Meeting 2022 03 17 - openpmix/openpmix GitHub Wiki
03/17/2022 OpenPMIx-devel call notes
Attendees
- Samuel Gutierrez (LANL)
- Matthew Baker (ORNL)
- Ralph Castain (Nanook)
- Thomas Naughton (ORNL)
- Michael Karo (Altair)
- Howard Pritchard (LANL)
Notes
-
Fixes in pmix/prrte
-
PMIx_Get
bug found/fixed that hit rare case where can get false not-found, only impacted the direct modex (dmodex). New dmodex test in place to test. Fixed on master. Partly in pmix and partly in prte -
Some OMPI schizo fixes to ensure the command line behaves as needed
-
Configury cleanups by Brian for pmix and prte, another coming to support static builds for pmix, also will impact prte
-
Need to decide what gets pulled to other release branches
-
Gist of branches
-
PMIx v4.2.x -- incorporates the ABI related changes,
- macros moved either to full functions in API or add few e.g., add argv backing functions as static inline functions in header
-
PRRTE v2.1 -- main change here was the command-line processing
- interest in having isolation between the different users of prte, e.g., ompi diff args than others.
-
master (PMIx and PRRTE)
- work to reduce code duplication (e.g., bugs must be fixed multiple times if found in MCA)
- prte becoming more dependent on PMIx, pushing out of PRTE into PMix to avoid having that code duplication
- This causes problems for porting fixes to prte to older branches b/c the newer PRTE does not include those code refactoring to leverage more of PMIx (to avoid prte dup code).
- if add these refactoring into branches, simplifies backport of maintenance code, but increases the amount of changes for the release
-
-
Q: Backport of master to v2.1.x and v4.2.x. Would like to avoid bringing those changes to v2.0.x/v4.1.x
-
Open MPI is currently using prte-v2.0.x / pmix-v4.1.x, but will need to shift to prte-v2.1.x / pmix-v4.2.x
-
SLURM breakage w/ PRRTE, not exactly sure when the break occurred. Related to slurm env vars.
-
Q: Have chance for attribute for pmix group construct?
- Ralph - mostly done, just need little more work to finish off
- Howard - have code for test/kick-tires when ready
- Note, OMPI issue related to this that could replace much code with single call to pmix_group w/ CID.
-
Sam working on improving the shared memory support (adding into master)
-
Added support to configury to avoid static build when using python bindings. Due to problems with cython not support a static build.