Skip to content

WeeklyTelcon_20200428

Jump to bottom
Geoffrey Paulsen edited this page Apr 28, 2020 · 1 revision

Open MPI Weekly Telecon ---

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

Attendees (on Web-ex)

  • Akshay Venkatesh (NVIDIA)
  • Austen Lauria (IBM)
  • Barrett, Brian (AWS)
  • Brendan Cunningham (Intel)
  • Edgar
  • Geoffrey Paulsen (IBM)
  • Harumi Kuno (HPE)
  • Howard Pritchard (LANL)
  • Jeff Squyres (Cisco)
  • Joseph Schuchart
  • Josh Hursey (IBM)
  • Joshua Ladd (Mellanox)
  • Matthew Dosanjh (Sandia)
  • Michael Heinz (Intel)
  • Noah Evans (Sandia)
  • Ralph Castain (Intel)
  • Thomas Naughton (ORNL)
  • Todd Kordenbrock (Sandia)
  • William Zhang (AWS)

not there today (I keep this for easy cut-n-paste for future notes)

  • Artem Polyakov (Mellanox)
  • Brandon Yates (Intel)
  • Charles Shereda (LLNL)
  • David Bernhold (ORNL)
  • Edgar Gabriel (UH)
  • Erik Zeiske
  • Geoffroy Vallee (ARM)
  • George Bosilca (UTK)
  • Mark Allen (IBM)
  • Matias Cabral (Intel)
  • Nathan Hjelm (Google)
  • Scott Breyer (Sandia?)
  • Shintaro iwasaki
  • Xin Zhao (Mellanox)
  • mohan (AWS)

New

sign-off checker - when edit code on github, they put "co-authored by" instead of "signed-off by"

  • Be nice to update automation to accept this.
  • Would need to update legal text saying that if users do this they're agreeing to the terms of the community.

Jeff presented Feature --net (and --tune)

  • use a new openmpi-params.ini instead of open-mpi-mca-params.conf
    • symlink for forward compatibility.
  • This is similar for --tune <file> to provide more command line and env parameters from file.
  • Two parts to "visible" rule:
    1. command line values takes precidence over file values.
    2. If there's a conflict between things you don't see, that's an error.
  • Already slated for v5.0.x, slightly different than v4.0.x
  • --net says take a single line from a parameter.
  • --tune says take an entire parameter file.

Markdown vs Nroff

  • Writing manpages in nroff is painful
  • Jeff wrote MPI_T.5.md man page in Markdown
    • Make converts markdown files to nroff via new tool pandoc
    • Don't want to require users to install pandoc, so will convert .md files to nroff in make dist
    • Configure will error if you don't have pandoc v1.19 in path
      • will check if we can lower requirements to v1.12.3 (Comes with CentOS 7.7)
      • Initial testing looks good, verifying now.
  • Native nroff and markdown can co-exist, so don't need to do them all at once.
  • Can we suppress generation of manpages if there is no pandoc?
    • No. Don't want to support dist without "full" contents.
    • --without-manpages - maybe could make this work.
  • Jeff will send something to packagers downstream
  • Worst case, we could pull this for v5.0

Old

  • MTT -
    • If you change your MTT to startup PRRTE at begining of session, and just use prun.
    • Can see times cut in half or more.
    • This is good, but also need to test mpirun wrapper.
    • Cisco is converting half of MPI installs to use prrte/prun

OMPI submodules

  • OMPI master submodule pointers setup to track PMIx and PRRTE master.

Release Branches

Blockers All Open Blockers

Review v4.0.x Milestones v4.0.4

  • v4.0.4 Break
  • v4.0.4rc1 - This Monday
  • 7616 - ABI break introduced in OMPI v4.0.3 for some f08 symbols.

v5.0.0

  • Schedule:
    • Feature Freeze: April 30
      • NEW slipping 2 weeks (now: May 14)
      • Please Post a PR ASAP as place holder
    • Release: End of June
  • PMIx v4.0.0 - on track
    • Schedule: Still a number of issues, but probably not blocking
  • PRRTE v2.0 -
    • Went through issues to discuss remaining issues.
    • MCA usage is very different in PRRTE than in ORTE.
      • ORTE was a "one-shot" launcher, but PRRTE is persistant.
      • When launching PRRTE you can set "defaults" for the deamon
      • individual pruns override these defaults via command line parameters not mca parameters.
    • Now have two MCA users in the job. OPAL / PRRTE - if setting something in the wrong one, then it gets ignored and is confusing.
    • There will be a lot of mca param files, won't do what people expect them to do.
      • Might want to Open some issues on OMPI side to do some docs.
    • report bindings doesn't make sense to set this as a "default setting" in PRRTE, so is always a per-job basis.
    • RC1 Blockers things to get done before RC1 (Maybe 2-3 weeks?)
      • Need to get User-facing stuff done to reduce use confusion.
      • Binding reporting should be done (confusing) 523 - Needs thinking/careful work.
      • A bunch of knarly issues in here:
        • Call tomorrow?
        • Socket -> Package name change - Should we do now or later?
          • Already a lot of change, but hwloc has already moved onto the new name.
        • Also what to do with NUMA? - Doesn't even make sense anymore on some archetecture.
      • Depends on what versions of hwloc we support. Will be tricky (or more expensive) to support hwloc both 1 and 2.
      • Is there a list of distros and hwloc versions? Brian will put together list.
  • Discussing Features on google sheets document
  • Today with libevent, we default to prefering libevent if it's version is newer (we redistribute 2.0.22)
    • Still a bunch of distros that ship a libevent not newer than 2.0.22, but works.
  • For v5.0 we're continuing down the path of NOT encouraging users to use the internal libs.
    • So probably should just use external if it's found, as long as it's newer than 2.0.21 (trusted version)
  • Issues not tracked on spreadsheet.
    • libopal isn't slurped into Open-MPI correctly (related to 7560)
      • Jeff and Brian will meet Friday

master

  • Heriarchacal collectives

    • If someone wants to do, PMIx has much of this information already.
    • Not too hard to do, and they're much faster. Will be in next version of competitor MPI
    • Probably not for v5.0

  • SLURM PMIx plugin has been locked on PMIx v2 for some time.

    • There are some NEW PMIx calls that SHOULD be added to bring it up.
      • Ralph has started a PR, but needs help.
    • So for now, there's some optional info that won't be passed correctly.
      • No OMPI_INFO for now.
      • Ralph gets pinged occasionally.
    • Not sure priority of this.

  • MTT on master is looking pretty good.

Face to face

  • Defered.

Infrastrastructure

  • scale-testing, PRs have to opt-into it.

Review Master Master Pull Requests

CI status

Depdendancies

PMIx Update

  • CI testing only tests build and did it run, but doesn't test HOW it ran.
    • Environment setup can be a bit different.
    • For example no-permissions in /tmp. Might pass on one machine, and fail on another without /tmp permissions.

ORTE/PRRTE

MTT

Back to 2019 WeeklyTelcon-2019

Clone this wiki locally