Conference call notes 20180606

(back to Conference calls)

Notes on the 103rd EasyBuild conference call, Wednesday June 6th 2018 (5pm - 6pm CEST)

Attendees

Alphabetical list of attendees (7):

• Damian Alvarez (JSC, Germany)
• Adam Huffman (Big Data Institute, University of Oxford)
• Kenneth Hoste (HPC-UGent, Belgium)
• Victor Holanda (CSCS, Switzerland)
• Davide Vanzo (Vanderbilt University, US)
• Bart Oldeman (ComputeCanada)
• Fotis Georgatos (SDSC, Switzerland)

Agenda

• 2018b common toolchains
• Python in GCCcore
• Q&A

2018b common toolchains

• jump to GCC 7.3.0 as base for both foss/2018b and intel/2018b

  • Kenneth got feedback from Intel support that this should be OK for Intel compilers 2018 update 3 (even though it'll only be officially supported in 2019 version of Intel compilers)

• foss/2018b proposal

  • GCC 7.3 (Jan 25th 2018) + binutils 2.30 (Jan 28th 2018)
  • OpenMPI 3.1.0 (May 7th 2018)
  • OpenBLAS 0.3.0 (May 23rd 2018) + ScaLAPACK 2.0.2 (May 1st 2012)
  • FFTW 3.3.8 (May 28th 2018)
  • cfr. https://github.com/easybuilders/easybuild-easyconfigs/pull/6424

• intel/2018b proposal

  • GCC 7.3 + binutils 2.30 as a base
  • icc + ifort 2018 update 3 (a.k.a. 2018.3.222) (May 29th 2018)
  • impi 2018 update 3 (a.k.a. 2018.3.222) (May 29th 2018)
  • imkl 2018 update 3 (a.k.a. 2018.3.222) (May 29th 2018)
  • cfr. https://github.com/easybuilders/easybuild-easyconfigs/pull/6409

• include GCCcore/6.4.0 as build dependency for GCCcore/7.3.0 to fix installation on CentOS 6?

  • cfr. https://github.com/easybuilders/easybuild-easyconfigs/issues/6374
  • -lto needed for MPFR
    • Davide: dynamically include GCCcore 6.4.0 as build dep for GCC 7.3 via easyblock?
    • Bart: drop -lto for GCC stage 1

• CUDA 9.2 supports GCC 7.3, but not Intel 2018 (only 2017)

Python in GCCcore

• cfr. https://github.com/easybuilders/easybuild-easyconfigs/pull/5072 + https://lists.ugent.be/wws/arc/easybuild/2018-04/msg00045.html
• Python-core is basically a replacement for the system Python, but we're in control over the version
• Bart: seems orthogonal to Python setup @ ComputeCanada
• should get (more) feedback from Markus & Jack
• Davide: potential issues with situations where both are mixed?
• Kenneth: we could add a test that ensures that Python-core is only used as a build dep
• Davide: could also help with situation where python is re-built in SciPy bundle?
  • Damian: was a problem related to performance, python + numpy with Intel is faster
  • mostly synthetic benchmarks, so actual need is unclear...
• Damian: jupyterhub (needs numpy) on top of hierarchy is not very convenient
  • looking into alternate approaches for that, ComputeCanada approach seems interesting
  • have everything built at GCCcore level?
    • have a (hidden) imkl (without FFTW interfaces) at GCCcore, which takes out compiler+MPI out of the equation
      • OK since imkl is a binary installation
• Bart: be careful with libpython.so being RPATH'ed to Python-core...
• Python environment mess: https://m.xkcd.com/1987/; pushing down stuff to GCCcore isn't helping either

Q&A

• https://jitsi.org/ as potential replacement for Google+ Hangouts?
• Fotis: alternative for imkl to ensure openly available toolchains compoents?
  • cfr. iomkl can be used without Intel MPI (Intel compilers + OpenMPI + imkl)
• Victor:
  • shift EB conf call with 1 week to avoid clash with GROMACS conf call?
  • load a module with lots of dependencies with Lmod takes a long time
    • http://lmod.readthedocs.io/en/latest/210_load_storms.html
    • cfr. --recurvsive-module-unload in EasyBuild, or depends_on (supported since EasyBuild v3.6.0)