Telcon: 2016 04 21

Participants

• Todd Gamblin (LLNL)
• Jim Amundson (Fermi)
• Jim Galarowicz (Krell)
• Patrick Gartung (Fermi)
• Peter Scheibel (LLNL)
• Robert French (ORNL)
• Matt Belhorn (ORNL)
• Ben Boeckel (Kitware)
• Elizabeth Fischer (NASA)

News & Updates

1. Todd is back but recovering from a cold/flu/fever, will be merging lots of PRs very soon. Thanks for all the activity!
2. TOSS RPM generation update (Peter Scheibel)

• Peter working on "redirection"
  • Part of creating RPM is doing the build in a chroot and a special build directory
  • Not the same as the final install prefix (DESTDIR)
  • Mostly delegating to CMake, auto tools for DESTDIR
  • Also offering autodestdir for Makefile builds, special cases for distutils and boost build (bjam)
  • Peter is redirecting Spack commands like install and others to point to the staging area and not the install prefix
• Rough support for a yaml file describing custom RPM naming schemes.
  • allow you to specify how Spack should name multi-versioned RPMs
  • Manages trade off between combinatorial naming and in-place upgrading
  • can specify version granularity per package, per package's dependencies.

1. package.py hashing (Peter Scheibel)
2. Kitware update (Ben Boeckel)
3. News from HEP folks (Jim Amundson)

• Meeting of HEP software foundation in May
• Session on Spack led by Patrick Gartung
• Jim working on a build environment that uses Spack to get external packages
  • half fleshed out but would like to present this at a meeting
  • Elizabeth implemented something like this in PR #543
  • LLNL's code teams would like something similar to this too

1. Cray & newarch status
2. ORNL has a stopgap Cray compiler PR #803
3. LLNL merge plan: * newarch is pretty much done PR #561
   • newarch provides Cray support, module-detected compilers, better architecture specifiers
   • uses old syntax with a - to specify os and architecture:

     spack install foo =cnl-x86_64
     

* NERSC wants to switch over when new architecture *syntax* is in
  * doesn't want to train consultants twice
  * New syntax is in `cflags` PR [PR #360](/LLNL/spack/pull/360)
  * allows key/value attributes to be provided on specs:
    ```
    spack install foo os=cnl target=x86_64 cflags='-O3 -g3'
    ```
  * also allows referring to specs by hash
    ```
    spack uninstall /abc123
    ```
* **The plan right now:**
  1. merge cflags
  1. release v0.9
  1. merge newarch
  1. release v1.0

1. What should we do about ORNL's PR? * Can we merge it first to get them started? * Should we just let ORNL use this then migrate them to new arch?
   • Robert French: if new arch is getting merged in a month then go with newarch

PR Review requests

• From @citibeth:

  1. PR #543 ("spack spconfig"): I'd like to determine any final changes needed for this, and then get this merged.
  2. PR #663 (Documentation Improvements): OK docs are not sexy. But I've written a bunch of stuff, much of it based on past discussions online.
  3. PR #721 (Load Recursive Modules): I think this is good for what it does, and should be merged. I know it was not well loved, but I think it deserves a second chance to explain why it should be merged. It could, for example, obviate the need to build spack profiles.
  • Patrick Gartung has used this to set up environments on OS X
  • Lmod and tcl modules automatically load things recursively
  • relies on module purge (or logging in again)
    • module purge doesn't work on cray (:()

• @alalazo

  1. #744 : customization of modules from config files
  • ORNL sounds very happy about these additions.
    • Robert French to look at this and provide comments
  • @citibeth wants more docs on this.
  • @alalazo: can you add docs?
  1. #775 : fixes a bug found while installing things for multiple architectures
  2. #375 : flake8 check of modified sources, code coverage on test results (using coveralls)

Discussion

• RPATH for Python extensions (currently there is none; see PR #722 for the problem, not a solution)
• How to properly handle LD_LIBRARY_PATH when Spack sees it's been set by the user.