Conference call notes 20190515

(back to Conference calls)

Notes on the 124th EasyBuild conference call, Wednesday May 15th 2019 (17:00 - 18:00 CEST)

Attendees

Alphabetical list of attendees (7):

• Damian Alvarez (Jülich Supercomputing Centre)
• Markus Geimer (Jülich Supercomputing Centre)
• Fotis Georgatos (SDSC, Switzerland)
• Kenneth Hoste (HPC-UGent)
• Adam Huffman (Big Data Institute, Oxford)
• Bart Oldeman (ComputeCanada)
• Davide Vanzo (Vanderbilt University

Agenda

• update on EasyBuild 3.9.1
• update on road to EasyBuild 4.0
• dealing with the OpenBLAS problems in foss/2018b on Intel Skylake (not discussed)
• Q&A

Outlook to EasyBuild v3.9.1

• ETA: early next week

  • all PRs should be merged end of this week

• framework:

  • https://github.com/easybuilders/easybuild-framework/milestone/63
  • a couple of minor additional bug fixes
  • (general) support for fix_python_shebang_for and fix_perl_shenang_for to make scripts use #!/usr/bin/env python (mostly relevant for multi-Python installations done via multi_deps)
  • open issues:
    • clarify use of multi_deps in generated module file: https://github.com/easybuilders/easybuild-framework/issues/2860 => AP Åke

• easyblocks

  • https://github.com/easybuilders/easybuild-easyblocks/milestone/54 (to be cleaned up)
  • ELPA easyblock PR would be nice to get in
  • fix embedding of zlib in binutils https://github.com/easybuilders/easybuild-easyblocks/issues/1350
    • has been reporting for a while now, and several people have been bit by it
    • Markus is using a custom binutils easyblock to work around the problem
    • the planned changes should be safe to make w.r.t. backward compatibility
      • libbfd.so will have a runtime requirement in libz.so
    • fixing all binutils easyconfigs shouldn't be too much trouble
  • Damian: getting https://github.com/easybuilders/easybuild-easyblocks/pull/1727 which fixes prefix in impi wrappers included would be nice

• easyconfigs

  • https://github.com/easybuilders/easybuild-easyconfigs/milestone/57 (to be cleaned up)
  • Markus' PR for Cube/Score-P & co

• easyconfigs merge sprint (Wed May 8th 2019)

  • 110 easyconfig PRs dealt with (64 merged + 46 closed)

    • see also https://github.com/easybuilders/easybuild-easyconfigs/labels/merge%20sprint
  • • 4 easyblock PRs

  • merge sprint worked out really well, focused effort clearly makes a difference

  • another sprint in June'19?

    • @boegel will send a doodle to plan another sprint

  • Davide: would be nice to be able to tell if someone else is already working on a PR

    • self-assign? but will people pay attention to it
    • Bart: may be more important for framework/easyblocks, most time dealing with easyconfig PRs is spent waiting on test reports

  • Davide: could make sense to plan sprints more regularly

    • could be triggered by easyconfig PR count getting too high?

The road to EasyBuild v4.0

• ETA for EasyBuild 4.0 is (still) by summer 2019...

• rename of dummy to system https://github.com/easybuilders/easybuild-framework/pull/2877 MERGED

  • a bit more work is done on top to add support for "eb --fix-deprecated"
  • Fotis: should be renamed to --fix-deprecated-easyconfigs (but that's a bit long?)

• Bart: can we distinguish between toolchain for binary installations vs build done with system compilers?

  • toolchain = None vs toolchain = SYSTEM?
    • or more explicit like toolchain = BINDIST
    • implies making toolchain optional?
  • why does using dummy/system toolchain even work when EB f*cks up build env?
  • Damian: can we clean up the use of True in dependency specs, and use SYSTEM instead
    • example: ('Java', '11', '', SYSTEM) or ('Java', '11', SYSTEM) or ('Java', '11', '', toolchain=SYSTEM)

• deprecation of runtest in favor of testopts

  • cfr. https://github.com/easybuilders/easybuild-framework/issues/2868

• TODO: OpenMPI easyblock

Q&A

• Intel MPI 2019 update 3 that occurred today at JSC

  • libfabric (pre-compiled with impi) wasn't working properly (slow, always using TCP rather than IB)
    • not an issue on Omnipath systems...
  • fixed by untarring included libfabric and building that on own system
  • Bart: other option is to use libfabric release
    • Damian: libfabric shipped with impi has additional patches
  • 'tricky' bit: configuration of libfabric
    • MLX (Mellanox) provider configuration failed (even if UCX is not installed)
      • Bart: MLX provider is not that good (and low priority anyway)
    • if both IB & Omnipath libraries are available on system, build works but runtime fails because Omnipath bits are missing in system
  • should be integrated in impi easyblock, ideally
  • but not trivial to get configuration right based on what's available on the system
    • introduce libfabric_configopts for impi easyblock so you can push in custom configure options if needed
    • separate parameter to control rebuilding of shipped libfabric (rebuild_libfabric=True by default)
  • Damian: please open issue in easyblocks repo on this with more details
  • are the scaling issues fixed?
    • Damian: not sure, will check
    • previous issues with collectives got fixed in impi 2019.2, much better now

• Fotis: binaries = [...] to list names of expected binaries

  • also run ldd on them and capture the output in the EB log
  • could also be done for (dynamic) libraries
  • default logic could be to use (lowercase?) software name in binaries (which implies checking for bin/<name>)?
    • current default sanity check is to require non-empty bin & lib directories
  • Davide: checking ldd output could help with catching unintended linking to system libraries

• Markus: using Clang as a library doesn't really work right now, because it's registered as a compiler

  • KH: related to recent discussions with Alan & Mikael?