Repository navigation
PR Snapshot 2026 10 08
Tomislav Janjusic edited this page Oct 8, 2026
·
1 revision
Snapshot date: 2026-10-08
Total open PRs: 89
PRs that left the open list (merged or closed) and PRs opened since then. Every snapshot records this, so the dated pages form a running history.
| # | Target | Title | Outcome |
|---|---|---|---|
| #14557 | v6.0.x | [v6.0.x] btl/tcp: let the kernel pick the source port at connect() time | Merged 2026-10-05 |
| #14556 | main | btl/sm: add the creator's PID to the backing file name | Closed unmerged 2026-10-07 |
| #14555 | main | pml/ob1: defer accelerator stream and event creation to first use | Merged 2026-10-06 |
| #14550 | main | CI: add CFLAGS to capture format-security | Closed unmerged 2026-10-06 |
| #14542 | main | fortran: don't copy out values the C call never wrote | Merged 2026-10-05 |
| #14494 | main | Add memory clobber to opal_atomic_ll for arm and PPC | Merged 2026-10-05 |
| #14377 | main | Establish connections lazily: defer the modex wait and add_procs out of MPI_Init | Merged 2026-10-05 |
| #14332 | main | Work around buggy libofi CXI provider | Closed unmerged 2026-10-06 |
| #14195 | main | btl/tcp: fix the multi-module wire-up race (replace the #5917 workaround) | Closed unmerged 2026-10-05 |
| #14163 | v5.0.x | [v5.0.x] Fix file descriptor leaks and teardown ordering in MPI instance finalize | Closed unmerged 2026-10-06 |
| #13969 | main | opal/lifo: fix missing acquire barriers in pop on weak-memory architectures | Closed unmerged 2026-10-07 |
| #13789 | main | opal/sys: add RISC-V 64-bit atomic LL/SC and full atomic support | Closed unmerged 2026-10-07 |
| #13195 | main | btl/uct: allow connections to be formed using a separate memory domain | Closed unmerged 2026-10-05 |
| #12259 | main | Fix incorrect TCP connections. | Closed unmerged 2026-10-05 |
| #12119 | main | Allow MPI_Comm_disconnect on intracoms. | Closed unmerged 2026-10-05 |
| #9934 | main | tuned decision for single node communicators | Merged 2026-10-07 |
| #6426 | main | Add MCA parameters to define the size of memcpy chunks. | Closed unmerged 2026-10-05 |
| # | Target | Title | Author |
|---|---|---|---|
| #14592 | v6.0.x | [v6.0.x] coll/tuned: Change the allreduce default collective algorithm selection | app/ompi-backport-bot |
| #14590 | main | PRRTE/PMIx: advance shas | yinliaws |
| #14585 | main | Fix noncommutative operator issues | devreal |
| #14581 | v6.0.x | [v6.0.x] Backport/pr 14092 to v6.0.x | hppritcha |
| #14580 | main | BINDINGS: make prefixing method more flexible | hppritcha |
| #14578 | v6.0.x | [v6.0.x] info: report only the MPI_INFO_ENV fields that were populated | hppritcha |
| #14577 | v6.0.x | [v6.0.x] comm: return an error when the nvlink domain MCA var has no value | hppritcha |
| #14576 | v6.0.x | [v6.0.x] Backport/pr 14424 to v6.0.x | hppritcha |
| #14575 | v6.0.x | [v6.0.x] pml/ob1: defer accelerator stream and event creation to first use | app/ompi-backport-bot |
| #14570 | v6.0.x | [v6.0.x] comm_cid: fix leak of the PMIx_Value_string() result | hppritcha |
| #14569 | v6.0.x | [v6.0.x] ompi/comm: release c_keyhash instead of destructing it | hppritcha |
| #14568 | main | dpm: name each peer once when disconnecting an intracommunicator | bosilca |
| #14566 | v6.0.x | [v6.0.x] Fix singleton tmp files cleanup | hppritcha |
| #14564 | main | MPI_T: fix object-bound handle validation, index error codes, and monitoring add_procs | bosilca |
| #14563 | main | pml/ob1: add MPI_T event producers for point-to-point | bosilca |
All PRs below are open/unmerged — "merged" in any column always refers to a related PR, never the row itself.
| Field | Meaning |
|---|---|
| Target branch (section header) | PR's actual base branch — authoritative |
Target:/target: label |
Auto-label mirroring base branch (redundant, just two casings) |
backport: vX.Y.x label |
Intent to backport after merge — not yet done |
| Created (age) | Open date; age = days/months/years since snapshot |
| LOC (+/-) | Lines added / deleted |
| Merge/Backport Status | Source PR + sibling-backport state, or why it's standalone |
| Review Status | ⏳/✅/🔴 decision, who approved/requested changes/is awaiting review |
| Draft/WIP | Split into its own section at bottom |
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #14590 | PRRTE/PMIx: advance shas | yinliaws | 2026-10-07 (1d) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +2 / -2 | ⏳ Review required — no reviewer assigned, no reviews yet | advance openpmix sha to 1244a4498 advance prrte sha to 6213a36bd1 Picks up the PRRTE CALLED_ABORT termination fix (openpmix/prrte#2815): a process that calls abort is now counted as terminated, so prterun no longer wa... |
| #14585 | Fix noncommutative operator issues | devreal | 2026-10-07 (1d) | Target: main, backport: v5.0.x, backport: v6.0.x | Not yet backported - pending merge to main first. | +149 / -11 | ⏳ Review required — awaiting: @bosilca | An audit by Claude found issues with noncommutative ops in allreduce: - ompi_coll_base_allreduce_intra_ring and ompi_coll_base_allreduce_intra_ring_segmented do not support noncommutative operators but did not fal... |
| #14580 | BINDINGS: make prefixing method more flexible | hppritcha | 2026-10-06 (2d) | Target: main | - | +49 / -12 | ⏳ Review required — awaiting: @devreal | to allow for MPI_, MPI_T, and MPIX_ Also need to update MPI_T Makefile.am |
| #14568 | dpm: name each peer once when disconnecting an intracommunicator | bosilca | 2026-10-05 (3d) | Target: main | - | +10 / -6 | ✅ Approved — approved by hppritcha | MPI_Comm_disconnect() on a dynamic intracommunicator hands PMIx_Fence() every peer twice. ompi_comm_set() aliases c_remote_group to c_local_group for an intracommunicator, and ompi_dpm_mark_dyncomm() fla... |
| #14564 | MPI_T: fix object-bound handle validation, index error codes, and monitoring add_procs | bosilca | 2026-10-02 (6d) | Target: main | - | +487 / -36 | ⏳ Review required — awaiting: @hppritcha, @jsquyres | Five correctness fixes in the MPI_T implementation and the monitoring component, plus tests for them. They were all found while probing what MPI_T can actually report about point-to-point traffic, which is why they ar... |
| #14563 | pml/ob1: add MPI_T event producers for point-to-point | bosilca | 2026-10-02 (6d) | Target: main | - | +794 / -41 | 🔴 Changes requested — awaiting: @hppritcha, @jsquyres | ob1 reports nothing through the MPI_T event interface. The event framework landed in #14083, but the point-to-point producers that were meant to use it did not: they were part of [#13133](https://github.com/open-mpi/o... |
| #14559 | Remove PERUSE interface | hppritcha | 2026-10-02 (6d) | Target: main | - | +7 / -939 | ⏳ Review required — no reviewer assigned, no reviews yet | The PERUSE (Performance Revealing Unified Subscription Environment) interface has been superseded by the MPI_T events implementation and is removed. The --enable-peruse configure option is no longer available. Applica... |
| #14553 | fix: ireduce nonblocking tests failure | Saikiran-hpe | 2026-10-01 (7d) | Target: main | - | +8 / -1 | 🔴 Changes requested — changes requested by bosilca — awaiting: @edgargabriel | LibNBC’s scheduled OP step calls ompi_op_reduce() on the host. When its operands are CUDA device buffers, this can pass device pointers to a host reduction kernel and crash. Detect device operands and stage supported ... |
| #14534 | config: Bump minimum Autoconf requirement to 2.69 | hppritcha | 2026-09-29 (9d) | Target: main | - | +6 / -1 | ⏳ Review required — no reviewer assigned, no reviews yet | Open MPI has required Autoconf 2.60 since 2007. Autoconf 2.69 was released in 2012 and is widely available. This update enables the adoption of modern Fortran compiler detection macros that improve |
| #14521 | BTL/UCT: fix verbosity issue | hppritcha | 2026-09-28 (10d) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +2 / -2 | ⏳ Review required — awaiting: @janjust | fixes #14086 |
| #14519 | coll/xhc: new features/developments | gkatev | 2026-09-25 (13d) | Target: main | - | +3399 / -2119 | ⏳ Review required — commented only: github-advanced-security, gkatev | New features & developments for XHC Main changes: - Run-time tuning with preliminary but quite comprehensive auto-tuning rules |
| #14506 | update-my-copyright.pl: keep file permissions, use the nearest base, and honor @protected from any directory | jsquyres | 2026-09-24 (14d) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +57 / -22 | ✅ Approved — approved by hppritcha | Three bugs in contrib/update-my-copyright.pl, the first two found while preparing #14504 and #14505. Many of us run this script from a git pre-commit hook with --check-only, which is where the second one bites. Th... |
| #14499 | Tell a bridged intercomm's peers which of them share our node | rhc54 | 2026-09-23 (15d) | Target: main | - | +154 / -26 | ⏳ Review required — no reviewer assigned, no reviews yet | MPI_Intercomm_create between two groups that have never been connected (for example two jobs spawned by one parent and bridged through it) has the two leaders swap their sides' procs with ompi_proc_pack()/`ompi_pr... |
| #14495 | Add LL/SC support for RISC-V platforms | devreal | 2026-09-23 (15d) | Severity: critical, Target: main | - | +323 / -11 | 🔴 Changes requested — changes requested by jsquyres | - Introduces explicit acquire-ll, linked load with acquire semantics - Adds implementation of ll/sc for 64bit RISC-V platforms. Note: The RISC-V "A" extension spec (Volume I, §A.6 "Eventual Success of Store-Conditiona... |
| #14493 | osc/ucx: fence between endpoint close and worker destroy at finalize | bwestheimer | 2026-09-22 (16d) | Target: main | - | +82 / -12 | ⏳ Review required — awaiting: @janjust | ompi/mca/osc/ucx/osc_ucx_component.c:component_finalize() closes the UCP endpoints and then destroys the UCP workers back to back, with nothing synchronizing the ranks in between. |
| #14492 | comm_revoke_local: Check for null before revoking c_coll | Matthew-Whitlock | 2026-09-22 (16d) | Target: main | - | +12 / -2 | ✅ Approved — approved by bosilca | There's a window during communicator creation that a failure can lead to the communicator being constructed being revoked before it has been assigned a coll component in c_coll. Open MPI then segfaults. This fix just ... |
| #14474 | Rename --enable-standard-abi to --enable-forum-abi, and use "MPI Forum ABI" consistently | jsquyres | 2026-09-21 (17d) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +160 / -146 | ⏳ Review required — commented only: hppritcha, jsquyres | Follow-on to #14467 (the ompi-forum-abi*.pc rename). The docs settled on "the Open MPI ABI" and "the MPI Forum ABI" as the names of Open MPI's two ABIs, but the configure option, the installed header directory, and ... |
| #14470 | opal/util: strip trailing separators in opal_os_dirpath_create() | tonyskjellum | 2026-09-20 (18d) | Target: main | - | +93 / -7 | ⏳ Review required — commented only: hppritcha | Fixes #14469. opal_os_dirpath_destroy() was reworked in #14257 to ignore trailing separators before it opens anything, because a trailing separator makes the kernel resolve the final component as a directory and fol... |
| #14460 | DOCS: add blurb about creative commons | hppritcha | 2026-09-17 (21d) | Target: main | - | +64 / -12 | ⏳ Review required — commented only: hppritcha, jsquyres | this is need for HPSF activity |
| #14458 | common/ofi: select a device near the accelerator by default | yinliaws | 2026-09-17 (21d) | Target: main | - | +48 / -4 | ⏳ Review required — no reviewer assigned, no reviews yet | Device selection already knows how to find the NIC closest to the accelerator a process is using: find_nearest_provider_from_accelerator() compares the PCI locality of each candidate against the accelerator's own and ... |
| #14449 | coll/ucc: pass void pointers to %p | yqtian-se | 2026-09-16 (22d) | Target: main | - | +17 / -17 | ⏳ Review required — no reviewer assigned, no reviews yet | UCC_VERBOSE ultimately passes its variadic arguments to the printf-checked opal_output_verbose() function. Seventeen persistent-collective logging calls use %p for an mca_coll_ucc_req_t * argument. The C forma... |
| #14429 | Add Bine tree collective algorithms | Dragoned | 2026-09-14 (24d) | Target: main | - | +6187 / -234 | ⏳ Review required — commented only: Dragoned, hppritcha | Add implementations of the BINE (Binomial Negabinary tree) family of collective algorithms, based on "Bine Trees: Enhancing Collective Operations by Optimizing Communication Locality" (De Sensi et al., SC'25, https://... |
| #14426 | Add --disable-ompi-abi configure option for ABI-only installations | hppritcha | 2026-09-13 (25d) | Target: main | - | +195 / -26 | ⏳ Review required — commented only: bwbarrett | This commit adds support for a new configuration mode that produces MPI ABI-only installations containing only the standard MPI ABI library (libmpi_abi.so), the standard mpi.h header, and standard wrapper compilers. W... |
| #14425 | btl/tcp: arbitrate an inbound connection from the component | bosilca | 2026-09-13 (25d) | Target: main | - | +7013 / -1197 | ⏳ Review required — no reviewer assigned, no reviews yet | Depends on #14377, and through it on #14423 and #14424. The branch is cut from topic/delayed_connections, so this PR necessarily shows all 24 commits of the stack below it — a pull request's base branch has to l... |
| #14420 | topo/base: drop the parent's topology when creating a dist graph | metereconsulting | 2026-09-11 (27d) | Target: main | - | +21 / -0 | ⏳ Review required — no reviewer assigned, no reviews yet | A distributed graph communicator created from a Cartesian or graph communicator is treated by Open MPI as the parent's topology type. Open MPI v5.0.10, MPI_Dist_graph_create_adjacent called on a 3-D periodic Cartesi... |
| #14331 | accelerator/cuda: fix memory pool and memory context handling | devreal | 2026-08-18 (1m) | Target: main, backport: v5.0.x, backport: v6.0.x | Not yet backported - pending merge to main first. | +27 / -12 | 🔴 Changes requested — changes requested by bosilca — awaiting: @Akshay-Venkatesh, @janjust | This PR addresses two issues I encountered recently: 1) When we detect a context for the memory we're probing in check_addr we would just return without setting the output device ID and after setting the memory cont... |
| #14306 | Emit MPI Standard ABI handles in MPI_T event payloads | hppritcha | 2026-08-10 (1m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +2490 / -79 | 🔴 Changes requested — awaiting: @devreal | Open MPI's MPI_T event producers publish MPI object handles (communicator, window, session, error handler, file) in their event payloads. Those values were always the internal object pointer, which is correct only for... |
| #14262 | OSC/UCX: Fix non-contiguous origin iovec traversal in MINLOC/MAXLOC accumulate | bwestheimer | 2026-08-02 (2m) | Target: main | - | +432 / -87 | 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust | The three MINLOC/MAXLOC reduce loops in osc_ucx_comm.c walked the origin buffer through the iovec list built by create_iov_list(), stepping an absolute pointer by the staging buffer's element stride and handing ra... |
| #14260 | OSC/UCX: Size accumulate staging buffers with the datatype span | bwestheimer | 2026-08-02 (2m) | Target: main | - | +43 / -29 | 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust | Three OSC/UCX accumulate paths allocated primitive staging arrays as true_extent * count, even though the data is produced and consumed at the datatype extent. For padded pair types and counts greater than one, that |
| #14256 | Fix CodeQL integer-multiplication overflow alerts | jsquyres | 2026-08-02 (2m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +91 / -65 | 🔴 Changes requested — changes requested by bosilca — approved by devreal, edgargabriel, hppritcha | This triages and fixes 34 of the 38 open cpp/integer-multiplication-cast-to-long CodeQL code scanning alerts (the remaining 4 are dismissed as bounded by construction — see below). Every site was individually triage... |
| #14254 | Fix CodeQL snprintf/scanf alerts in monitoring, tuned, and comm_method | jsquyres | 2026-08-02 (2m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +44 / -18 | ✅ Approved — approved by devreal | This fixes the three remaining small CodeQL code scanning alerts, plus two adjacent bugs found while fixing them. Four commits: common/monitoring ([alert 41](https://github.com/open-mpi/ompi/security/code-scanning... |
| #14253 | Fix CodeQL snprintf-overflow and format-specifier alerts in coll | jsquyres | 2026-08-02 (2m) | Target: main, backport: v5.0.x, backport: v6.0.x | Not yet backported - pending merge to main first. | +50 / -30 | ✅ Approved — approved by bosilca — awaiting: @devreal | This fixes four CodeQL code scanning alerts, in two commits. Commit 1 resolves alert 40 (cpp/overflowing-snprintf) in coll/han: the loop building t... |
| #14207 | osc/ucx: Honor origin datatype lower bounds in accumulates | jeffhammond | 2026-07-24 (2m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +6 / -3 | ⏳ Review required — awaiting: @devreal | Contiguous derived datatypes may use a nonzero true lower bound, including absolute hindexed displacements paired with MPI_BOTTOM. The UCX accumulate reduction paths classified these datatypes as contiguous but passed... |
| #14204 | osc/rdma: Preserve remote base for local window peers | jeffhammond | 2026-07-24 (2m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +12 / -7 | ⏳ Review required — awaiting: @devreal | MPI_Win_allocate maps a node-local shared segment into every process. When CPU and NIC atomics cannot be mixed, local accumulates are sent through the node leader's BTL endpoint and must use the leader's registered vi... |
| #13958 | mca: fix deprecated variable source warning | francdoc | 2026-06-11 (3m) | Target: main | - | +35 / -19 | ✅ Approved — approved by rhc54 | Linked to #12437. This PR fixes deprecated MCA synonym warning source reporting. Previously, if mca_base_component_show_load_errors was set in an MCA params file and the deprecated mca_component_show_load_errors n... |
| #13949 | coll/ucc: give pset-derived communicators their own OOB and context | bosilca | 2026-06-09 (3m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +510 / -145 | ⏳ Review required — awaiting: @Sergei-Lebedev, @janjust | UCC previously bootstrapped a single, process-global UCC context whose out-of-band (OOB) channel was bound to MPI_COMM_WORLD, and built every team on that context using an endpoint map into the world rank space. That ... |
| #13942 | Add pdeb modular process debugger helper | bosilca | 2026-06-05 (4m) | Target: main, backport: v6.0.x | Not yet backported - pending merge to main first. | +3402 / -0 | 🔴 Changes requested — awaiting: @jsquyres | I developed these tools over the years to compensate for the lack of parallel debuggers and to help debug OMPI internals. It now seems like the right time to contribute them back to the community, in case others find ... |
| #13903 | rocm/accelerator: add HIP Virtual Memory Management (VMM) IPC support | mshanthagit | 2026-05-21 (4m) | Target: main | - | +646 / -66 | 🔴 Changes requested — changes requested by bosilca | Implements HIP VMM-based IPC for the ROCm accelerator component and fixes misc bugs. |
| #13896 | osc/rdma: extend MPI_Win_shared_query to MPI_Win_create windows via smsc | devreal | 2026-05-14 (4m) | Target: main | - | +139 / -39 | ⏳ Review required — commented only: copilot-pull-request-reviewer, devreal, hppritcha | Follow-up to https://github.com/open-mpi/ompi/pull/13330 to support MPI_Win_shared_query on windows created using MPI_WIN_CREATE in osc/rdma. This is supported if xpmem is available via smsc. The mapping is created ... |
| #13876 | Add hook/fault_injector debugging tool | Matthew-Whitlock | 2026-05-06 (5m) | Target: main | - | +338 / -0 | 🔴 Changes requested — awaiting: @bosilca | Any interest in making this a core ompi feature? edit: I'm really asking if there's any willingness to let this be an ompi feature - I can supply the interest :) It's a helpful tool for debugging fault tolerance funct... |
| #13830 | HPE Cassini (CXI) Hardware Performance Counters (HWPC) Feature | sadlochr | 2026-04-20 (5m) | Target: main | - | +7192 / -20 | 🔴 Changes requested — approved by ryanhankins | This feature allows users to track and output various Cassini hardware performance counters over execution time. During MPI_Init(), it will determine which hardware counters to track depending on user input. During th... |
| #13667 | Properly del_proc spawned procs upon instance finalize | Dupratj | 2026-01-20 (8m) | Target: main | - | +153 / -17 | ⏳ Review required — awaiting: @hppritcha | Upon finalize the ressources allocated for a spawned proc are not all released. This Pull request is a take over of 13445 |
| #13654 | wipe the generalized request upon demand | bosilca | 2026-01-14 (8m) | Target: main | - | +19 / -12 | ⏳ Review required — no reviewer assigned, no reviews yet | If the user calls MPI_Request_free on a generalized request, make sure the request, both user side and MPI side, is properly released. However, make sure the generalized request free_fn callback function will be calle... |
| #13527 | pml/ob1: introduce a new protocol for intermediate-sized messages | hjelmn | 2025-11-21 (10m) | Target: main | - | +670 / -142 | ⏳ Review required — commented only: bosilca, hjelmn | This commit introduces the multi-eager protocol to ob1. This protocol works by fragmenting into multiple eager-sized messages and sending them in parallel to the destination. On the receiver the first fragment is matc... |
| #13526 | btl/uct: update btl default limits | hjelmn | 2025-11-21 (10m) | Target: main | - | +10 / -8 | 🔴 Changes requested — changes requested by bosilca | When a module is created for a UCT transport we take the min of the transport limits and whatever the default is in the module. This is done to support user provided limits. The issue is the limits are too low and the... |
| #13519 | ucx: meet at the pmix fence before disconnecting to avoid an infinite loop | jeking3 | 2025-11-14 (10m) | Target: main | - | +8 / -2 | 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust | While moving a job from a small number of GPU nodes to a larger number of CPU nodes, I was able to reliably reproduce #11087 in my environment. While in the debugger, I found that opal_common_ucx_mca_pmix_fence was ... |
| #13433 | interlib: fix thread safety issue | ggouaillardet | 2025-10-10 (11m) | bug, Target: main | - | +6 / -3 | ⏳ Review required — no reviewer assigned, no reviews yet | getenv() is not thread-safe so it should not be invoked within a callback subroutine. So read the environment at initialization and cache the result. Thanks Zdeněk Salvet for reporting this issue. Refs #13428 |
| #13147 | Extract the install prefix from the shared library. | bosilca | 2025-03-17 (1y6m) | Target: main | - | +225 / -22 | ⏳ Review required — commented only: bosilca, jsquyres | This is part of a multi-project effort, a similar PR will be created in OpenPMIX and OMPI. The goal of each of these changes is the same: instead of using build-time generated prefix that ignore a project rebase, take... |
| #13088 | RFC: Provide equivalence of MPICH_ASYNC_PROGRESS | hominhquan | 2025-02-11 (1y7m) | Target: main | - | +548 / -123 | 🔴 Changes requested — changes requested by jsquyres | This PR is follow-up of #13074. |
| #13039 | part-persist: implement message aggregation | AxelSchneewind | 2025-01-15 (1y8m) | Target: main | - | +369 / -43 | 🔴 Changes requested — commented only: AxelSchneewind, mdosanjh | Created a component for partitioned communication that supports message aggregation, based on the existing part-persist component. The goal is to prevent drops in effective bandwidth when using too fine-grained partit... |
| #12935 | part: multithreading deadlocks fixes and safety checks | Keluaa | 2024-11-26 (1y10m) | Target: main | - | +172 / -148 | ⏳ Review required — no reviewer assigned, no reviews yet | This PR fixes multiple deadlocks and issues encountered with partitioned communications. The first deadlocks occur when one thread is in opal_progress and others are working on the partitioned request: - `MPI_Pready... |
| #12826 | mca/base: add a new MCA variable type for include lists | hjelmn | 2024-09-24 (2y) | Target: main | - | +910 / -26 | 🔴 Changes requested — changes requested by jsquyres — awaiting: @bwbarrett | It is not uncommon in Open MPI to register a string variable to allow specifying an exclude or exclude list (every framework registers one). Given this common patten it is worthwhile to formalize an MCA include list v... |
| #12796 | Created the part direct module | mdosanjh | 2024-09-05 (2y1m) | Target: main | - | +2444 / -4 | ⏳ Review required — commented only: AxelSchneewind | Created the part direct module. This adds a module for partitioned communication that uses RMA similar to better match the original research. This has performance benefits but places additional constraints on the user... |
| #12725 | opal_thread integration with libevent | npe9 | 2024-08-01 (2y2m) | Target: main | - | +187 / -2 | ⏳ Review required — no reviewer assigned, no reviews yet | This is an initial stab at changing the Open MPI libevent integration to use opal threads instead of pthreads for alternative threading libraries. Currently, although opal threads are now replaceable via an mca, opal_... |
| #12586 | Rework the MPI connect/accept code | bosilca | 2024-05-29 (2y4m) | Target: main | - | +26 / -29 | ⏳ Review required — no reviewer assigned, no reviews yet | I'm puzzled by this code working. It does all the right things, but it was calling PML->add_procs with zero new procs, ALWAYS (the ilist was empty by the time we got to the add_procs call). But somehow it still work... |
| #12570 | Make datatype copy stream-aware | devreal | 2024-05-23 (2y4m) | Target: main | - | +87 / -34 | ⏳ Review required — awaiting: @bosilca | This is the third chunk of #12318, which makes datatype copying stream-aware. This will be needed in the coll components to order data movement with kernel invocations. Adds `ompi_datatype_copy_content_same_ddt_stream... |
| #12569 | Add CUDA/HIP implementations of reduction operators | devreal | 2024-05-23 (2y4m) | Target: main | - | +8490 / -52 | 🔴 Changes requested — awaiting: @Akshay-Venkatesh | This is the second part #12318, which provides the device-side reduction operators and adds stream semantics to ompi_op_reduce. As usual, the operators are generated from macros. Function pointers to kernel launch f... |
| #12453 | mca/coll: Add any radix k for alltoall bruck algorithm | jiaxiyan | 2024-04-05 (2y6m) | Target: main | - | +130 / -45 | ⏳ Review required — awaiting: @juntangc, @lrbison | This method extends ompi_coll_base_alltoall_intra_bruck to handle any radix k. |
| #12285 | test/datatype: do not use installed components | ggouaillardet | 2024-01-28 (2y8m) | bug, Target: main | - | +2 / -1 | ✅ Approved — approved by jsquyres | By default, the tests will use the components in the build directory but also in the installation directory if any. That can cause some bizarre issues when they are not compatible. Thanks Kook Jin Noh for the initial ... |
| #12122 | mpi-ext.h: include <mpi.h> | jsquyres | 2023-11-23 (2y10m) | Target: main | - | +3 / -0 | ✅ Approved — approved by devreal — awaiting: @bosilca | Ensure to #include <mpi.h> at the top of mpi-ext.h, just so that it includes files that it depends on. Thanks to @eschnett for raising the issue. Refs #12111 |
| #11628 | Topic/improved alltoallv | bosilca | 2023-04-27 (3y5m) | Target: main | - | +228 / -17 | ⏳ Review required — awaiting: @wzamazon | Instead of a large number of persistent requests (for both sends and receives) work on a predefined size window. This allows us to control the number of concurrent communications per process. |
| #8597 | decide CONVERTOR_CUDA flag based on first byte of buffer | markalle | 2021-03-12 (5y6m) | - | - | +70 / -18 | ⏳ Review required — awaiting: @bosilca | There's a cuda check in the converter that checks one address and caches a flag for CONVERTOR_CUDA. But it was just checking pUserBuf which isn't necessarily anywhere near the data, and doesn't even need to be a legal... |
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #14592 | [v6.0.x] coll/tuned: Change the allreduce default collective algorithm selection | app/ompi-backport-bot | 2026-10-08 (0d) | Target: v6.0.x, target:v6.0.x | Backport of merged main PR #12200 (merged 2026-04-21). v5.0.x sibling backport not opened yet. | +191 / -1 | ✅ Approved — approved by jiaxiyan | Backport of #12200 to v6.0.x. --- Include the new allreduce algorithm allgather_reduce in the default algorithm selections. The default algorithm selections were out of date and not performing well. After gathering ... |
| #14581 | [v6.0.x] Backport/pr 14092 to v6.0.x | hppritcha | 2026-10-06 (2d) | Target: v6.0.x | Backport of merged main PR #14092 (merged 2026-06-30). v5.0.x sibling backport not opened yet. | +87 / -12 | ⏳ Review required — awaiting: @devreal | no description |
| #14578 | [v6.0.x] info: report only the MPI_INFO_ENV fields that were populated | hppritcha | 2026-10-06 (2d) | NEWS, Target: v6.0.x | Backport of merged main PR #14523 (merged 2026-10-01). v5.0.x sibling backport not opened yet. | +219 / -45 | ⏳ Review required — awaiting: @edgargabriel | MPI_Info_create_env describes how the process was started, so it has to answer the same way before MPI_Init, while MPI is up, and after MPI_Finalize. On macOS it reported maxprocs 0, soft 0 and host "Redshift.local" b... |
| #14577 | [v6.0.x] comm: return an error when the nvlink domain MCA var has no value | hppritcha | 2026-10-06 (2d) | Target: v6.0.x | Backport of merged main PR #14259 (merged 2026-08-03). v6.0.x #14570 carries the other commit of #14259; no v5.0.x backport yet. | +4 / -1 | ⏳ Review required — awaiting: @jsquyres | ompi_comm_split_type_get_nvlink_domain() lumped the "get_value failed" and "no value available" cases into a single return statement that returned OMPI_SUCCESS in the latter case, even though no NVLink domain had been... |
| #14576 | [v6.0.x] Backport/pr 14424 to v6.0.x | hppritcha | 2026-10-06 (2d) | Target: v6.0.x | Backport of merged main PR #14424 (merged 2026-09-23). v5.0.x sibling backport not opened yet. | +590 / -279 | ⏳ Review required — no reviewer assigned, no reviews yet | no description |
| #14575 | [v6.0.x] pml/ob1: defer accelerator stream and event creation to first use | app/ompi-backport-bot | 2026-10-06 (2d) | Target: v6.0.x, target:v6.0.x | Backport of merged main PR #14555 (merged 2026-10-06). v5.0.x sibling backport not opened yet. | +301 / -194 | ⏳ Review required — no reviewer assigned, no reviews yet | Backport of #14555 to v6.0.x. --- Previously, mca_pml_ob1_accelerator_init() unconditionally created two accelerator streams and 2*N events (default N=400) at MPI_Init time whenever a non-null accelerator component ... |
| #14570 | [v6.0.x] comm_cid: fix leak of the PMIx_Value_string() result | hppritcha | 2026-10-05 (3d) | Target: v6.0.x | Backport of merged main PR #14259 (merged 2026-08-03). v6.0.x #14577 carries the other commit of #14259; no v5.0.x backport yet. | +5 / -1 | ⏳ Review required — awaiting: @edgargabriel | PMIx_Value_string() returns an allocated string that the caller must free. ompi_comm_get_remote_cid_from_pmix() called it inside OPAL_OUTPUT_VERBOSE() on the type-mismatch error path and leaked the result. Save the st... |
| #14569 | [v6.0.x] ompi/comm: release c_keyhash instead of destructing it | hppritcha | 2026-10-05 (3d) | Target: v6.0.x | Backport of merged main PR #13846 (merged 2026-05-04). MERGED sibling backport [v5.0.x] #13867 (merged 2026-06-15). | +8 / -4 | ⏳ Review required — awaiting: @bosilca | c_keyhash is allocated via OBJ_NEW in ompi_attr_hash_init, so cleanup must use OBJ_RELEASE — not OBJ_DESTRUCT, which runs the destructor without freeing the heap-allocated opal_hash_table_t struct itself. This leaks 7... |
| #14566 | [v6.0.x] Fix singleton tmp files cleanup | hppritcha | 2026-10-03 (5d) | Target: v6.0.x | Backport of merged main PR #13261 (merged 2026-01-12). likely duplicate: the same commit already merged to v6.0.x via #14584 (2026-10-06). | +62 / -24 | ⏳ Review required — awaiting: @janjust | Only in singleton mode, directory cleaning needs to be done by the program itself. There are some problems with these parts of the code that cause the directory to not be cleaned. This commit fixes some of these iss... |
| #14473 | [v6.0.x] Rename the MPI Forum ABI pkg-config files to ompi-forum-abi*.pc | app/ompi-backport-bot | 2026-09-21 (17d) | Target: v6.0.x, target:v6.0.x | Backport of merged main PR #14467 (merged 2026-09-21). | +89 / -72 | ⏳ Review required — no reviewer assigned, no reviews yet | Backport of #14467 to v6.0.x. --- Stacked on #14466 — please review/merge that one first. |
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #14560 | [v5.0.x] btl/tcp: let the kernel pick the source port at connect() time | app/ompi-backport-bot | 2026-10-02 (6d) | Target: v5.0.x, target:v5.0.x | Backport of merged main PR #14541 (merged 2026-10-01). open sibling backport [v6.0.x] #14557. | +32 / -1 | ⏳ Review required — no reviewer assigned, no reviews yet | Backport of #14541 to v5.0.x. --- mca_btl_tcp_endpoint_start_connect() binds every outgoing socket to one of the addresses of its btl module. The intent is to pin the source address, so the peer can pair the connect... |
| #14543 | [v5.0.x] btl/ofi: fix missing EREMOTEIO on OSX | app/ompi-backport-bot | 2026-09-30 (8d) | Target: v5.0.x, target:v5.0.x | Backport of merged main PR #14518 (merged 2026-09-29). MERGED sibling backport [v6.0.x] #14527 (merged 2026-09-29). | +2 / -0 | ⏳ Review required — no reviewer assigned, no reviews yet | Backport of #14518 to v5.0.x. --- Quick fix for compiling with OFI on Macs (and probably other unix OSes). |
| #14516 | [v5.0.x] btl/tcp: do not use an interface that has no carrier | app/ompi-backport-bot | 2026-09-24 (14d) | Target: v5.0.x, target:v5.0.x | Backport of merged main PR #14508 (merged 2026-09-24). MERGED sibling backport [v6.0.x] #14515 (merged 2026-09-28). | +67 / -1 | ⏳ Review required — no reviewer assigned, no reviews yet | Backport of #14508 to v5.0.x. --- A TCP wire-up between two ranks can complete on both sides and still leave one pair of BTL modules unconnected, after which the first message routed to the mispaired module never ar... |
Pulled out of the tables above because they are marked as a GitHub draft, or carry a WIP/DNM label, and are not realistically mergeable as-is. Same columns, same newest-to-oldest ordering within each target branch.
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #14461 | dpm: connect the two sides with one PMIx group, not publish/lookup | rhc54 | 2026-09-18 (20d) | DRAFT, Target: main | - | +992 / -195 | ⏳ Review required — no reviewer assigned, no reviews yet | > CI will fail until the PMIx and PRRTE submodules are advanced to their current masters. That update is being done separately; this is up so the changes can be reviewed. See Dependencies. `ompi_d... |
| #14263 | OSC/UCX: Defer nonblocking-accumulate finalize to top-level progress | bwestheimer | 2026-08-02 (2m) | DRAFT, Target: main | - | +628 / -163 | ⏳ Review required — no reviewer assigned, no reviews yet | With osc_ucx_enable_nonblocking_accumulate=1, accumulate finalize issues ucp_worker_fence() and then calls ucp_worker_progress() from inside a UCX send completion callback, re-entering ucp_worker_progress. Som... |
| #13943 | accelerator/cuda: skip VMM/mpool fallback for unrecognized pointers and add tunables to disable them entirely | bosilca | 2026-06-05 (4m) | DRAFT, Target: main | - | +604 / -11 | ⏳ Review required — awaiting: @devreal, @edgargabriel | Builds on a312180b43 ("defer VMM/mpool checks in check_addr fast path"), which short-circuits classification of cudaMalloc pointers reported as CU_MEMORYTYPE_DEVICE with a valid context. Two further reductions to the ... |
| #13656 | coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback | QiaoK | 2026-01-15 (8m) | DRAFT, bug, Target: main | - | +52 / -0 | ⏳ Review required — awaiting: @Sergei-Lebedev | This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t... |
| #13216 | OSC/UCX: explicitly flush default endpoints | MamziB | 2025-04-28 (1y5m) | DRAFT, Target: main | - | +26 / -0 | 🔴 Changes requested — commented only: MamziB, bosilca, devreal | This PR fixes a hang in osc/ucx observed during MPI_Finalize with Graph500 by ensuring no outstanding operations remain on default endpoints before they are destroyed. |
| #13188 | configury: enhance short float detection with clang | ggouaillardet | 2025-04-10 (1y5m) | DRAFT, Target: main | - | +13 / -5 | 🔴 Changes requested — changes requested by bwbarrett | With clang, short float support requires --rtlib=compiler-rt to be passed to the linker. In order to avoid a false positive, it is necessary have the runtime called in the test program. The test program was hence upda... |
| #12991 | coll/accelerator: allow to select functions to register | edgargabriel | 2024-12-19 (1y9m) | - | +123 / -12 | ✅ Approved — approved by devreal | This PR introduces the ability to register the component only to the select functions specified by an MCA parameter string. The idea and the code is based on the UCC component, and some of the bits might be moved late... | |
| #12745 | configury: fix IGNORE_TKR check | ggouaillardet | 2024-08-12 (2y1m) | DRAFT, Target: main | - | +1 / -1 | ✅ Approved — approved by jsquyres | As reported by Chris Parrott in #12681, the enhanced test was buggy. Pass the expected argument to make it Fortran compliant and LLVM 17 and above happy compilers |
| #9867 | Remove all ops with 3 buffers | bosilca | 2022-01-12 (4y8m) | - | +25 / -2083 | ⏳ Review required — no reviewer assigned, no reviews yet | We planned to remove support for all the MPI_Op with 3 buffer arguments, but this feature might be useful for the MPI extension described in https://github.com/mpi-forum/mpi-issues/issues/530. Full discussion in #9719. |
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #13768 | coll/ucc: fix ucc finalize leak on MPI_Finalize | janjust | 2026-03-17 (6m) | direct cherry-pick of a main commit, not tied to an open/tracked main PR | +26 / -45 | ⏳ Review required — awaiting: @bosilca, @edgargabriel, @hppritcha | commit b79004e77f intentionally skips user-defined attribute delete callbacks on MPI_COMM_WORLD during MPI_Finalize to fix an issue where PETSc attribute callbacks invoked MPI after finalize had started (PR #12072 / i... | |
| #13658 | v6.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback | QiaoK | 2026-01-16 (8m) | DRAFT, bug, Target: v6.0.x | parallel direct submission to v6.0.x (no separate main PR); twin of open PR #13657 on v5.0.x | +52 / -0 | 🔴 Changes requested — changes requested by jsquyres — awaiting: @janjust | This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t... |
| # | Title | Author | Created (age) | Labels | Merge/Backport Status | LOC (+/-) | Review Status | Description |
|---|---|---|---|---|---|---|---|---|
| #13953 | v5.0.x: docs: add MPI-4.0 and MPI-4.1 conformance gap analyses | jsquyres | 2026-06-10 (3m) | DRAFT, Target: v5.0.x | standalone v5.0.x-only docs change, bot:notacherrypick | +382 / -34 | ⏳ Review required — no reviewer assigned, no reviews yet | Similar to #13954, I had Claude analyze the v5.0.x branch vs. the MPI Forum 4.0 and 4.1 PDFs, and update our RST docs based on the actual conformance in the git branch. This PR represents the results. I have not (ye... |
| #13657 | v5.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback | QiaoK | 2026-01-16 (8m) | DRAFT, bug, |
parallel direct submission to v5.0.x (no separate main PR); twin of open PR #13658 on v6.0.x | +52 / -0 | 🔴 Changes requested — changes requested by jsquyres — approved by janjust | This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t... |
| #12253 | v5.0.x: Propagate the error up to the user. | jsquyres | 2024-01-19 (2y8m) | bug, |
Backport of merged main PR #11683 (merged 2024-01-19). no separate v6.0.x backport PR found (source PR predates current v6.0.x branch point). | +47 / -20 | ⏳ Review required — awaiting: @bosilca | Make sure to reset the generalized request to guarantee not to call the free callback a second time. (cherry picked from commit ac3647e16d2b9765fec2e60a61c92e9b8661efc4) This is the v5.0.x PR corresponding to main PR ... |