Conference call notes 20190710 - easybuilders/easybuild GitHub Wiki

(back to Conference calls)

Notes on the 128th EasyBuild conference call, Wednesday July 10th 2019 (17:00 - 18:00 CEST)

Attendees

Alphabetical list of attendees (6):

  • Damian Alvarez (Jülich Supercomputing Centre)
  • Fotis Georgatos (SDSC, Switzerland)
  • Victor Holanda Rusu (CSCS, Switzerland)
  • Kenneth Hoste (HPC-UGent)
  • Bart Oldeman (ComputeCanada)
  • Davide Vanzo (Vanderbilt University)

Agenda

  • update on progress towards EasyBuild 4.0: what's left to do for an alpha release?
  • 2019b toolchains
  • Q&A

Update on progress towards EasyBuild 4.0

Other

  • 2019b toolchains

    • WIP foss/2019b: GCC 8.3, OpenMPI 3.1.4 (not 4.x) + latest versions
    • intel/2019b: latest versions
    • Davide: timeline?
    • vetting of proposed toolchain definitions
      • test with existing easyconfigs via eb *foss-2018b*eb --try-toolchain-version 2019b --robot --job

  • Davide: removing binutils version from GCC toolchain?

    • GCC/8.3.0-2.32 -> GCC/8.3?
    • we only make a single combo of GCC+binutils
    • Victor: in CSCS different combos of GCC & binutils are made
    • similar for iccifort version...
      • drop -GCC-* part of iccifort modules
      • Damian: not done yet at JSC because of concerns related to CUDA compatibility
      • take opportunity to rename to human readable: iccifort/2019_update2
    • Damian: dropping binutils versionsuffix has been done for a while already, no major issues
    • only use binutils version as exception to the rule
    • adding exception could cause issues in HMNS setup w.r.t. $MODULEPATH expansion for GCC built with different binutils
      • just using a different binutils as build dep could be enough to make an exception
    • we could apply these changes for the 2019b common toolchains

  • Damian: better way to deal with MPI updates in HMNS setting at JSC

    • avoid massive rebuild when MPI update is installed
    • example: impi 2019 update 1 -> 2019 update 2
    • planning to change naming scheme for MPI modules, to drop minor versions from module tree subdirectory
      • same $MODULEPATH extensions, but different impi installations
    • possible pitfalls
      • installations with RPATH/hardcoded paths
      • unintended breaking of API/ABI with updated MPI lib
      • hopefully only exceptions
    • Bart: also applying the "one digit rule" for GCC (for example. updating of GCC/8 to a newer GCC 8.x)
      • software gets installed to <install_prefix>/Core/GCC/8/...
      • Damian: bit more conservative with compilers at JSC because of runtime libraries vs compiler version used to install the software with (been bitten by issues in the past)

  • Victor: performance change for GROMACS built with and without EB?

    • w/ GROMACS 2018.4
    • slower when built with EB (50%!), because of CMake bug?
      • 18ns/day -> 37ns/day
      • GROMACS uses (incorrectly compiled) cpuinfo tool to drive CMake
    • updating to newer CMake doesn't help
    • didn't see the problem when building by hand...
    • cross-compilation on (Cray) Skylake login nodes for Haswell host
      • also happens when compiling on Haswell for Haswell host
    • Damian: also seen with GROMACS 2019 at JSC (not Cray)
    • GROMACS issue tracker URL: https://redmine.gromacs.org/issues/2927
    • ask Åke?
    • Fotis: check GCC spec strings for possible culprits?

  • Davide: Lmod question

    • GCCcore is hidden (only GCC & iccifort are visible)
    • ml spider suggests loading GCCcore for some things (e.g. Perl), which confuses user
    • some workaround in place via avail hook: merging of modules for GCCcore with those of GCC or iccifort
    • better approach to change output of ml spider to replace GCCcore with GCC or iccifort?
    • Damian: preloading of GCCcore+binutils at JSC
    • ask Ward?

  • Damian: toolchain for AMD-based systems?

    • no support yet for AMD compiler in EasyBuild
    • ping Miguel? ask EB ML?