-
Notifications
You must be signed in to change notification settings - Fork 148
Conference call notes 20251105
Kenneth Hoste edited this page Nov 19, 2025
·
7 revisions
(back to Conference calls)
Notes on the 282st EasyBuild conference call, Wednesday 05 Nov 2025 (08:00 UTC / 10:00 CET)
List of attendees (16):
- Bob Dröge (Univ. of Groningen, The Netherlands)
- Alex Domingo (Vrije Universiteit Brussel, Belgium)
- Loris Ercole (CECAM)
- Davide Grassano (CECAM)
- Jasper Grimm (University of York, UK)
- Jorge Delgado Guerrero (University of Luxembourg)
- Leonardo Honfi (Wageningen University)
- Georgios Kafanas (University of Luxembourg)
- Kurt Lust (UAntwerpen / LUMI)
- Alan O'Cais (freelancer)
- Mikael Öhman (Chalmers University of Technology, Sweden)
- Jarne Renders (Vrije Universiteit Brussel, Belgium)
- Jan Reuter (JSC, Germany)
- Jörg Sassmannhausen (Imperial College London)
- Helena Vela (DoITNow)
- Cintia Willemyns (Vrije Universiteit Brussel, Belgium)
- overview of recent developments
- update of common toolchains
- Q&A
- latest EasyBuild release: EasyBuild v5.1.2 (26 Sept 2025)
- next (stable) EasyBuild release: EasyBuid v5.2.0 (ETA: mid Nov'25)
- incl. support for LLVM-based toolchain, NVHPC toolchain, etc.
-
LLVM-based toolchain
- GCCcore as base
- LLVM C and FORTRAN compilers on top
- Main framework PR in https://github.com/easybuilders/easybuild-framework/pull/4914
- The basisc are there, can already be tested and used to build stuff
- (Alan) will this lead to an explosion of PRs? variants of the same software for foss, intel, llvm?
- (Jorg) current situatuin with Intel is that it is only used with things that needed
- (Jan) taking into account nvidia-compilers and variants with/without CUDA that can also lead to even more easyconfigs
- (Alex) easyconfig will still be different enough to not allow for toolchain-agnostig solutions to be implemented
- (Miket) we also want to test each combination of version/toolchain that we ship with EB
-
NVHPC toolchain rework
- change on default cuda detection will be merged and adapted to open PRs (https://github.com/easybuilders/easybuild-easyblocks/pull/3924)
- following-up on Slack discussion: GCCcore is already a dep of
nvidia-compilersso that avoids relying on host system - new nvidia-compilers/NVHPC toolchian is well tested in VUB and JUWELS, already work being done to use in ESSII as well
- modifying the NVHPC ecosystem with custom MPI implementation or external numerical libraries is painful and breaks very easily
-
CUDA compute capabilities
- (Caspar) compute capabilities from version 9.x can have a revision component (e.g.
9.0a)- nothing: architecture specific features are disabled
-
a: architecture specific, just current minor supported -
f: family specific, current minor and future ones
- explanation in https://docs.nvidia.com/cuda/cuda-c-programming-guide/#feature-set-compiler-targets
- (Caspar) what should EB do to handle all these new exytra options?
- proposal to just use current
cuda_compute_capabilitiesframework using definition including those letters - (Alexander) not trivial to implement as current code does not assume letters in
cuda_compute_capabilitieselements. This applies at all levels, not just framework code in EB, but support in other tools like CMake or any specific easyconfig - (Alex) the default should change as well, so we need to decide how EB will handle those letters and what default will be set by EB
- (Caspar) handling this through the existing
cuda_compute_capabilitiesallows to have complex definitions with different settings for PTX and device code parts - (Alex) positive thing is that compute capabilities without the letter continue to work, so this change can be gradually implemented wherever it is supported
- (Mikael) might not be trivial to populate templates with complex definitions of cuda compute capabilities
- (Caspar) one option would be to bypass build systems and just directly pass those flags to nvcc
- proposal to just use current
- (Caspar) compute capabilities from version 9.x can have a revision component (e.g.
- future easyconfigs merge sprints planned:
- Mon 15 Dec 2025
- Mon 16 Feb 2026
- Mon 13 Apr 2026
- Mon 15 Jun 2026
- Mon 17 Aug 2026
- aiming for semi-fixed schedule every other month: 3rd Monday in even months
(changed made in PRs marked with * are included latest EasyBuild stable release)
-
blog/docs (merged PRs)
- ...
-
framework (merged PRs)
-
easyblocks (merged PRs)
-
bug fixes
- Honor user-defined configopts in LLVM easyblock (PR #3967)
- fix crash in LLVM easyblock when doing module-only or sanity-check-only rebuild (PR #3977)
- Show summary of relevant LLVM test failures (PR #3962)
- Move setting of $PYTHON* environment variables to load_module method in Python easyblock (PR #3958)
-
enhancements
- ...
-
updates
- ABAQUS: add missing question pattern when building 2025 without hotfixes (PR #3980)
-
changes
- ...
-
new
- ...
-
code cleanup
- ...
-
CI
- ...
-
bug fixes
-
easyconfigs (merged PRs)
- Around 100 easyconfig PRs were merged since last conf call \o/
-
bug fixes/reports
- ...
-
enhancements
- ...
-
(noteworthy) new software
- ...
-
noteworthy software updates
- ...
-
cleanup
- ...
-
changes
- ...
-
blog/docs (open PRs + issues)
- add blog post about reproducible tarballs (blog PR #9)
- add FlexiBLAS blog post (blog WIP PR #13)
- Added documentation for entrypoints (docs WIP PR #331)
-
framework (open PRs + issues)
-
bugs
- pass dependencies to
toolchain.preparewhen setting up build environment for extensions (PR #5023)- fixes a problem with not finding header files provided by dependencies when using
--search-path-cpp-headers=include_pathsbecause$C_INCLUDE_PATH& co set by toolchain only include paths for toolchain components
- fixes a problem with not finding header files provided by dependencies when using
- Pass correct
-marchoption on RISC-V systems whenoptarchtoolchain option is enabled (PR #5029)
- pass dependencies to
-
enhancements
- Make specifying
exts_defaultclassoptional (PR #4800) - new LLVM based toolchain (PR #4914)
- make EasyBuild plugin-able through entrypoints (PR #4918)
- make NVHPC a full toolchain with nvidia-compilers, NVHPCX and NVBLAS (PR #4927)
- refactor NVHPC easyblock into NvidiaBase, nvidia-compilers and NVHPC (easyblocks PR #3788)
-
clickCLI wrapper around normal CLI (PR #4961) - Add
--keep-goingoption to fail at the end not at first failing installation (PR #5022) - Add templates for source with just version numbers (PR #5025)
- Add templates for patch versions (PR #5028)
- Make specifying
-
code cleanup
- ...
-
changes
- Deprecate support for running EasyBuild with Python < 3.9 (PR #4966)
-
tests
- Add decorator to ignore PR test failures due to rate limit (PR #5019)
-
bugs
-
easyblocks (open PRs + issues)
- bug fixes/reports
- Potentially wrong
cblas.hin OpenBLAS installations with 64bit support (issue #3872) - Update Clang easyblock for fixed 'finalpath' source attributes (PR #3961)
- Potentially wrong
- enhancements
- updates
-
changes
- ...
-
code cleanup
- ...
-
new easyblocks
- new easyblock for
systemd_wrapper(WIP PR #3963)- use case for this was xorg-server, but also other things like nvtop eed
udev
- use case for this was xorg-server, but also other things like nvtop eed
- new easyblock for
- bug fixes/reports
-
easyconfigs (open PRs + issues)
- bug fixes/reports
-
enhancements
- ...
- (noteworthy) new software
- software updates
-
changes
- ...
- ETA early 2026
- will likely use GCC 15.x as base...
- (Davide) https://github.com/easybuilders/easybuild-framework/pull/5038
- (Davide) rebuilds can break whenever the the linker picks the "old" lib still present in the install directory instead of the new one just rebuilt (in build/test directory)
- (Alex) changing that precedence order seems to defeat the purpose of having those absolute RPATHs to install dire at all, since the relative ones (through ``$ORIGIN`) will be used every single time. Also not clear if this can break other situations