Conference call notes 20220817 - easybuilders/easybuild GitHub Wiki

(back to Conference calls)

Notes on the 204th EasyBuild conference call, Wednesday 17 August 2022 (8:00 UTC)

Attendees

List of attendees (8):

  • Jorg Sassmannshausen
  • Kurt Lust
  • Mikael Ohman
  • Thomas Hayward-Schneider
  • Alexander Grund
  • Jasper Grimm
  • Simon Branford
  • Caspar van Leeuwen

Agenda

  • overview of recent developments
  • Q&A

Recent developments

  • AMD toolchains contribution
  • release timeline
    • latest release: EasyBuild v4.6.0 (8 July 2022)
    • ETA next release (EasyBuild v4.6.1?): end of Aug'22
  • recent changes
    • framework
      • bug fixes
        • Autopep8 1.7.0 + Python 2.7 caused failures in CI. Fixed to autopep8 < 1.7.0 in PR 4055
      • enhancements
      • changes
        • ...
    • easyblocks
      • bug fixes
        • update PyTorch easyblock to capture regex for tests failing with signals (PR #2768)
        • Change to PythonPackage to make sure output from the test step always ends up in EB logs, even for return_output_ec = True (only used in PyTorch EasyBlock) (PR #2770)
      • enhancements
        • ...
      • updates
        • ...
      • changes
        • ...
      • new software
        • ...
    • easyconfigs
    • framework
      • reported bugs / bug fixes
        • How to support the use of ifx in intel toolchains, specifically with mpiifort (issue #4056)
        • EasyBuild seems to rebuild dependency when --wait-on-lock-limit and --wait-on-lock-interval is used, even though dependency has been build when lock is released (PR #4059)
          • Probably, the main of the 2nd process (waiting on the dep) has passed the point where the decision to skip a dep is made. Maybe we can do the wait-on-lock earlier? Order should probably be: build dependency list, check if it's there, or if there is a lock, then wait. Might not be easy to fix.
        • EasyBuild 4.6.0 not building on beegfs due to hard-link usage. Furthermore, source here seems to contain the hardlinks, but this does not.
      • enhancements
        • add support for easystack file that contains easyconfig filenames + implement parsing of configuration options (PR #4021)
        • adding support for the use of --from-pr _and other options) in EasyStack files. Changed from previous PR, as we now reconfigure at higher level (in main.py) in between items in the EasyStack file. Very much WIP (issue #4057)
      • changes
        • ...
    • easyblocks
      • bug reports/fixes
        • Some issue with detecting C++11 support when building CMake. Seems quite specific to this user though. (issue # 2769)
        • Issue with cleanup in post install step of fftw.mpi on certain distros (PR #2771)
      • enhancements
        • New (dedicated) EasyBlock for RELION, WIP (PR #2773)
        • Better detection of CMake version (PR 2772)
        • Avoid superfluous warning for NVHPC (PR #2767)
      • updates
        • ...
      • new software
        • ...
      • changes
        • ...
    • easyconfigs

Q&A

  • Should we use Java suffix? issue #16038

    • Maybe for System toolchains, yes, and for others no? We can discuss in the issue
    • Which versions should we use? Java 13-16, don't use it for anything, support is dropped upstream. Java 17 seems to work with the Eclipse version. Let's see if we can go for this in the next toolchain.
    • What about using e.g. multideps for Java, for software where the version of Java is not critical? (e.g. in PR #16034)
      • Might be good if we just add sanity check, then people can easily duplicate the EC and change Java version

  • yaff has problems with HDF5 1.12, Jorg contacted the devs, but hasn't heard back from them. (PR #15854)

    • There's a compatibility option that might help.
    • It might also be that the issue is actually in h5py?

  • netCDF test suite, there are tests that are failing on Generoso with Intel, other user has seen it fail with foss. It's running some parallel I/O tests.

    • What filesystem does Generoso mount?
    • Maybe we should comment out the parallel tests, which are actually more like benchmarks. Michael wil make a PR for this. ...