Conference call notes 20200708

(back to Conference calls)

Notes on the 151st EasyBuild conference call, Wednesday July 8th 2020 (08:00 UTC - 10:00 CEST)

Attendees

Alphabetical list of attendees (7):

• Kenneth Hoste (HPC-UGent, Belgium)
• Jörg Saßmannshausen (NIHR Biomedical Research Centre, UK)
• Simon Branford (Univ. of Birmingham, UK)
• Alexandre Strube (JSC, Germany)
• Adam Huffman (Big Data Institute, Oxford, UK)
• Mikael Öhman (Chalmers University of Technology, Sweden)
• Leo Rojas (University of Sussex, UK)

Agenda

• update on EasyBuild v4.2.2 release
• 2020b update of common toolchains
• Q&A

Update on EasyBuild v4.2.2 release

• EasyBuild v4.2.2 will be released today
• project to track issues/PRs that should be tackled for this release: https://github.com/orgs/easybuilders/projects/6
• recent changes
  • framework
    • always take into account builddependencies when generating template values, also when we're not iterating over builddependencies (PR #3346)
    • fixes for --try-update-deps (still experimental) (PR #3350, PR #3353)
    • fixed regression in check for available modules, when using ancient modules tool version (PR #3379)
    • fix for use of template values in exts_list (PR #3377)
    • retain order of paths in when generating prepend_path statements for module file (don't sort them alphabetically) (PR #3367)
    • also put easyblocks used by extensions in reprod directory (PR #3375)
  • easyblocks
    • see https://github.com/easybuilders/easybuild-easyblocks/pulls?q=is%3Apr+sort%3Aupdated-desc+is%3Aclosed+milestone%3A4.2.2
  • easyconfigs
    • stop relying on Mercurial repositories in Bitbucket (see issue #10868)
    • cleanup of 2020a branch: not finished yet, but made some progress
      • see issue #10753
    • see https://github.com/easybuilders/easybuild-easyconfigs/pulls?q=is%3Apr+sort%3Aupdated-desc+is%3Aclosed+milestone%3A4.2.2

2020b update of common toolchains

• we should start looking into this...
• GCC 10.1 (wait for 10.2)
• OpenMPI 4.1.0 (RC)
• no updates for Intel yet?

Q&A

• Jörg: ICE with GCC 9.3.0 when compiling Salmon (see https://github.com/easybuilders/easybuild-easyconfigs/pull/10631)
  • also hitting this with manual make build
  • doesn't seem to be a memory issue
  • check if others can reproduce it
  • issue seems to be with LTO, try to disable that?
    • requires patching
  • could also try lower optimization level (-O1)
• Adam: we should discuss creating a CUDAcore + CUDA (similar to GCCcore + GCC)
  • see https://github.com/easybuilders/easybuild-easyconfigs/pull/10935
  • main motivation is avoiding multiple installations of CUDA like we have now
• Mikael: "dropping" last version part of CUDA version, good idea?
  • see https://github.com/easybuilders/easybuild-easyconfigs/pull/10935 (starting point for fosscuda/2020a)
  • do we really want to do in-place updates of CUDA?
  • CUDAcore 11.0 wrapper should be at Core level
    • may get us in trouble later (fosscuda/2020b)
  • do we trust NVIDIA to follow semantic versioning so in-place updates are safe?
  • Jörg: providing this flexibility clashes with reproducibility...
  • could still swap back to previous CUDA 11.0.x versions if needed
• Alexandre: similar discussion for NCCL version in TensorFlow
  • see https://github.com/easybuilders/easybuild-easyblocks/pull/2085