Hi Kenneth,

No problem to modify the sanity check, did this before.

Querying the installer for available components may be a lot harder, unless
Intel provides a clean mechanism.  I can look into it, but I doubt there
will be an reasonable solution except for parsing the provided install.sh,
but that is likely a very brittle.

Adding COMPONENTS=ALL to the generated silent.cfg in the imkl.py EasyBlock
is the most straightforward solution, but it will potentially unstall too
much (e.g., support for the PGI compiler suite, no problem for us, since we
have it, but many won't).

The absence of ScaLAPACK was also noted here, which triggered the "quest".

Best regards, -gjb-



On Mon, Mar 17, 2014 at 10:54 AM, Kenneth Hoste <[email protected]>wrote:

>  Hi Geert Jan,
>
> On 17/03/14 10:41, Geert Jan Bex wrote:
>
> Dear all,
>
>  I've got a problem installing Intel MKL version 11.1.1.106.
>
>  When I just install using EasyBuild, I don't get the libraries for the
> components that depend on MPI, such as BLACS and friends.  When I do the
> install by hand, I get a similar set of library files, but I can optionally
> install more components by selecting them at some stage of the
> installation.  A difference between the silent.cfg generated during that
> installation and what seems to be produced by the EasyBlock for imkl is the
> COMPONENTS value.
>
>  My question is two-fold:
>
>    1. Do I miss something that makes the current EasyConfig/EasyBlock
>    make install the MPI dependent libraries (i.e., the cluster components?
>    2. If not, then I assume Intel changed the default behaviour of its
>    installer for this MKL release, and the EasyConfig file and imkl EasyBlock
>    should be modified.  The cleanest way to do this seems to be to introduce a
>    variable akin to 'interfaces' that holds the components  to be installed in
>    the EasyConfig, and make the imkl.py add it to the silent.cfg it generates.
>     Or are there alternatives?
>
>   This problem is new to me, so it seems like Intel changed the default
> behavior.
>
> There were other changes to silent.cfg since Intel MKL 11.1 (e.g. renaming
> keys like from 'ACTIVATION' to 'ACTIVATION_TYPE', etc.), so it's likely
> this change is related.
>
> Apparently the sanity check for imkl is missing this aspect, so it should
> definitely be enhanced to also check for these MPI dependent libraries.
>
> Adding a custom easyconfig parameter like 'components' seems to make
> sense, yes.
> Apparently, something like " COMPONENTS=ALL" is also supported (cfr.
> http://software.intel.com/en-us/articles/intel-mkl-111-install-guide).
> It seems like that should be the default (I assume that corresponds to the
> behavior like it was before).
>
> If possible, we should verify what is being passed to the "components"
> easyconfig parameter, before we add it to the silent.cfg. That seems like a
> whole can of worms though, unless we can query the installer for supported
> components somehow...
>
> Geert Jan: are you up for looking into this? The first step should
> probably be to enhance the list of sanity check paths, to avoid this
> passing unnoticed... Feel free to ping me over IM if you have any questions.
>
> BTW, this also explains the issue reported by Ward last Friday about the
> missing ScaLAPACK libs and the CP2K build failing because of that. See
> https://github.com/hpcugent/easybuild-framework/issues/883 .
>
>
> regards,
>
> Kenneth
>
>
>
>
>  Best regards, -gjb-
>
>  --
>  - dr. Geert Jan BEX
>  - HPC analyst/consultant Hasselt University & University of Leuven
> - email:                [email protected]
>  - homepage:         http://alpha.uhasselt.be/~gjb/
> - PGP public key: http://alpha.uhasselt.be/~gjb/pgp.txt
> ---------------------------------------------------------
>
>
>


-- 
- dr. Geert Jan BEX
- HPC analyst/consultant Hasselt University & University of Leuven
- email:                [email protected]
- homepage:         http://alpha.uhasselt.be/~gjb/
- PGP public key: http://alpha.uhasselt.be/~gjb/pgp.txt
---------------------------------------------------------

Reply via email to