Hi Damian,
On 28/06/16 17:10, Alvarez, Damian wrote:
Hi Kenneth,
My only concern is that this means that CUDA won’t work in any of these toolchains,
since CUDA 7.5 doesn’t support GCC > 4.9 or Intel > 2015. Presumably, CUDA 8
(to be released in August) solves that, but for the moment this won’t work for CUDA
users.
That's indeed a problem, but I don't think that should block progress on
the update of the common toolchains.
People can stick to 2016a for the time being (or another toolchain based
on GCC 4.x), until NVIDIA releases a CUDA version that is officially
compatible with GCC 5.x.
Also, it seems that people have been fairly successful in combining CUDA
7.5 with GCC 5.x (even though it may not be officially supported yet),
cfr. https://lists.debian.org/debian-devel/2015/12/msg00134.html .
Of course they can define their own toolchains as a workaround.
Other than that looks good to me. Thank you for taking the time to test all
those packages!
Thanks for the feedback!
regards,
Kenneth
PS: PRs are open for the 2016b toolchains, and I'm slowly getting to
opening additional PRs for easyconfigs using those toolchains too:
* intel/2016b: https://github.com/hpcugent/easybuild-easyconfigs/pull/3270
* foss/2016b: https://github.com/hpcugent/easybuild-easyconfigs/pull/3271
* Python 2.7.12 w/ intel/2016b:
https://github.com/hpcugent/easybuild-easyconfigs/pull/3277
* Python 3.5.2 w/ intel/2016b:
https://github.com/hpcugent/easybuild-easyconfigs/pull/3278
Damian
On 28/06/16 16:55, "[email protected] on behalf of Kenneth Hoste"
<[email protected] on behalf of [email protected]> wrote:
Dear all,
I spent some time the last couple of weeks evaluating what I consider
the prime candidates for the foss/2016b and intel/2016b 'common' toolchains.
The intention is to fix the definition for the 2016b toolchains by very
soon (hopefully by the end of this week), and include them in the next
EasyBuild release which will be released shortly after (my mid July).
Please provide feedback on this ASAP, especially if you're interested in
using these toolchains in the near future.
I consider the most recent updates to the foss & intel toolchains prime
candidates for 2016b (basically all latest & greatest, other than GCC 6.1):
* foss/2016.06
* GCC 5.4 + binutils 2.26
* OpenMPI 1.10.3
* OpenBLAS 0.2.18 + LAPACK 3.6.0 + ScaLAPACK 2.0.2
* FFTW 3.3.4
* intel/2016.03-GCC-5.4
* GCC 5.4 + binutils 2.26
* icc & ifort 2016.3.210
* impi 5.1.13
* imkl 11.3.3.210
I evaluated these thoroughly, by rebuilding all foss/2016a & intel/2016a
easyconfigs we have included in EasyBuild (using a naive
'--try-toolchain'). Here are the results:
* foss/2016.06: success for 278/279 easyconfigs
* OK: BioPerl; Boost; BWA; GMP; GROMACS; GSL; HDF5; IPython; LLVM;
libX*; MariaDB; Mesa; netCDF*; Perl 5.22.1 w/ extensions; Python 2.7.11
& 3.5.1 w/ numpy/scipy/pandas/...; R 3.2.3 w/ extensions; QT 4.8.7 &
5.6.0; SAMtools; etc.
* only RMBlast failed due to a hard check on the GCC version
* intel/2016.03-GCC-5.4: success for 384/413 easyconfigs
* OK: BioPerl; BWA; Boost; CLHEP; FSL; GAMESS-US; GATE 6.2 & 7.1;
GMP; GROMACS; GSL; Geant4; HDF5; IPython; LLVM; libX*; MariaDB; Mesa;
NGS; netCDF*; OpenFOAM 2.2.2 & OpenFOAM-Extend 3.2; Perl 5.20.3 &
5.22.2; Python 2.7.11 & 3.5.1 w/ numpy/scipy/pandas/scikit-learn/...; Qt
4.8.7; R 3.2.3 w/ extensions; ROOT; SAMtools; WRF; etc.
* 11 failed builds + 18 more due to missing deps:
* HISAT2: failed due to linking with g++ to Intel-built
libraries (easy fix?)
* Amber: failed due to hard check on GCC version (by CUDA)
* mysqlclient & MySQL-Python: weird symbol issue with
libmysqlclient library (ld.gold issue?)
* PCMSolver: error: namespace "std" has no member "nullptr_t"
* GATE 7.2: error: no operator "<<" matches these operands
(missing "#include <string>" ?)
* Qt 5.6.0: error #809: exception specification for virtual
function ... is incompatible with that of overridden function
* HarfBuzz & GST-plugins: error: identifier
"__builtin_uadd_overflow" is undefined (not implemented by Intel)
Several of the failed builds should be easy to fix, as the need arises
to fix them, and all seem to indicate problems with the software itself
rather than with the Intel compilers or GCC 5.4 being used.
Besides this, most concern 'minor' software packages; the 'big ones'
build without any problems, so these toolchains can be considered
stable/mature imho.
Unless some major technical arguments are raised, I propose we go
through with using foss/2016.06 and intel/2016.03-GCC-5.4 for the 2016b
updates of the common toolchains.
regards,
Kenneth
------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------
Forschungszentrum Juelich GmbH
52425 Juelich
Sitz der Gesellschaft: Juelich
Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498
Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher
Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender),
Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt,
Prof. Dr. Sebastian M. Schmidt
------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------
