On 22/08/16 15:45, Maxime Boissonneault wrote:
Don't Numpy and Scipy require MKL or some optimized BLAS version though (it should) ?
Also, doesn't mpi4py require MPI ?
How are these not "additional deps" ?

Because both MPI & BLAS are included in the toolchain already (at least the one typically used for installing Python).

For Python packages, R packages, Perl packages, Lua packages, etc., we usually instruct our users to either install them in their home (using pip --user for Python for example), or using a virtual environment (for Python). In order to avoid build headaches for our users with this strategy, we use python wheels (http://pythonwheels.com/) to pre-build most used Python packages. This way, installing the packages (either in their home or a virtual environment) takes mere seconds.

I guess this is a matter of site policies, but we try to avoid letting users handle this themselves, for a variety of reasons (one being that our setup involves a single (set of) login nodes for multiple clusters with different operating systems).

It would be nice to have wheel support for Python packages within EasyBuild.

It is already supported, see for example https://github.com/hpcugent/easybuild-easyconfigs/blob/master/easybuild/easyconfigs/t/testpath/testpath-0.3-intel-2016a-Python-2.7.11.eb .




On 16-08-22 09:40, Ward Poelmans wrote:
On 22-08-16 15:27, Joachim Hein wrote:

Looking into e.g. Python-2.7.12-intel-2016b.eb it seems this is building
stuff like scipy, numpy and mpi4py (sometimes old versions). Why are
the separate easyconfigs for those kind of packages as well? Are the
later obsolete?  Someone minds clueing me?
The current policy is to include 'extensions' in the main easyconfig if
they don't require extra dependencies. If they need additional deps, we
create a separate easyconfig. The reason that there something are
separate easyconfigs for mpi4py and such is if we need a newer or older
version then included in the main easyconfig.


Reply via email to