WeeklyTelcon_20170314

Open MPI Weekly Telcon

Thought we are ready today, BUT ...
Nathan found a blocker Bandwidth regression for the OB1 RDMA (openib) and ugenie BTLs:
- PR3164
- OB1 and Ugenie caches leave-pinned, which affects add/procs (somehow).
- KNEM uses registration, so it would be affected.
- Master and v2.1
- Why is MTT not seeing this? Because we don't test performance.
At a minimum another MTT run, and another RC.
- Earliest could release is next Tuesday.
Do we really need another RC?
- Concerned about issues creeping into other components.
- It would affect "out of the box" experience.

A couple of PRs waiting. We probably want to do a v2.0.x release, because it has the madvise bug (serious)
Enough of a tipping point to begin a v2.0.3.
Some PRs waiting for review.
If SGE fix gets into v2.0.3 that'd be great. Not going to push for it on v2.1.0

PMIx - reason we're doing an accelerated v3.0
Whitelist Issue 3107
UCX has it's own Multithreading API that needs to be enabled. UCX is thread safe. Inside UCX PML
- allocator will be inside of OSHMEM.
- Sounds reasonable (component level stuff).
- Allocator is Merged into Master.
- Multithreading is going through Mellanox internal review. So maybe end of this week.
Still think we should branch TODAY to get it over with.
People can add it to their MTT tests.
What else is on Chopping block for v3.0.0?
- Remove Yoda.
- MxM MTL can go, but Yalla must stay.
- For OSes, we'll say whatever one is running on Nathan's little MTT box.
- 32bit Linux is a challenge.
- Should remove ARM? Solaris? - Howard and Brian will reach out.
- Compilers is a mix.
- which nVidia GPUs?
- Should we remove SM/BTLs?
Brian it's not in Master right now, and not a bug affecting customers, we should not do a bunch of code work to remove stuff for v3.0 due to aggressive schedule.
- Others liked this concept.

Should be able to Torque, SLURM, SGE.
IBM committed to testing Platform LSF.
IBM will check if they care about load-leveler. Think it's just a legacy product.
- If IBM doesn't care, Recommend removing this one because it's clunky interface.
Eventually it'd be nicer to have CI for these things, but for v3.0.0, Either CI or MTT might be enough.
- Testing matrix of schedulers is more nightly MTT.

We should begin thinking about scheduling our next face to face.
- Geoff will put out doodle for June and July and begin to nail down a schedule.