Skip to content

Telcon: 2016 06 02

Jump to bottom Edit New page
Todd Gamblin edited this page Jun 10, 2016 · 11 revisions

Participants

  1. Todd Gamblin (LLNL)
  2. Benedikt Hegner (CERN)
  3. Ben Boeckel (Kitware)
  4. Dan Udwary (NERSC)
  5. Jim Galarowicz (Krell)
  6. Joseph Ciurej (LLNL)
  7. Mario Melara (NERSC)
  8. Matt Legendre (LLNL)
  9. Mike Collette (LLNL)
  10. Patrick Gartung (Fermi)
  11. Steve Leak (NERSC)
  12. Veronica Vergara (ORNL)
  13. Yang (??)
  14. BT (??)

News

  1. Todd back from travel this week
  • Catching up with PRs
  1. Cray support nearly done github/NERSC/features/newarch
  • Things remaining to do:
    1. A few remaining tweaks to the spec.yaml formatting. This changes hashes but ensures maintainability.
    • more lazy binding on platform/os/etc. (too much currently being encoded in spec)

    1. New compilers.yaml format:

      compilers:
- compiler:
    spec: gcc@4.2.1
    alias: 'gcc421'
    operating_system: elcappitan
    paths:
      cc: /usr/bin/gcc
      cxx: /usr/bin/g++
      f77: null
      fc: null
    modules: []

    2. Allows compiler descriptor to be self-contained, easily copied/reused.

    3. A few small design tweaks

  1. Build dependencies (#378) next large PR on the list after new architecture support.
  • Sorry for delay on this one -- we needed to get Cray support out the door for NERSC and ORNL
  1. NERSC Spackathon!
  • Todd Visited NERSC Tuesday, spent 2 hrs giving an overview of spack to consultant team.
    • Nice to meet the NERSC consultants (10+ people)
    • very envious of the view from Bldg 59.
  • Mario Melara and others at NERSC helping to bootstrap the Cori phase II packaging effort
  • NERSC has lots of good suggestions, setting up channels for support.
  • Maybe get some ORNL folks involved?
  1. #676 (concretization slowness) partially addressed in #1015
  • Eliminated a lot of importing and O(num package.py files) from Spack.
  • See the PR description for details; still needs testing.
  • On my mac, this brings concretization time for dealii down to ~5-6 sec, 3 sec without openssl (see next bullet)
  1. Current OpenSSL version check pings the OpenSSL site; this is problematic for a few reasons.
  • Change openssl's url_for_version -> _url_for_version for a quick fix.
  • Slows down any concretization that has openssl (even things like spack spec that do not install)
  • Makes for kludgy code in url_for_version
  • See discussion here for proposed solutions.
  1. Compiler LD_LIBRARY_PATH issues
  • Partially addressed by new compilers.yaml format in Cray support branch
    • Allows modules
    • Will add env options for custom LD_LIBRARY_PATH too (Per @hegner's #943)
  1. NERSC wants #908 too (customize install location)
  • also wants to customize install tree layout
  • views would be better
  1. Intel compiler installation (Greg Lee)
  2. RPM building (Matt, Peter)
  3. Archive caching (Peter)
  4. Kitware update (Ben, DJ)
  • build deps rebased and squashed for develop
  • DJ still working on Qt and UVCDAT
  1. Update from HEP team
  • binary packaging (Gartung, Hegner)
  • views (Viren)
  • Sorry for delays on PRs -- hope to get them in faster.
  1. Others?
  • @citibeth?

Discussion

  1. Mario Melara: update on NERSC's Spack workflow

  2. NERSC consultants divvying up packages to maintain (10-ish people) 1. Work on packages in home directory 1. Install in /usr/common once working for a consultant. 1. First installing things with gcc & intel on haswell and KNL

  3. Potential issues/questions from consultants: 1. static vs dynamic linking on Cray

    • Todd suggests handling this with shared variant 1. auto-generation of module files
    • every variant currently produces a module file
    • LLNL in a similar situation with RPM-based system
      • spack can build more than we want to "publish" to users
      • took steps to decide whether a package is actually seen by users or "just installed"
      • "published" packages have module files, others are just installed but no module generated. 1. changing the install prefix from $spack/opt/spack to something else
    • In Benedikt's PR -- to be merged real soon. 1. consultant's spack build environment is breaking due to modules in .bashrc
    • Cray doesn't have a good way to get a vanilla environment
    • Need a better way to get a basic environment without using module purge.
      • or tell users to unload all modules that depend on the PrgEnv
    • Look into also omitting the user's .bashrc
      • could we just source the site-specific .bashrc

  4. veronica trying to use Spack to build on a Power machine * Sqlite and can't understand that the machine is POWER * look in package.py files and grep for filter_file for examples of how to customize build files.

  5. Joe Ciurej (LLNL): PR update 1. Better support for setup.py (ignoring user configuration) is ready to merge (PR #950) 1. Fetcher configuration

    • curl on LC systems sometimes fail because curl does not support the progress bar
    • probably need a notion of fetch dependencies in Spack
Add a custom footer
Clone this wiki locally