WeeklyTelcon_20170307

Open MPI Weekly Telcon

rcache 3013 - rcache broken in all verison.
- big-hammer workaround for v2.0 and v2.1 Don't hook madvise.
- For 3.0 need to
- Registration cache is a bottleneck, and will always hit issues like this if we hook madvise.
- PR 3013 - might not be necessary if create new PR that does NOT hook madvise (remove a function pointer).
  - PR 3013 is an optimization, not bad, but really need to remove madvise. Nathan will Merge PR3013 now.
  - PR 3013 is not NEEDED for v2.1, so just going into master / v3.0.
- Ask Nathan to create v2.x, and v2.0.x and master PRs that remove hook of madvise.
- Does this necessitate a v2.0.3 release? Have to be doing malloc and free in threads.
  - Failure-mode threads deadlock, not silent memory corruption. Does not affect 1.10 (using pmalloc hooks).
3 issues on BSD and various flavors.
- 3 are PMIx related. Include file missing, Josh already put up PR for.
- oob 3115 dlopen failing to find files, but happening 20%. Open BSD on i386. getcwd() is missing.
- NetBSD on AMD64 - not sure how common this is.
- Artem and Paul are looking into Issue 3117 - Waiting to hear if it's easy to fix.
Taking one from Edgar: https://github.com/open-mpi/ompi/pull/3105 since we're doing another RC (for rcache fix).
Only Blockers for v2.1: Issue 3117, and unhooking madvise.

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).
DELAYED TO v3.1 - Info Keys - IBM Do an Audit with what was posted last week from MPI Forum and rebase.
- PR 2941.
- Open-MPI currently doesn't implement.
  - Concerned about Don't want to implement something if it's NOT going to be solid.
- Nathan would like to have it into v3.0, but not necessary.
- Sounds like everyone is okay with delaying this to v3.1, but want to get it into Master soon.
Updated internal hwloc DELAYED to v3.1. Still support latest hwloc via external.
Big Elephant in room is PMIx v2.0, it's not released yet, but it's being Whitelitsted, but we need to branch v3.0 soon to make June 15 release.
43 Issues against v3.0 out there. Feature / Enhancements based ones just be punted to v3.1.
Schedule for branching v3.0 after next week's meeting.

Planning on doing a v2.1.2 release in next week or so.
- Don't want to hold up v2.1, would go into v2.1.1
PMIx v2.0 - looking good for early april release

OSHMEM testing in MTT via CISCO is greatly improved. (Lookat what CISCO did if you're interested in OSHMEM)
- Gone from many thousands of false failures to a few hundred.