WeeklyTelcon_20220111

Open MPI Weekly Telecon ---

Need ROMIO update.
- Open an issue to track this.
What's the status of MPI Sessions?
- It's ready for review, but it's a very big PR.
- Nathan pulled attributes out
- v5.0 release managers should look at it closely. Two main things in it:
  1. Reorganization for MPI_Init/Finalize to be able to be called multiple times
  - Attributes is used for this
  1. Extended CID thing. Read this section and this link for cluster 19 paper.
  - Allow us to use PMIx process set stuff more efficently.
    - Request a unique 64bit PMIx CID as part of a PMIx group join.
      - Expensive to get this from PMIx for large jobs.
  - UCX isn't able to do Sessions (because not using extended CID thing now)
    - For these PMLs just need to look at Init/Finalize work.
  - If you get a Comm outside of MPI_Comm_from_Group, goes back to
- Nice to have in v5.0, but since we won't have all other
- Howard is fixing one conflict on PR.
- Deadline of Feb 1st Review - Send email to devel
Thinking about an RC before and after Sessions.
- Well as far as tracking, we have nightly tarballs, and it'll be clear in git
Docs rework
- We made a lot of progress on revamping the docs with restructured text.
- Might actually be able to get this done by v5.0.x
- Dont go review yet, but lots of good progress.

Fortran
- Howard put a comment in PR #9837, We'll need a different type that uses assumed rank instead of assumed ____
  - based on email comment on standard.
  - Lists the types of dummy args when the actual arguments are async, and assumed size is not one of them.
    - Assumed size is Type(*) - so can't use asyncronous.
  - Anything that takes a buffer calls it Type(*), Dimension(..) and Dimension is the rank.
  - If it's Type(*), Dimension(*) then can't do asyncronous.
  - For the moment this only affects master. How can we fix master so it can be brought to v5.0.x
- Jeff will fix on a new PR, but doesn't want it on #9837.
- Tackle Omnibus one that's still in flight. Unclear if this would make it to v5.0.x
Some other MPI-4.0 standard things that are small. If we want to get these into v5.0, we need reviews.
- Howard pinged

Cisco test build failures still pending fixes.
- Cisco re-enabled MTT but had a typeo
- Did we change minimum hwloc allowable by master?
  - Yes, but can't remember what.
  - Found a bug, will file an issue.
IBM has an OMPI build failure with XL compiler on ppc64le.
- IBM's looking to workaround with Open MPI code change.
- Looking at this morning.
  - Looks like XL didn't like combo of -O2 + -g
- CFLAGS is for -g and doesn't like changing default when --debug is used.
- Shouldn't undo autoconf-isms.
Test related - Late last year, we found a number of collectives that had issues with large payloads, and push against MPI-3's large buffer.
- Will be pushing to the public repo.