (Dec 2022)): The duplicate modex/barrier fences are still there. v5.0.x instance.c:602 and :762 fence in init_common, then ompi_mpi_init.c:482 and :548 fence again. The fix #11305 was reverted. |
| #11148 | mpirun can now invoke mpirun | rhc54 | documentation | Keep | (rough pass; reported on v5.0.x (feature note from rhc54)): Doc request: say that nested mpirun (mpirun calling mpirun) works in v5. Neither upstream/v5.0.x nor main docs mention it. |
| #11107 | List the available MPI variables in the documentation | cponder | documentation | Keep | (docs request; verified 2026-10-02): v5.0.0 added docs/tuning-apps/environment-var.rst (6 runtime vars) via #11531/#11615, and wrapper/MCA envars were already in customizing-wrappers.rst and mca.rst. Still missing: one complete env-var list, the env var matching each mpirun option, and runtime vars PMIx still sets (OMPI_WORLD_SIZE, OMPI_APP_CTX_NUM_PROCS, OMPI_FIRST_RANKS, ...). Close as not planned if no consolidated list is wanted. |
| #10962 | OSC header has outdated doxygen comments | wckzhang | documentation | Keep | (rough pass; reported on main (2022-10)): Doc request still valid: query_fn_t doxygen still documents only win/info/comm, not base/size/disp_unit/flavor (main osc.h:~105-126, v5.0.x osc.h:118). |
| #10947 | Reduction latency variation with coll/tuned | devreal | - | Keep | (rough pass; reported on main (2022-10)): Attempted fix (#11023, 'Fixes #10947') was reverted days after merging; v5.0.x barrier decision still flips tree/bruck at 512/1024/4096 (coll_tuned_decision_fixed.c ~464-500). |
| #10759 | Collective CUDA operations have dangerous pointer operation that could corrupt memory | qkoziol | - | Keep | (rough pass; reported on main (2022-09)): Staging copies use raw memcpy of span bytes from sbuf/rbuf, not sbuf+gap, so any datatype with true_lb != 0 copies the wrong bytes (v5.0.x coll_cuda_allreduce.c:55-57,70-82). |
| #10732 | Regression when HOME is not set | haampie | - | Keep | (rough pass; reported on v4.1.4 and main (2022-08)): Still present: mca_base_var_init returns OPAL_ERROR when HOME is unset (v5.0.x opal/mca/base/mca_base_var.c:271-275, main :273-277), breaking env -i builds. |
| #10560 | Make --enable-wrapper-rpath --disable-wrapper-runpath pass -Wl,--disable-new-dtags when supported |
haampie | question | Keep | (rough pass; reported on v4.1.4): No --disable-new-dtags logic exists in config/ on upstream/v5.0.x or main; jsquyres' 6-case proposal was never implemented. |
| #10526 | Unify integer constants between mpi.h and the Fortran interfaces | jsquyres | feature | Keep | (rough pass; reported on v5.0.x / main design): Constants still duplicated: e.g. MPI_ERR_ERRHANDLER hard-coded in upstream/main ompi/include/mpi.h.in:775 and mpif-values.py:305. |
| #10522 | btl/uct not compatible with AM RDMA? | devreal | bug | Keep | (rough pass; reported on main (2022-06), UCX 1.11.2): btl/uct AM handler still passes endpoint=NULL, so AM-RDMA fallback aborts (upstream/v5.0.x btl_uct_component.c:319, main :238). |
| #10402 | ompi_info_t does not follow OMPI object semantics | devreal | - | Keep | (rough pass; reported on master-2022, v5.0.x): the user-facing bug was fixed by #10349 (main) and #10420 (v5.0.x), but devreal kept this open for the design problem: ompi_info_t is an opal object, yet it must be created with ompi_info_allocate() so the instance is retained and released. That split is unchanged: ompi_info_allocate() still calls ompi_mpi_instance_retain() (ompi/info/info.c:405 on v5.0.x, :410 on main). |
| #10259 | Make ompi_mpi_java_eager MCA var show up in ompi_info | jsquyres | - | Keep | (rough pass; reported on all releases with Java bindings): ompi_mpi_java_eager still registered only inside libmpi_java at Java MPI.Init (v5.0.x ompi/mpi/java/c/mpi_MPI.c:278), so ompi_info cannot show it; Java removed on main (#13977). |
| #10152 | opal_path_nfs test fails with (timed-out) autofs mount | gkatev | - | Keep | (reported on 2022 master/v5 era): opal_path_nfs test trusts mount output; an autofs entry for a real NFS path causes a false failure. Also tracks #628 (chroot/mock case, closed as duplicate 2026-09-28). Fix for both: in get_mounts(), skip entries whose mount fs type disagrees with statfs(); skipping autofs alone fixes only this case. |
| #10110 | Comm_spawn/accept using inefficient (and frail) rendezvous mechanism | rhc54 | - | Keep | (rough pass; reported on main (2022)): rhc54 enhancement still unimplemented: dpm still rendezvous via PMIx publish/lookup (opal_pmix_base_exchange) at v5.0.x dpm.c:232 and main dpm.c:424. |
| #10053 | Lock files generated during parallel read of a NetCDF file | dqwu | - | Keep | (rough pass; reported on v4.1.1, v4.1.2, v4.1.3): Partial fix #10006 only fixed close-time cleanup; v5.0.x still creates <file>-<jobid>-<pid>.lock beside the data file at every open and leaks it on failure. |
| #10042 | Asynchronous Cuda Copy in pml ob1 implementation fails to check for copy completion for dtoh copy. | wckzhang | bug | Keep | (rough pass; reported on master (Feb 2022)): ob1 still sets an async dtoh copy stream in the ACK callback but never records or queries dtoh events; on v5.0.x the dtoh event array is only allocated and freed. |
| #9912 | Errors ignored in ft com_ft_detector code | bwbarrett | - | Keep | (rough pass; reported on master (Jan 2022)): fd_progress still ignores the return codes of the PML irecv and of request cancel/wait. The warning was silenced, not fixed (comm_ft_detector.c:583-608 on v5.0.x). |
| #9845 | CUDA-aware Ireduce and Iallreduce operations for GPU tensors segfault | jmerizia | question | Keep | (rough pass; reported on v4.0.7, v4.1.0-v4.1.2 (CUDA-aware build)): Ireduce/Iallreduce on device buffers still fall to libnbc, which runs host ompi_op_reduce on GPU memory; coll/cuda (v5.0.x) and coll/accelerator (main) do not cover them. |
| #11237 | MTT failure for spawn_multiple | rhc54 | - | Needs check | (rough pass; reported on main-202212220241-4d32d2d (IBM MTT)): The MTT spawn_multiple failure (children got the wrong app's argv) was never diagnosed. hppritcha (2025-02) linked it to the multi-namespace locality work in #13059, which closed unmerged. |
| #11087 | Hang in libucs on mpi_finalize | strocode | question | Needs check | (rough pass; reported on v4.1.2rc2 (MOFED 5.5, HCOLL); 2025 repro on HPC-X OMPI 4.1.x + UCX 1.20): Intermittent finalize hang in opal_common_ucx_mca_pmix_fence. Root cause unknown (bosilca 2025-11). Only v4.1-based repros. HCOLL off avoided the original hang. Needs a v5.0.10+ repro. |
| #11047 | Cuda Managed Memory fails with cuMemcpyAsync | wckzhang | - | Needs check | (rough pass; reported on main (2022-11)): Unresolved root cause: v5.0.x accelerator_cuda.c:641 still does cuMemcpyAsync for all device copies incl. managed memory; needs osu_bw M M run with pml ob1. |
| #10860 | MCA_MTL_BASE_FLAG_ACCELERATOR_INIT_DISABLE does not work with non contiguous buffers | wckzhang | - | Needs check | (rough pass; reported on main (2022-09)): v5.0.x pml_cm_request.h:61-67 only sets CONVERTOR_SKIP_ACCELERATOR_INIT for contiguous layouts, which seems to avoid the described failure; needs a test to confirm. |
| #10559 | osc/ucx: MPI_Win_flush sometimes hangs on intra-node | s417-lama | - | Needs check | (rough pass; reported on v5.0.0rc7, UCX 1.11.0/1.12.1): Intermittent MPI_Win_flush hang with two windows on v5.0.0rc7; osc/ucx worker handling was reworked before v5.0.0 but nobody retested. |
| #10410 | v4.1: UCX onesided crash (ibm/onesided/1sided) |
jjhursey | - | Needs check | (rough pass; reported on v4.1.x branch (0987319a4a), UCX 1.11.2, ppc64le): v4.1.x MTT crash in osc/ucx fence/accumulate with old UCX 1.11.2, no follow-up; v5 osc/ucx was heavily rewritten. Needs a rerun on current v4.1.x/v5.0.x. |
| #10374 | --prefix is not properly propagated to 3rd party libraries | bosilca | - | Needs check | (rough pass; reported on main (fa08eac79bc, 2022-05), macOS M1): Likely generic autotools behavior (libevent dylibs not relinked after --prefix change without make clean); needs a rebuild test on macOS to confirm. |
| #10297 | mca_io_ompio_file_write_at_all() failed during parallel write of PnetCDF lib | dqwu | - | Needs check | (rough pass; reported on v4.1.3): OMPIO write_at_all failure at 512 ranks never root-caused; io=^ompio workaround; rerun the scorpio pioperf_rearr 512-rank test on v5.0.x with OMPIO. |
| #10048 | Open MPI memory leaks on Ubuntu | dqwu | - | Needs check | (rough pass; reported on v4.1.2 (Ubuntu 18.04/20.04, --enable-embedded-mode)): Only 2 of the reported leaks (common/ucx variable registration) were fixed. Whether Init/Finalize is valgrind-clean on v5.0.x is unknown. |
By category: Decision needed 2; Keep 13; Needs check 5.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #9690 | cuda-aware build fails cuda runtime linking with undefined symbol cuIpcOpenMemHandle_v2 | mlohry | question | Decision needed | Undecided (4.1.1; same report on 4.1.8 in 2025): undefined cuIpcOpenMemHandle_v2 because the loaded libcuda.so.1 (older driver) lacks it, a driver/toolkit or LD_LIBRARY_PATH mismatch in the v4.1 common/cuda dlsym path. v5.0.x accelerator/cuda links libcuda at build time. Closing comment drafted; related #13326. |
| #9080 | Mixing UCX pml and osc performance degradation | devreal | - | Decision needed | Undecided (main-2021, 4.1.1, UCX 1.10): RMA+P2P slowdown at 64+ nodes. devreal reproduced it with MPICH+UCX and profiled the stalls inside UCX (uct_ep_put_zcopy on dc_mlx5), which points to UCX. No UCX issue filed; no activity since 2021-08. Held open pending a decision. |
| #9795 | mpi_f08 configure tests failing for NAG | jsquyres | bug | Keep | (4.1.2, 5.0.0rc2, 5.0.1): NAG cannot build mpi_f08 because the configure ignore-TKR test has invalid Fortran (type() without dimension(), real vs complex, missing interface); still on v5.0.x and main. Fix PR #9812 closed unmerged. |
| #9764 | SHMEM Fortran library only builds with --enable-oshmem | bwbarrett | bug | Keep | (master-2021): OSHMEM Fortran bindings build only with an explicit --enable-oshmem, because the AM_CONDITIONAL tests enable_oshmem = yes instead of != no. |
| #9722 | Add simple opal atomic store/load? | gkatev | - | Keep | (master-2021): request for relaxed opal_atomic_load/store helpers. None exist on v5.0.x, main or v6.0.x, but coll/xhc (#11418) and coll/acoll now carry their own wrappers waiting for them, so the 'no in-tree use case' objection no longer holds. Status comment posted. |
| #9307 | Lazy intialization of frameworks is not thread safe | nysal | - | Keep | (master-2021): lazy framework initialization is not thread-safe; the topo lazy init is still unsynchronized on v5.0.x and main. |
| #9260 | Using MPI_T for monitoring | jychoi-hpc | bug | Keep | (4.1.1, confirmed on master-2021): monitoring pvars (pml_monitoring_flush, messages_count) are not found through MPI_T; no fix since. |
| #9051 | Memory Leak using Cuda-Aware MPI_Send and MPI_Recv for large packets of data | geohussain | question | Keep | (4.0.5; reproduced again on 5.0.5 with UCX master in 2024): GPU memory grows with CUDA-aware Send/Recv of large buffers. Root cause is the UCX cuda_ipc cache keeping freed peer buffers mapped; workaround UCX_CUDA_IPC_CACHE=n. UCX 1.21 added UCX_CUDA_IPC_CACHE_MAX_REGIONS/MAX_SIZE (default unlimited). Related open #12849 and #12971. |
| #9050 | OMPI_{CC,CXX}=mpi{cc,cxx} is a fork-bomb | PHHargrove | - | Keep | (reported on 4.0.5): OMPI_CC=mpicc makes mpicc exec itself recursively (fork bomb) on v5.0.x and main; opal_wrapper has no recursion guard. Cheap hardening. |
| #9022 | Need a way to disable use of -mcx16 compile flag | opoplawski | - | Keep | (reported on 4.1.1): configure still adds -mcx16 to the global CFLAGS on x86_64 with no opt-out; the proposed --disable-cx16-atomics PR #9023 was closed unmerged. |
| #9021 | Failures on MPICH tests (large_count, structpack2, struct_no_real_types, get_elements_pairtype, indexed_misc) | karasevb | - | Keep | (master-2021, 4.1.x): MPICH datatype tests large_count, struct_no_real_types and get_elements_pairtype were confirmed failing on master (2022); no fix found. |
| #8820 | Building of openmpi rpm hits issue with python scripts in 3rd-party modules | acgoldma | - | Keep | (reported on 5.0.0a1): the bundled libevent event_rpcgen.py still has a '#!/usr/bin/env python' shebang and is installed to bin, which breaks RHEL/Fedora RPM builds. |
| #8720 | Derived types with CUDA-Aware MPI | ShatrovOA | - | Keep | (reported on 4.0.5, NVHPC 21.2): non-contiguous datatypes on GPU buffers are still packed with one accelerator memcpy per contiguous block on v5.0.x, so derived types on device memory are very slow. |
| #8467 | Compiling with AOCC (Clang-9) | adrianjhpc | - | Keep | (reported on 4.1.0, Cray ftn wrapping AOCC flang): libtool left wl empty for ftn. Tarballs are still built with Libtool 2.4.6, and the autogen.pl flang patch (6f8010c685) matches only flang, not ftn; Libtool 2.5.1 added ftn. Fix: add ftn to the autogen.pl patch or build tarballs with Libtool 2.5.1+. Status comment drafted, not posted. |
| #8439 | Detect (and ignore) stale Open MPI components | jsquyres | feature | Keep | (RFC, master-2021): detect and ignore stale component DSOs left by an older install. v5.0.x mca_base_component_repository.c still compares only the MCA version (2.1), with a TODO for a project-version check. Since v5.0.0 most components are static, so same-name stale DSOs are skipped, but DSOs of components that exist only in older releases still load. Status comment posted; cross-linked with #14299 and #475. |
| #9599 | mca_pml_ucx_preq_completion: Assertion `preq->tmp_req != ((void *)0)' failed | jeromerobert | question | Needs check | (4.1.x, UCX 1.10/1.11): rare preq->tmp_req NULL assert in mca_pml_ucx_preq_completion; the same code is on v5.0.x and maintainers never reproduced it. Test: persistent send/recv stress on a debug v5.0.x build across nodes. |
| #9479 | openMPI 4.1 + UCX 1.11 OMPIO hangs writing shared file on Lustre | denisbertini | - | Needs check | (4.1.x with UCX 1.11/1.13): OMPIO MPI_File_write_all hangs or runs 10x slower than ROMIO on Lustre 2.12. OMPIO is still the default on v5.0.x; never retested. Test: BTIO or write_at_all on Lustre with v5.0.x, OMPIO vs ROMIO. |
| #9062 | Large message collective performance drops when using coll/han | hjelmn | - | Needs check | (reported on 4.1.1): large-message (256K-1M) collectives much slower with coll/han than with tuned. HAN is the default on v5 (priority 35). Test: osu_allreduce on v5.0.x, defaults vs --mca coll ^han. |
| #8982 | btl/tcp: Assert with MPI_Sendrecv() | awlauria | - | Needs check | (master-2021, 4.1.x): assert(endpoint_sd_next == -1) in mca_btl_tcp_endpoint_accept is still in btl_tcp_endpoint.c on v5.0.x; seen only in IBM MTT (ibm/random/alloc_mem over tcp), never reproduced. main has connection-race fixes (d5919501db, a1616b1da1) that v5.0.x lacks. Test: loop alloc_mem on a debug v5.0.x build. |
| #8334 | AVX-based MPI_OP performance regression | rajachan | - | Needs check | (reported on 4.1.0): op/avx still enables AVX512 by default on v5.0.x; the ~20% LAMMPS slowdown on Skylake with gcc was never root-caused. PR #8376 (AVX512 off by default) closed unmerged. Test: LAMMPS on Skylake with v5.0.x, default op_avx_support vs 0x3f. |
By category: Decision needed 1; Keep 11; Needs check 2.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #8010 | reduce_scatter and reduce_scatter_block algorithms incorrectly handling noncommutative ops | wckzhang | bug | Decision needed | Duplicate of #7765 (pending decision) (reported on 4.1.x-2020): Default selection is fixed: 57b95bcb45 (first in v5.0.0, also on v4.1.x) makes the fixed rules use non-overlapping for RS and basic_linear for RSB with non-commutative ops (coll_tuned_decision_fixed.c:853 and :999 on v5.0.x, :1121 and :1268 on main). The remaining failures only occur when an algorithm is forced or chosen from a rules file. coll_tuned_reduce_scatter_decision.c:135-145 and coll_tuned_reduce_scatter_block_decision.c:128-138 still dispatch RS 2/3/4 and RSB 2/4 without a commutativity check. That is the root cause tracked in #7765. RS 2/3 are documented commutative-only (coll_base_reduce_scatter.c:129,409). RS 4 butterfly and RSB 2/4 claim non-commutative support in their headers but failed the ibm nocommute_stride tests. Whether those three have a real ordering bug could not be decided from source without running the tests. |
| #7958 | Undefined behaviour reported by clang UBSan in pml_ob1_hdr.h | cniethammer | - | Keep | (reported on master-2020): mca_pml_ob1_rdma_btls() leaves reg_handle = NULL for BTLs without btl_register_mem (ompi/mca/pml/ob1/pml_ob1_rdma.c on v5.0.x, loop around the if (btl->btl_register_mem) block). mca_pml_ob1_send_request_start_rdma() then passes that NULL handle with reg_size = btl_registration_handle_size (0) to mca_pml_ob1_rget_hdr_prepare (pml_ob1_sendreq.c:902 v5.0.x, :903 main), which does an unguarded memcpy(hdr + 1, local_handle, local_handle_size) (pml_ob1_hdr.h:266 v5.0.x, :267 main). memcpy with a NULL source is UB even for size 0, which is what UBSan flags. The same pattern exists in mca_pml_ob1_rdma_hdr_prepare (pml_ob1_hdr.h:421/422). No fix in git history; functionally benign. |
| #7766 | Option to enable two level namespace on Mac OS | srherbener | - | Keep | (reported on 4.0.3): config/ompi_setup_fc.m4 on v5.0.x lines 321-349 (main 323-351) still, on apple-darwin, appends -Wl,-flat_namespace to LDFLAGS for building all of Open MPI's libraries and adds it to the mpifort wrapper and ompi-fort.pc FCFLAGS via OPAL_WRAPPER_FLAGS_ADD (v5.0.x :341), with no configure opt-out; it dates from 6164dc4960 (2014, fix for #259, Fortran sentinels such as MPI_BOTTOM). The later -Wl,-commons,use_dylibs (+ -ld_classic) probe (v5.0.x :98-117, called at :408) only adjusts the build LDFLAGS and is not added to the wrapper, so flat_namespace remains what makes sentinels work for user programs. Keep vs Stale: Keep argues it is a real, still-present leak of a global linker mode into every mpifort-linked user dylib on a supported platform (Homebrew-style clang C++ plus gfortran builds are common). Stale argues it has only two reporters since 2020, no maintainer reply, last activity 2021-11, and removing flat_namespace risks silently wrong MPI_BOTTOM or MPI_IN_PLACE detection that only a macOS Fortran tester can validate. A no-rebuild workaround exists: mpifort compiler flags can be overridden with the OMPI_FCFLAGS env var (opal_wrapper.c:649-651) or by editing mpifort-wrapper-data.txt and ompi-fort.pc. Stale wins on balance. |
| #7765 | Dynamic Tuning (coll/tuned) behavior doesn't have any non commutative checks | wckzhang | bug | Keep | (reported on master-2020): Neither the forced-algorithm path nor the rules-file path checks op commutativity before dispatch. The ompi_coll_tuned_*_intra_do_this switches go straight to base algorithms: coll_tuned_allreduce_decision.c:134-148 (ring 4, segmented ring 5), coll_tuned_reduce_decision.c:158-178 (chain/pipeline/binary/binomial), coll_tuned_reduce_scatter_decision.c:135-145 and coll_tuned_reduce_scatter_block_decision.c:128-138 on v5.0.x, with identical code on main. The base ring allreduce (coll_base_allreduce.c:342) and reduce_scatter recursive-halving/ring (coll_base_reduce_scatter.c:129,409) are documented commutative-only and have no fallback. Only rabenseifner allreduce/reduce (coll_base_allreduce.c:991, coll_base_reduce.c:830) and RSB recursive halving (coll_base_reduce_scatter_block.c:419) guard themselves, and only decision_fixed.c (v5 :80,:691,:853,:999) checks ompi_op_is_commute. Main has a newer problem that also affects defaults. The disjoint allreduce fixed rule picks alg 7 allgather_reduce for non-commutative ops (coll_tuned_decision_fixed.c:263-270, from 539038a4c8). That algorithm reduces in reverse rank order (coll_base_allreduce.c:1310-1315: buf0 = b[t] op buf0), so it returns b[n-1] op ... op b0. |
| #7693 | btl/uct using wrong topology | rhc54 | bug | Keep | (reported on master-2020): mca_btl_uct_component_open() still sizes contexts from the whole-node core count: hwloc_get_nbobjs_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE) at opal/mca/btl/uct/btl_uct_component.c:124 (v5.0.x) and :166 (main), then divides by num_local_peers+1. opal_hwloc_topology is the node topology shared from the PRRTE daemon, so a job-level cpuset/cgroup or a binding is ignored. The 2025 btl/uct rework on main (112b4f813b) did not touch this heuristic. Impact is performance only (too many UCT contexts) and btl_uct_num_contexts_per_module is a workaround. Note PMIX_JOB_CPUSET does not exist in the bundled PMIx headers. |
| #7672 | Inconsistent & Perhaps buggy dynamic tuning behavior | wckzhang | - | Keep | (reported on master-2020): Gather/scatter were fixed by #7758 (2b1f053, first in v5.0.0), and the per-collective message-size table is now documented in docs/tuning-apps/coll-tuned.rst:172-189 (v5.0.x) and docs/tuning-apps/collectives/tuned.rst:224-237 (main). Scan and Exscan still compute the rule lookup size as dtype size * comm size and ignore count: coll_tuned_decision_dynamic.c:714 (exscan) and :758 (scan) on v5.0.x, :723 and :767 on main. bosilca confirmed in-thread that these two are copy/paste outliers. Because comm size is already the outer key in the rules file, Scan/Exscan rules cannot vary by message size at all. |
| #7526 | Disable cuda coll parameter doesn't work | souravzzz | - | Keep | (reported on 4.0.3): On v5.0.x, coll_cuda_disable_cuda_coll is registered at ompi/mca/coll/cuda/coll_cuda_component.c:83-88 into mca_coll_cuda_component.disable_cuda_coll (coll_cuda.h:138). Nothing reads it, and mca_coll_cuda_comm_query (coll_cuda_module.c:80-90) always returns a module. On main, the renamed coll_accelerator_disable_accelerator_coll (coll_accelerator_component.c:91-96, coll_accelerator.h:171) is also never read. mca_coll_accelerator_comm_query (coll_accelerator_module.c:74-90) only disqualifies itself for the null accelerator component or the mpi_assert_no_accel_buf assertion. The help text ("Automatically handle the CUDA buffers...") also describes the opposite of the parameter name. --mca coll ^cuda (v5) or ^accelerator (main) is the working way to disable the component. |
| #7481 | Configure doesn't fail when enable-mca-no-build mca and component names are invalid | Akshay-Venkatesh | bug | Keep | (reported on master-2020, 4.0.2): config/opal_mca.m4:102-115 on v5.0.x (103-116 on main) splits each --enable-mca-no-build item and blindly sets DISABLE_<fw> or DISABLE_<fw>_<comp>, which is only consumed by MCA_COMPONENT_BUILD_CHECK (v5.0.x :922-923, main :970-971); a typo such as btl-utc sets an unused variable and configure succeeds. The --enable-mca-dso (:195) and --enable-mca-static (:223) lists are equally unchecked; only --enable-mca-direct is validated (:465-470, direct component requested but not found). No fixing commit or PR found. Caveat: the same flag is forwarded to the bundled PMIx/PRRTE sub-configures (3rd-party/openpmix/config/pmix_mca.m4, 3rd-party/prrte/config/prte_mca.m4 have their own --enable-mca-no-build), so names valid only for PMIx/PRRTE (e.g. plm-slurm) must not become a hard error in the OMPI configure. |
| #7472 | Delete smcuda BTL and put CUDA functionality in vader BTL | jsquyres | - | Keep | (roadmap) (reported on master-2020): smcuda still ships on v5.0.x and main (opal/mca/btl/smcuda, owner.txt NVIDIA/active) and the CUDA docs still list smcuda as the ob1 CUDA-aware shared-memory path (main docs/tuning-apps/accelerators/cuda.rst:187). v5.0.x btl/sm has no accelerator code and no smsc/accelerator component. On main/v6.0.x the plan was realized differently: smsc/accelerator (6264853bf6, PR #12949, first in v6.0.0rc1) lets btl/sm set MCA_BTL_FLAGS_ACCELERATOR_GET (main opal/mca/btl/sm/btl_sm_component.c:353-354), but it is opt-in (smsc_accelerator_component.c:24 default priority 0, disqualifies itself at :79) and docs/tuning-apps/accelerators/rocm.rst:81-100 still offers smcuda as the alternative. |
| #7423 | Segv / freeing memory from dlclose()'ed common libraries | jsquyres | bug | Keep | (reported on master-2020): The root cause is unchanged: var_destructor frees *mbv_storage->stringval for string vars (opal/mca/base/mca_base_var.c:1863-1869 on v5.0.x, :1884 on main), and nothing deregisters vars that a common library registers under its own group (e.g. opal_common/ucx) before that library is unloaded along with its last component DSO. Every in-tree common library with string vars now works around it. common/monitoring registers under the pml/monitoring group (common_monitoring.c:259-265, from #7284). common/ucx heap-allocates the storage pointer (common_ucx.c:106-113 on v5.0.x, from 32aba0bdf2 for #11632, v6.0.0rc1). common/ofi does the same (common_ofi.c:342-351). So there is no reachable in-tree trigger left, and the issue has no reproducer or activity since 2020-02. Related but separate: v5.0.x mca_base_component_repository.c:299 still looks up the var group with a NULL project, so component vars are not deregistered before dlclose in that path; this is fixed on main by c454fd408f (#13897, with a v6.0.x cherry-pick 62f5f24742) and is a v5.0.x backport candidate. |
| #7394 | shmem related files left in /dev/shm after interrupt | manomars | - | Keep | (reported on 3.1.x, 4.0.1, 4.0.2): The leaked 4 MiB+8 files are the per-process btl/vader (now btl/sm) segment. On v5.0.x btl_sm_component.c:379-381 (main:395-397) registers sm_segment.<host>.<uid>.<jobid>.<noderank> for PMIx cleanup and then calls opal_shmem_segment_create(), but shmem/posix ignores file_name and instead creates /open_mpi.%04d (shmem_posix_common_utils.c:78 v5.0.x, :74 main; shmem_posix_module.c:168), so the registered cleanup path never matches the real /dev/shm object. The segment is only shm_unlink()ed at btl finalize (btl_sm_module.c:367), so a killed job leaks it; after OPAL_SHMEM_POSIX_MAX_ATTEMPTS=128 (shmem_posix.h:33) leaked names, posix selection fails. Only affects the non-default --mca shmem posix (mmap has higher priority). Not runtime-verified. |
| #7381 | OpenMPI OSHMEM 4.0.2 missing some pshmem functions | lriesen | - | Keep | (reported on 4.0.2): oshmem/shmem/c/shmem_info.c (v5.0.x:18,29; main:19,30) defines shmem_info_get_version/shmem_info_get_name with no OSHMEM_PROFILING block, unlike every other API file (e.g. shmem_finalize.c uses #pragma weak shmem_finalize = pshmem_finalize plus profile-defines.h). Neither oshmem/include/pshmem.h nor oshmem/shmem/c/profile-defines.h has pshmem_info_get_* entries on v5.0.x or main, so libshmem exports no pshmem_info_get_version/pshmem_info_get_name. The file has not changed functionally since 2dc2159b4c (spec 1.2 import); no fixing PR found. |
| #8098 | BTL UCT got active message id 26, but no handler installed | abouteiller | bug | Needs check | (reported on master-2020): btl/uct only registers AM ids 0x0d (MCA_BTL_UCT_FRAG) and 0x0e (MCA_BTL_UCT_CONNECT_RDMA) (v5.0.x opal/mca/btl/uct/btl_uct_types.h:32-34, same on main). Ids 24-28 are never valid. In each dump, the reported id equals the low 5 bits of the byte just before the payload in the repeating data pattern (0x58->24, 0x1c->28, 0x9b->27). So UCX received raw application data with no UCT header, which points to a corrupted or misrouted send rather than a missing registration. One relevant v5.0.x difference: the context-0 FRAG handler is registered with UCT_CB_FLAG_ASYNC (btl_uct_tl.c:473-474) on a UCS_THREAD_MODE_SINGLE worker (btl_uct_tl.c:327). mca_btl_uct_am_handler (btl_uct_component.c:310) calls into ob1, which may post sends. main/v6.0.x registers every FRAG handler with MCA_BTL_UCT_CB_FLAG_SYNC (btl_uct_device_context.c:60-61, via 112b4f813b/PR #13335) and also has endpoint race fixes (f99cbcd172, 2a128b1ec6, d586292e60) that are not on v5.0.x. I cannot run it, so I have not shown that any of these is the cause. |
| #7900 | Terminal size under OpenMPI | mrzv | feature | Needs check | (PRRTE) (reported on 4.0.4): Open MPI has no code in this path. On v5.0.x and main, PRRTE sets up the child's stdout. PTY support is on by default (prrte config/prte_configure_options.m4:175-184) and odls sets opts.usepty=PRTE_ENABLE_PTY_SUPPORT for every child (v5.0.x prrte a99bf3a006 src/mca/odls/base/odls_base_default_fns.c:1584,2238, main prrte f33bfe488 :2326,3063). prte_iof_base_setup_prefork (src/mca/iof/base/iof_base_setup.c:103-110 on both) copies TIOCGWINSZ from the daemon's own STDOUT_FILENO into pmix_openpty, the same logic v4.0.x had in orte. A remote prted's stdout is not a tty, so ioctl fails, wp stays NULL, and remote ranks get a pty with a 0x0 window. Local ranks copy prterun's size only if its stdout is a tty. The reporter hit 0x0 even locally with identical v4 code, so that case needs a runtime test. |
By category: Keep 4; Needs check 2.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #7143 | OSHMEM: SHPCLMOVE returns incorrect error code for 2 cases | sssharka | - | Keep | (reported on 4.0.1): SHPCLMOVE returns -2 instead of -4 (block already freed) and -5 (not at block start). v5.0.x oshmem/shmem/fortran/shpclmove_f.c still maps any shrealloc failure to -2 and never returns -4 or -5. Unchanged; no comments. |
| #6728 | shmem.h uses C++ <complex> type for C++, but standard uses _Complex from <complex.h> | sam6258 | bug | Keep | (reported on master-2019, 3.0.x-4.0.x): In C++, OSHMEM_COMPLEX_TYPE is std::complex<T>, so Sandia C++ tests calling shmem_complex*_to_all with _Complex fail to compile. hppritcha asked Mellanox about plans in 2021 with no answer. oshmem/include/shmem.h.in is unchanged in v5.0.x and main. |
| #6464 | DDT shows message queues with pml ob1 but not with ucx and yalla. | bartoldeman | bug | Keep | (reported on 2.1.1, 3.1.x): DDT message queues are empty with pml ucx because the msgq DLL walks mca_pml_base_send/recv_requests, which only ob1/cm populate. On v5.0.x and main, libompi_dbg_msgq still reads those lists and pml/ucx still does not use them (grep). Proposed opt-in MCA param was never done. yalla is gone. |
| #6366 | Unexpected behavior of DataType.getSize in Java | malinkallen | question | Keep | (reported on 3.1.3): Java Datatype.getSize/getExtent count elements, not bytes, which conflicts with byte-based createResized/struct. Options discussed: fix the docs, add byte APIs, or switch to bytes in v5. Last activity was a 2021 question about v5.0.0. v5.0.x Datatype.java still returns getSize(handle)/baseSize. |
| #6710 | MPI.COMM_SELF.Spawn() error | connorourke | - | Needs check | (reported on 3.1.4): mpi4py loop of COMM_SELF.Spawn fails after ~3276 spawns (proc on host unknown, ompi_dpm_dyn_init Unreachable), also on one node. No maintainer reply since 2019-06. Test: run the toy script (or a C loop of MPI_Comm_spawn+disconnect) for >5000 iterations on v5.0.x; close if it passes. |
| #6240 | TCP: unexpected process identifier in connect_ack | abouteiller | bug | Needs check | (reported on master-2019): TCP BTL deadlock with 'received unexpected process identifier' on multi-rail hosts with non-routable ifaces (virbr0, docker0); if_include worked around it. PR #7134 (reachable pairing, v5.0.0) meant to fix; mkre 2020 still hung on master without excluding docker0. Test: v5.0.x mpirun -mca btl tcp,self IMB pingpong across 2+ nodes with docker0/virbr0 up, no if_include/exclude. |
By category: Keep 1; Needs check 2.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #6143 | When using libtool to compile against Open MPI, libopen-pal and libopen-rte are explicitly linked in | jsquyres | bug | Keep | (reported on 3.1.x / 4.0.x (master-2018)): Libtool .la files make apps like HDF5 link libopen-pal/libopen-rte directly, which breaks cross-version runs. The Debian link_all_deplibs patch and dropping .la files were discussed with no resolution. libopen-rte is gone in v5, but libopen-pal 70:3:30 (so.40) became 80:0:0 (so.80) while libmpi stayed .40. |
| #5902 | TCP BTL progress thread segv's | jsquyres | bug | Needs check | (reported on master-2018): Segv in ~60% of runs at MPI_Finalize with btl tcp,self and btl_tcp_progress_thread=1 (2 nodes x16); suspected TCP progress thread still running after component dlclose. Maintainers pinged, no fix identified; option still exists in v5.0.x. Test: loop that config on v5.0.x/main, check for finalize segv. |
| #5172 | Need a way to limit vma tree size (specifically for rgpusm if that's possible) | Akshay-Venkatesh | feature | Needs check | (reported on 2.0.x, 3.0.x): Stale cuIpcMemHandles in rgpusm pile up on cudaMalloc/Isend/Irecv/cudaFree loops; asked for vma tree size knobs. hjelmn agreed 2018-06, jsquyres pinged 2018-09, no follow-up. v5.0.x still has smcuda+rgpusm (rcache_size_limit, leave_pinned, empty_cache) but vma_tree_items params are gone (38d9b10db8). Test: run the loop on v5.0.x with pml ob1 btl smcuda over 2 GPUs and watch GPU memory. |
By category: Keep 2; Needs check 1.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #3612 | XL Fortran Compiler F08 Bindings broken for OpenSHMEM | jjhursey | bug | Keep | (reported on master-2017, 3.0.0rc1, master-2021): xlf defaults to no-underscore Fortran symbols, and oshmem/shmem/fortran only emits _, __ and _f names. jjhursey found the fix sprawling. In 2021 on master he confirmed it again and showed the documented FC='xlf -qextname' workaround also fails to link. docs/release-notes/compilers.rst in v5.0.x/main still documents it and links #3612. |
| #3357 | --enable-script-wrapper-compilers and oshmem build failures | bwbarrett | bug | Keep | (reported on master-2017): --enable-script-wrapper-compilers breaks the oshmem wrappers. On v5.0.x and main, oshmem/tools/wrappers/Makefile.am still always symlinks shmemcc/oshcc etc. to opal_wrapper, but opal_wrapper is not built when script wrappers are on (opal/tools/wrappers/Makefile.am), so the links dangle. No comments. |
| #4610 | [v2.1.1] Crash at opal_convertor_generic_simple_position while sending/receiving GPU memory with multi-threads at comm_dup channel | woso | bug | Needs check | (reported on 2.1.1, 3.0.0, 1.10.7): Crash in opal_convertor_set_position with THREAD_MULTIPLE, comm_dup and cudaMalloc buffers over tcp and smcuda. Fragments from two threads appeared to overlap. Last activity 2018-04. To decide, run the attached reproducer on v5.0.x with --mca pml ob1 --mca btl self,tcp,smcuda and CUDA buffers. |
By category: Keep 7.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #2366 | MPI_Get_elements for data types used in reduction functions | nysal | bug | Keep | (reported on master-2016): OMPI returns 1 element per MPI_DOUBLE_INT, MPICH returns 2 (mpich test get-elements-pairtype.c). bosilca: OMPI treats pair types as unitary predefined, suggested asking the Forum; no follow-up since 2016-11. v5.0.x ompi_datatype_module.c DECLARE_MPI2_COMPOSED_STRUCT_DDT still sets PREDEFINED. Standard defines pair types as-if struct of 2 types. |
| #2335 | MPI_Reduce op x datatype parameter check needs improvement | jjhursey | feature | Keep | (reported on master-2016): Request to reject invalid op/datatype pairs such as MPI_SUM on MPI_BYTE. bosilca explained MPI_BYTE maps to the UINT8 id. ompi_datatype_internal.h in v5.0.x and main still defines OMPI_DATATYPE_MPI_BYTE as OMPI_DATATYPE_MPI_UINT8_T and ompi_op_is_valid is unchanged. |
| #2289 | openmpi 2.0.1 over/under linkage | opoplawski | - | Keep | (reported on 2.0.1): Fedora rpmlint underlink/overlink warnings on libmca_common_sm, libompitrace, etc. Dynamic components got project-lib LIBADD in e1d079544b (v3.1.0, #3705); libompitrace is gone in v5.0.x. But opal/mca/common/{sm,ucx,ofi} and ompi/mca/common/ompio Makefile.am on v5.0.x and main still do not LIBADD libopen-pal/libmpi. Confirm with ldd -r on install. |
| #2056 | [feature request] packaging for multiple Fortran compilers | jeffhammond | feature | Keep | (reopened 2026-09-28 by bosilca after jeffhammond's reply; reported on 2.0.x): request for one Open MPI C library usable with several Fortran compilers. Our 2026-09-28 close cited the docs' wait-and-see on the MPI Forum Fortran ABI; jeffhammond (reporter) says that ABI is settled (incomplete by design) and the real path is standalone Fortran bindings (e.g. VAPAA, MPICH work) on top of the C ABI. main/v6.0.x already implement the Fortran-related C ABI pieces those need (MPI_Abi_get/set_fortran_info, MPI_Abi_get/set_fortran_booleans, MPI_Comm/Type_toint/fromint); v5.0.x has none. Open items: whether OMPI's own Fortran bindings should be buildable per-compiler on top of libmpi_abi, and updating docs/building-apps/mpi-forum-abi.rst (says the Forum still discusses the Fortran ABI). |
| #2025 | can opal_progress() call opal_progress()? (test hanging in MPI_COMM_DUP) | jsquyres | bug | Keep | (reported on master-2016, 3.0.x, 3.1.x): Started as a keyval test hang, then became MT/idup CID deadlocks. bosilca confirmed overlapping MPI_Comm_idup (or idup+Ibcast) is valid MPI and can deadlock. ggouaillardet's PoC patch was never merged (last activity 2019). In v5.0.x/main, ompi_comm_nextcid_nb only uses the extended-CID path when comm==NULL, so idup keeps the old algorithm. Repro: two back-to-back MPI_Comm_idup on COMM_WORLD with 4 ranks. |
| #1763 | Tuned bcast data corruption | nysal | bug | Keep | (reported on master-2016 (2.x)): Tuned bcast segmented algorithms corrupt data when ranks use different count/datatype pairs with same type signature. bosilca 2020: cannot be fixed efficiently; PoC #1789 abandoned; only libnbc got a workaround (350ef67fe0). v5.0.x tuned bcast decision still picks pipeline/split-bintree for large msgs. Reproducer in issue. |
| #1666 | MPI_Accumulate Error when using datatype "MPI_SHORT_INT" | Zhiming-Wang | bug | Keep | (reported on master-2016 / 2.0.x): MPI_Accumulate MINLOC with MPI_SHORT_INT broken (padding/non-contig pair type). Forum agreed it must be supported; pt2pt fixed, rdma not; #2832 added an abort workaround. osc_base_obj_convert.c in v5.0.x, v6.0.x and main still has the TODO abort for these pair types. Last activity 2018. |
By category: Keep 2.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #1141 | corrected/expanded message about shmem temporary directory | loveshack | - | Keep | (reported on master-2015): Doc fix: shmem/mmap help text says $TEMPDIR but opal uses $TMPDIR. No comments. v5.0.x and main opal/mca/shmem/mmap/help-opal-shmem-mmap.txt line 34 still say $TEMPDIR. Trivial help-text PR; the suggested orte_tmpdir_base param name no longer exists (PRRTE era) so omit or update it. |
| #475 | No protection for opening plugins from wrong OMPI version | opoplawski | feature | Keep | (reported on master-2015): Stale plugins from an older install in the same prefix get dlopened and crash. Devs agreed in 2015 to check project version, but v5.0.x and main mca_base_component_repository.c still only compare MCA 2.1 and still say 'TODO -- add checks for project version'. v5 builds static by default, but DSO builds are still exposed. |
By category: Keep 6; Needs check 3.
| # | Title | Author | Type | Category | Note |
|---|---|---|---|---|---|
| #321 | Fix the attribute mess | bosilca | bug | Keep | (low priority): George (2014) said the ompi-release#142 attribute fix was a band-aid and the real fix is a per-object flag for attribute changes during callbacks. The only reply since was a 2016 milestone move. main still uses a single global attribute_lock, and MPI_Comm_free docs still say delete_fn must not add or delete attributes on the same object. |
| #160 | TCP BTL does not support Linux virtual interfaces | ompiteam | bug | Keep | TCP BTL gets confused by Linux virtual/alias interfaces (eth0:0) because it keys on the kernel index (kindex). bwbarrett reopened it in 2018 and explained the if_include/address-selection flaw. The same limitation is still described in a comment in mca_btl_tcp_create() in opal/mca/btl/tcp/btl_tcp_component.c on main. No fix found. |
| #139 | Use new Autoconf 2.69 Fortran m4 macros | ompiteam | feature | Keep | (low priority): Request to use the Autoconf 2.69 Fortran macros (AC_FC_MODULE_FLAG, AC_FC_IMPLICIT_NONE, etc.) once the minimum was raised. Main now requires 2.69 but still uses the custom config/ompi_fortran_find_module_include_flag.m4 and none of the macros. A configury cleanup. |
| #124 | mpi_show_handle_leaks=1 doesn't show all leaks | ompiteam | bug | Keep | (low priority): mpi_show_handle_leaks originally checked only comms, infos and files. Main now also checks windows (d7205f90f1, 2015). Datatypes, groups, ops, errhandlers and keyvals are still not reported. A valid enhancement. |
| #46 | Add FAQ entry about firewalls | ompiteam | documentation | Keep | request for firewall documentation: (1) disable firewalls between nodes, (2) how to run through a firewall using port ranges. Item 1 is covered in docs/launching-apps/troubleshooting.rst. Item 2 is not: docs/ never mentions port ranges (btl_tcp_port_min etc.) or firewalls in that sense. This is a small docs task. |
| #24 | Implement MPIX_GREQUEST_START | ompiteam | feature | Keep | request for an MPIX_Grequest_start user API with a progress callback. devreal asked to expose it in 2020 and jsquyres agreed. On main, ompi/request/grequestx.c exists behind --enable-grequest-extensions (PRs 5337/5353/6991/7546) but is internal only: there is no public MPIX API and nothing outside ompi/request uses it. |
| #194 | Non-uniform BTL usage not working | ompiteam | bug | Needs check | BTLs treated a missing peer modex key as a fatal error, which breaks mixed-BTL jobs. usnic was fixed in r31258 and tcp in r31753, and tcp on main now ignores OPAL_ERR_NOT_FOUND. But btl/ofi on main still aborts on any modex_recv failure (btl_ofi_module.c) and uct errors out. Test with an MPMD run where only one side enables ofi or uct. |
| #189 | Remove PERUSE in favour of upcoming MPI_T events implementation | ompiteam | - | Needs check | 2014 proposal to remove PERUSE in favour of MPI_T. bosilca objected. jsquyres (2022) said close when #8057 (MPI_T events) merges. #8057 and #13133 were closed as superseded by #14083, which merged (783ecc7605). PERUSE is still in ompi/peruse on main, but MPI_T events currently cover only comm/win/init/errhandler events, not PML-level ones. Maintainers need to decide whether to remove PERUSE now. |
| #123 | Fix MPI_File auto-close on Finalize | ompiteam | bug | Needs check | (v5.0.x): MPI_File left open at MPI_Finalize is auto-closed after ompi_instance_count drops to 0, and ROMIO's close still calls MPI_Barrier, which trips the after-finalize check. v5.0.x still ships ROMIO 3.4.1 (removed from main by 73a55fe4fe); default ompio is unaffected. Test the 2010 reproducer on v5.0.x with --mca io romio341. |
Issues that were closed since then, and issues opened since then. Every refresh records this, and the wiki keeps one page per day, so the pages form a running history.
| # | Title | Closed | Reason |
|---|---|---|---|
| #14558 | Heap buffer overflow in MPI_Comm_set_name(MPI_COMM_WORLD, …) with names longer than 15 characters (5.0.11, macOS arm64) | 2026-10-08 | completed |
| #14203 | Open MPI osc/rdma over OFI crashes after MPI_Win_allocate with multiple local ranks | 2026-10-08 | completed |
| #14181 | MPI_Accumulate loses updates with a one-element indexed datatype over UCX/InfiniBand |
2026-10-08 | completed |
| #14169 | --map-by dist not working (anymore) / --map-by device=gpu wanted | 2026-10-08 | completed |
| #14008 | MPI_Get_hw_resource_info returns an empty info object (no resource keys populated) | 2026-10-08 | completed |
| #13783 | Problems running CP2K-CI with OpenMPI parallelization | 2026-10-08 | completed |
| #13621 | Overflow in memheap_base_mkey.c mca_memheap_modex_recv_all | 2026-10-08 | completed |
| #13544 | Why does UCC consistently insist on using software to simulate PUT/GET operations? | 2026-10-08 | completed |
| #13451 | Regression: Spawned process are not killed on timeout | 2026-10-07 | completed |
| #13432 | MPI_Waitall doesn't return MPI_ERR_IN_STATUS for persistent request | 2026-10-07 | completed |
| #13399 | Variable forwarding in case of spawn not working for children | 2026-10-07 | completed |
| #13385 | Open-Coarrays and gfortran-15 with OpenMPI | 2026-10-08 | not planned |
| #13371 | Q: MPI-IO hint to disable/enable data sieving | 2026-10-08 | completed |
| #13358 | mpirun is terminated saying stack smashing detected when hostname is long | 2026-10-08 | not planned |
| #13347 | osc/ucx does not honor thread level requested | 2026-10-07 | completed |
| #13346 | openpmix 6.0.0 and openMPI 5.0.8 configure problem | 2026-10-08 | completed |
| #13336 | MPI_Type_get_envelope_c is wrong | 2026-10-07 | completed |
| #13326 | Is the error "Error string: /lib64/libcuda.so.1: undefined symbol: cuIpcOpenMemHandle_v2 CUDA-aware support is disabled." due to unavailability of module nv_peer_mem or nvidia-peermem in the nvidia-driver | 2026-10-08 | not planned |
| #13317 | Running openmpi with different rank on two docker on same host | 2026-10-08 | not planned |
| #13196 | mpiexec 4.1.3, how to avoid transfer PMIX_INSTALL_PREFIX to the child process of orted | 2026-10-08 | completed |
| #13191 | SEGV in ompi_request_default_test_all() when triggering IPoIB networking problem during ucc_perftest run | 2026-10-07 | completed |
| #13190 | Issue with NVFORTRAN With NVHPC 25.3 | 2026-10-07 | completed |
| #13183 | RDMA Read Error on DGX B200 systems | 2026-10-08 | not planned |
| #13170 | Memory leak with ROCM-aware OpenMPI with UCX 1.17.0 | 2026-10-08 | not planned |
| #13140 | Adding another example for testing connectivity | 2026-10-07 | completed |
| #13128 | MPI One-sided applications only work via Infiniband adapter as expected, but not via Ethernet adapter | 2026-10-07 | not planned |
| #13118 | Help with launching MPI application | 2026-10-08 | not planned |
| #13076 | Freeze or error using function MPI_Comm_connect | 2026-10-07 | completed |
| #13067 | Not all communicators subscribe to info keys | 2026-10-07 | completed |
| #13044 | Segfault sometimes when building with CUDA support | 2026-10-08 | completed |
| #13043 | static libraries, F90 module, and alignment warnings | 2026-10-07 | completed |
| # | Title | Author | Created |
|---|---|---|---|
| #14588 | flang: make check dies using flang-23 | hppritcha | 2026-10-07 |