-
Notifications
You must be signed in to change notification settings - Fork 145
Conference call notes 20171122
Kenneth Hoste edited this page Nov 22, 2017
·
3 revisions
(back to Conference calls)
Notes on the 90th EasyBuild conference call, Wednesday November 22nd 2017 (5pm - 6pm CEST)
Alphabetical list of attendees (5):
- Damian Alvarez (JSC, Germany)
- Fotis Georgatos (Illumina, UK)
- Kenneth Hoste (HPC-UGent, Belgium)
- Alan O'Cais (JSC, Germany)
- Davide Vanzo (Vanderbilt University, USA)
- report on EasyBuild activity at SC17
- outlook to EasyBuild v3.5.0
- discussion on (optionally) including glibc in toolchains
- workaround for future occurrences of RHEL vs Intel compilers issues?
- Q&A
- cake! http://easybuilders.github.io/easybuild/images/easybuild5.png
- Davide talked with Michigan State University (& others) at booth talks, who showed interest
- some interest with OpenPOWER
- current way they are managing software installations is... meh
- currently: large GitHub repository with bash scripts
- URL?
- suggested EasyBuild as a saner approach
- Davide may get access to a POWER system to help out with this
- Alan has some experience with this already
- CUDA is a pain, installation method used in EB is not available on POWER (and won't be)
- similar issue for Spectrum MPI (RPM install, non-relocatable)
- Alan's workaround was to use SystemMPI easyblock
- no extra info about inclusion of EasyBuild in latest Bright Computing release
- may be interesting to invite BC to EUM'18 (also for Fotis himself)
- people who attended EasyBuild booth talks were asking about
- module naming schemes, what is being exposed to users (concerned with end user experienced)
- how to integrate EasyBuild in current setup
- should be better documented?
- "How To Get Started With EasyBuild" page?
- also experiences/use cases from other sites
- big gap in docs: using custom module naming scheme
- ETA: 1st full week of Dec?
- framework
- https://github.com/easybuilders/easybuild-framework/milestone/50
- several minor bug fixes & enhancements merged
- support for customising EasyBuild via hooks (WIP)
- https://github.com/easybuilders/easybuild-framework/pull/2343
- implementation mostly done, docs TODO
- easyblocks
- https://github.com/easybuilders/easybuild-easyblocks/milestone/42?closed=1
- try and make sure BLAS/LAPACK configuration for R installation is correct
- fixes for CP2K easyblock w.r.t. used compiler options (WIP)
- https://github.com/easybuilders/easybuild-easyblocks/pull/1293
- easyconfigs
- https://github.com/easybuilders/easybuild-easyconfigs/milestone/45?closed=1
- mostly software updates & easyconfigs for new software
- triggered by problems with CentOS 7.3/7.4 updates vs Intel compilers
- also helps with build reproducibility
- experiences with building own
glibc
? - Damian: what about security aspects of this?
- less of an issue compared to system
glibc
, which is used for everything incl. stuff run as root -
glibc
used in toolchain & for scientific software is only used in 'user land'
- less of an issue compared to system
- Damian: is this going to be a problem in the future, are we solving a problem that won't happen again?
- in any case, inclusion of
glibc
provided through EasyBuild should be made optional through configuration option? - should use RPATH for glibc libraries
- helps (a little bit) with security concerns
- avoids breaking system tools like
cp
& co
- compatibility of glibc in toolchain with Linux kernel on updates?
- feedback from ComputeCanada who are using own glibc via Nix layer
- Fotis: GNU Guix does a trick with including placeholder paths that are patched later...
- used for relocation of installations
- ref?
- Nix talk @ HUST'17
- presenter is also going to join EUM'18 in Amsterdam
- Fotis: Shahzeb's buildtest (https://github.com/shahzebsiddiqui/buildtest) at FOSDEM/EUM?
- Shahzeb contacted Kenneth, plans to submit talk proposal for buildtest to FOSDEM
- can maybe also be presented at EasyBuild User Meeting?
- Davide:
- sharing
Core
installations between system with different processor architectures? - motivation is mostly reducing of used disk space
- maybe binary installations should be installed somewhere else (other than
Core
)- e.g.
Binary
(all binary installations) &Core
(compilers, that open up subdir inCompiler
level) on top of hierarchy ->Compiler
->MPI
-
Binary
installations only need to be installed once, can be shared across architectures -
Core
installations are still architecture-specific - e.g. CUDA, MATLAB, ...
- e.g.
- also helps discriminating between compilers (that give access to other stuff) & pure binaries
- GCC build is built for specific architecture by default (beyond our control?)
- cross-compilation support?
- potential impact on performance of GCC itself?
- sharing
- question from Maxime on making exceptions to dependencies being filtered out via --filter-deps
- cfr. https://github.com/easybuilders/easybuild-framework/issues/2351
- use case:
- EasyBuild is configured with --filter-deps=flex,... to leverage flex provided by OS (or actually Nix)
- some builds don't work with too recent flex, so then EasyBuild-provided flex could/should be used
- seems like a case for configuring with --hide-deps=flex instead of --filter-deps=flex?
- add support to specify exception in easyconfig to overrule --filter-deps feels backwards...
- feedback on idea to add toolchain option to specify whether linking to sequential or multi-threaded BLAS should be done?
- https://github.com/easybuilders/easybuild-framework/issues/2350
- basically leveraging the proposed change in R easyblock to a more generic solution
- setting
mt_blas_lapack
toolchain option would result in making$LIBBLAS
equivalent to$LIBBLAS_MT
- Alan: already done in QuantumESPRESSO easyblock, based on
openmp
toolchain option- probably not the right approach
- also applies to
$LIBFFT
, so multiple separate toolchain options needed?