Skip to content

WeeklyTelcon_20151215

Jump to bottom Edit New page
Geoff Paulsen edited this page Dec 15, 2015 · 5 revisions

Open MPI Weekly Telcon Minutes 12/15/2015

  • Dialup Info: (Do not post to public mailing list or public wiki)

Attendees

  • Jeff Squyres
  • Edgar Gabriel
  • Geoffroy Vallee
  • Howard
  • Joshua Ladd
  • Nathan Hjelm
  • Ralph
  • Ryan Grant
  • Todd Kordenbrock

Agenda

Review 1.10

  • Milestones: https://github.com/open-mpi/ompi-release/milestones/v1.10.2
  • 1.10.2 - still 3 PRs waiting to go into 1.10.2
    • Ibarrier thrown to Nathan. Found through MPICH test suite.
    • Jim Sharp reported, Ralph cleaned it over and put into 1.10, and threw this to Jeff S.
    • Integer Overflow - Thrown to George. Ralph will ping him.
      • in coll/allreduce - From Jeff Hammonds big MPI thing
      • Recasts to size_t to do math, and then recasts down to int.
      • Nathan - should doublecheck math, since might still overflow.
      • Should evaluate these codepaths a bit better.
    • PR on master, but tagged with 1.10.2
      • Jeff S will look at today, and may then be able to PR to 1.10.2
    • Nathan has one more, unmemmap a pointer belonging to OSHEM.
      • Oneline change, and will bring it over soon.
    • Subarray 1191 on master. Jeff hasn't been following.
      • Need to fork off to George.
    • Edgar - Email about ROMIO / Luster issue
      • Issue is fixed in OMPIO Master, but not on 1.10. 1.10 OMPIO is vastly out of sync with Master.
      • QUESTION: should we update OMPIO on 1.10?
        • Some changes in the Framework stuff, but if we pull it over it will drag a lot of other items.
        • DECISION: Lets NOT update OMPIO on 1.10.x for now, encourage people to
  • After these are done will roll an RC later this week.

Review 2.0.x

  • Wiki: https://github.com/open-mpi/ompi/wiki/Releasev20

  • Blocker Issues: https://github.com/open-mpi/ompi/issues?utf8=%E2%9C%93&q=is%3Aopen+milestone%3Av2.0.0+label%3Ablocker

    • PMIx is Howards #1 blocker right now. We need to decide what we want to do.
      • Putting off supporting external PMIx in Release Canidate.
        • Distros won't pick it up if we don't support it in 2.0.0
      • What can we do with PMI-x for 2.0 RC?
        • Ralph - It's relatively clean. Ralph will pull Master 1.1.2 over to OMPI 2.0 branch later today.
    • News and shlib version stuff.
      • Howard will do News, and share with others to review.
    • Addprocs == 0 discovery.
      • Running out of resources in a different way.
      • Only happens with perpair Queues in openib.
        • Thought we'd gotten rid of those years ago, no performance advantage.
      • Not really a blocker then, the blocker would be, ensure that we got rid of non-srq mode.
        • Nathan will review old email and code.
        • Anywhere we have free_list_wait, we get into infinite loops.
    • Debugger attachment issue broken on master and 2.x branch
      • Processes don't progress until debugger progresses, which can't attach until proctable has been created.
      • Some debuggers provide a flag that they turn on when they're attached.
      • So we added code that rank0 won't progress until it gets an RML message from mpirun saying that debuger has attached to mpirun.
        • So with PMI-x we removed that RML message, which we only need for this.
      • Ralph is proposing we use PMI-x error handling to get this message.
        • Problem is that PMI-x error hanlding code is in PMI-x 1.2.0, but Master / 2.0 is currently at 1.1.2.
      • This is an issue tagged somewhere else. Might be on the Totalview side.
      • QUESTION: SHOULD we update PMI-x on master / 2.0 branch to be PMI-x 1.2.0?
      • Jeff's been using DDT, but we missed that we broke totalview.
        • Totalview still uses MPIR_Being_deubugged.
        • two different ways to tell MPI implementation that we're being debugged.
      • Ralph will try to setup a time on Doodle to further discuss.

  • Milestones: https://github.com/open-mpi/ompi-release/milestones/v2.0.0 *

RFCs

  • RFC: remove embedded libevent and hwloc

Development Tips

  • Protip - You can add "Fixes: ISSUE#" to PR, then when PR is merged, it will close the issue.
    • Yes it works across github repos.
  • Howard, When creating a downstream PR, they add a

Review Master?

MTT status:

  • Travs testing

Status Updates:

Status Update Rotation

  1. Cisco, ORNL, UTK, NVIDIA
  2. Mellanox, Sandia, Intel
  3. LANL, Houston, HLRS, IBM

Back to 2015 WeeklyTelcon-2015

Add a custom footer
Add a custom sidebar
Clone this wiki locally