-
Notifications
You must be signed in to change notification settings - Fork 145
Conference call notes 20211013
Kenneth Hoste edited this page Oct 26, 2021
·
5 revisions
(back to Conference calls)
Notes on the 183th EasyBuild conference call, Wednesday Oct 13th 2021 (15:00 UTC)
Alphabetical list of attendees (10):
- Sebastian Achilles (Jülich Supercomputing Centre, Germany)
- Simon Branford (Univ. of Birmingham, UK)
- Jasper Grimm (University of York, UK)
- Kenneth Hoste (HPC-UGent, Belgium)
- Kurt Lust (Univ. of Antwerp, Belgium + LUMI User Support Team)
- Robert Mijakovic (LuxProvide)
- Mikael Öhman (Chalmers University of Technology, Sweden
- Bart Oldeman (Compute Canada)
- Åke Sandgren (Umeå University, Sweden)
- Alexandre Strube (Jülich Supercomputing Centre, Germany)
- overview of recent developments
- support for using Intel SVML in numpy
- broken PMIx detection in OpenMPI 4.x
- Perl-minimal as wrapper around system Perl in the style of OpenSSL?
- Q&A
- release timeline
- latest release: EasyBuild v4.4.2 (Sept 7th 2021)
- next release
- EasyBuild v4.5.0 (some significant enhancements already in the pipeline)
- ETA: mid Oct'21 (?)
- project with target issues/PRs for this release: https://github.com/orgs/easybuilders/projects/14
- recent changes
- framework
-
easyblocks
-
bug fixes
- fix installation of
libcp2k
for recent CP2K versions (PR #2585)
- fix installation of
-
enhancements
- update
MotionCor2
easyblock to consider to locations for built binary (PR #2541 + PR #2598) - enhance FlexiBLAS easyblock to support building with Intel MKL (imkl) as backend (PR #2588)
- related easyconfigs PR: #14082
- requires pulling
imkl
to compiler-only toolchain, to avoid having to use an MPI toolchain for FlexiBLAS... - add separate easyconfig for MKL FFTW wrappers (
imkl-fftw
?) - is it worth renaming
imkl
tomkl
orintel-mkl
?- is possible, but rather painful
-
imkl
should only be added as a dependency for FlexiBLAS onx86_64
- can be done via
arch=
in dependency version
- can be done via
- update MATLAB easyblock for 2021b (jre is no longer included) (PR #2589)
- update OpenCV easyblock to detect GTK3 and GTK2 dependencies (next to GTK+) (PR #2591)
- enhance COMSOL easyblock to change permission on build directory during extract step (to allow using patches) (PR #2598)
- update
- new easyblocks
- (none)
-
changes
- (none)
-
bug fixes
- easyconfigs
- to merge/fix/tackle soon
-
framework
-
reported bugs / bug fixes
- sources for extensions are still downloaded with
--module-only
(issue #3849)
- sources for extensions are still downloaded with
-
enhancements
- use separate different progress bars for different aspects of the installations being performed (WIP) (PR #3844)
-
changes
- ...
-
reported bugs / bug fixes
-
easyblocks
- reported bugs / bug fixes
-
enhancements
- enhance GCC easyblock to add support for AMD GPU offloading (PR #2578)
-
changes
- don't use
--config=mkl
for TensorFlow 2.4+ (PR #2583)- cfr. reported performance problems for CPU-only TensorFlow installations (issue #2577), which can worked around via
export OMP_NUM_THREADS=1
- blocked by broken TensorFlow tests when not using
--config=mkl
(see https://github.com/tensorflow/tensorflow/issues/52151) - should we make not using
--config=mkl
opt-in for now, so we can switch to it for selected (latest) TensorFlow versions?
- cfr. reported performance problems for CPU-only TensorFlow installations (issue #2577), which can worked around via
- don't use
-
new software
- (nothing major?)
-
easyconfigs
-
bug reports & fixes
- remove superfluous
-DCMAKE_BUILD_TYPE
(PR #13384) - TensorFlow
tf.matmul
ends up using CPU backend for 32bit floats (issue #14120)- low/wrong performance for matrix multiplication with certain data types (32-bit float, probably also 16-bit)
- TensorFlow seems to favor MKL on CPU over GPU, unclear why...
- should we auto-disable use of
mkl-dnn
when building on GPU? - working fine in TensorFlow 2.6.0 (check changelog?)
- remove superfluous
-
enhancements
- Add CI check for
CMAKE_BUILD_TYPE
(PR #14008)
- Add CI check for
- changes
- update to UCX 1.11.2 as dependency for OpenMPI 4.1.1 (PR #14090)
- new software
- noteworthy software updates
-
bug reports & fixes
-
framework
- for now:
foss/2021.07
andintel/2021.09
(candidates for2021b
after testing confirms they work well)-
foss/2021.07
: included with EasyBuild v4.4.2 release -
intel/2021.09
: WIP at PR #14085- includes
intel-compilers
2021.4 release, which support GCC 11.2 - includes
impi
2021.4 on top of UCX 1.11.2
- includes
-
-
PR #14090 (+ PR #14091) suggests bumping UCX to 1.11.2 (from 1.11.0) for the OpenMPI involved in
foss/2021.07
- failing tests for
SciPy-bundle
withfoss/2021.07
(PR #13789) - toolchain working group to follow up on this (?)
- see https://www.phoronix.com/scan.php?page=news_item&px=Intel-Numpy-AVX-512-Landed
- easyconfigs issue #14121
- nice performance boost for AVX-512 systems
- will be included in next numpy release
- seems to be opt-in?
- some of these changes affect accuracy as well (but still good enough?)
- seems similar to precision when using Intel MKL
- some tweaks needed for numpy
- see https://github.com/open-mpi/ompi/issues/8823
- easy to patch, cfr. https://github.com/artpol84/ompi/commit/47202ca8191ba857fb7707223118c5800310e1ed
- for a wide range of OpenMPI versions (except v4.1.1)
- only relevant for stuff building with
system
toolchain - assume that necessary Perl modules are available
- only symlink
bin/perl
to system Perl? - fall back to building Perl from source
- should be hidden
- Kurt: other option could be to add a module that should be installed if OS dependency is missing
- we should hide more build dependencies like binutils with system, Python-bare, etc.
- only going forward, since there would be impact for existing stuff
- Anyone tried to build TensorFlow on top of ROCm with EasyBuild?
- Kurt: not yet in LUMI...