- Update to version 4.1.1
- Fix a number of datatype issues, including an issue with
improper handling of partial datatypes that could lead to
an unexpected application failure.
- Change UCX PML to not warn about MPI_Request leaks during
MPI_FINALIZE by default. The old behavior can be restored with
the mca_pml_ucx_request_leak_check MCA parameter.
- Reverted temporary solution that worked around launch issues in
SLURM v20.11.{0,1,2}. SchedMD encourages users to avoid these
versions and to upgrade to v20.11.3 or newer.
- Updated PMIx to v3.2.2.
- Disabled gcc built-in atomics by default on aarch64 platforms.
- Disabled UCX PML when UCX v1.8.0 is detected. UCX version 1.8.0 has a bug that
may cause data corruption when its TCP transport is used in conjunction with
the shared memory transport. UCX versions prior to v1.8.0 are not affected by
this issue. Thanks to @ksiazekm for reporting the issue.
- Fixed detection of available UCX transports/devices to better inform PML
prioritization.
- Fixed SLURM support to mark ORTE daemons as non-MPI tasks.
- Improved AVX detection to more accurately detect supported
platforms. Also improved the generated AVX code, and switched to
using word-based MCA params for the op/avx component (vs. numeric
big flags).
- Improved OFI compatibility support and fixed memory leaks in error
handling paths.
- Improved HAN collectives with support for Barrier and Scatter. Thanks
to @EmmanuelBRELLE for these changes and the relevant bug fixes.
- Fixed MPI debugger support (i.e., the MPIR_Breakpoint() symbol).
Thanks to @louisespellacy-arm for reporting the issue.
- Fixed ORTE bug that prevented debuggers from reading MPIR_Proctable.