Just to mention, an approach like Stefans does not necessarily mean users have to load another module to get access to software. For users, you only care about where the software for the architecture lives. In the default profile you can define this with, e.g.,:

export DEFAULT_ARCHITECTURE_MODULEPATH=${SITE_EB_PREFIX}/${ARCH_SUBDIR}/modules/core
module use ${DEFAULT_ARCHITECTURE_MODULEPATH}

where ${SITE_EB_PREFIX} is common to all archs and ${ARCH_SUBDIR} needs to be defined/derived for each architecture.

Then, developers could still use `hpc-env/<ver>` modules if those include the Lmod/Tmod equivalent of `module unuse ${DEFAULT_ARCHITECTURE_MODULEPATH}` at the beginning.

Alan

On 22-Sep-23 9:15 AM, Loris Bennett wrote:
Dear Stefan,

"Dr. Stefan Harfst" <[email protected]> writes:

Dear Loris,

will you not overwrite the existing module if you use --force?
I.e. the module will always point to the software directory of the
architecture you have build last.
I am mounting the directory on each nodes with the architecture-specific
software under

   /sw/sc/easybuild

with each node mounting the correct branch of the NFS directory for its
CPU.  But as you say, --force will overwrite modules, so I need to be
able to load EasyBuild separately from the other software.

We have done something similar and yes, we treat EasyBuild itself differently. 
I will try to outline our setup:
* we have a basepath /cm/shared/uniol
* in the basepath we have modules and sw
* initially, our module path only consists of /cm/shared/uniol/modules/core in 
this path we put the EasyBuild modules and modules we call hpc-env
* if you load hpc-env/<ver> the module path is extended with paths like 
/cm/shared/uniol/<arch>/<ver>/<cat>, where <arch> can be zen3, zen4, ice, sky or so. <ver> 
is actually referring to the GCCcore version but this is not so relevant. <cat> are the module classes like bio, 
chem and so on
* the hpc-env module uses an environment variable to know which is the 
architecture of the current node, so it shows only that part of the module tree
This is an approach that I have seen used elsewhere.  My thinking in
using the mount approach instead was that the users don't have to load
an extra module and that the modules don't have some slightly arbitrary
path, such as '/cm/shared/' (or in our case it is actually
'/trinity/shared/') baked into the 'root' variable.

* the sw-directory has the same structure, so .../sw/<arch>/<ver> and we use 
--subdir-modules and --subdir-software to install the software and module in the right paths (we 
have setup an alias that also identifies <arch> from the env variable)
* we also have a SYSTEM arch for everything build with toolchain = SYSTEM (../sw/SYSTEM and 
correspondingly ../modules/SYSTEM), we link the SYSTEM-modules into all 
<arch>/<ver> (hidden, if there is non-SYSTEM version)
I hadn't thought about dealing with SYSTEM-modules separately.  How does
this work when you build an architecture-dependent, piece of software
with a specific toolchain and '--robot', but which has a dependency
which uses the SYSTEM-toolchain?  Everything will be installed in the
same architecture-dependent subdirectory, won't it?

* Easybuild itself is installed directly in sw but it could also go into SYSTEM 
(or your generic).
OK, so treating EasyBuild differently from the bulk of the software seems
indeed to be way to go.

It means that you need a lot of space and every module is installed
several times, but you are only reusing a working Easyconfig and it
can be somewhat automatized.
We are already resigned to the fact that a lot of space will be needed,
but it looks this sort of approach is the only way to ensure that each
node will see software that will definitely run on it.

Hope this helps.
Thanks for the detailed explanation of your approach.

Cheers,

Loris

PS: See you in Oldenburg at the AK Supercomputing, maybe?

Best wishes
Stefan

--
Scientific Computing

Carl von Ossietzky University Oldenburg
School of Mathematics and Natural Sciences
26111 Oldenburg, Germany

Office: W03 1-139
Phone: +49-441-798 3147
E-Mail: [email protected]
www: http://www.uni-oldenburg.de/fk5/wr

-----Ursprüngliche Nachricht-----
Von: [email protected] <[email protected]> Im 
Auftrag von Loris Bennett
Gesendet: Donnerstag, 21. September 2023 17:12
An: easybuild <[email protected]>
Betreff: [easybuild] Installing EasyBuild as 'generic' architecture?

ACHTUNG! Diese E-Mail kommt von Extern! WARNING! This email originated 
off-campus.

Hi,

EasyBuild plus all our software built with EasyBuild is currently installed 
under

   /nfs/easybuild/software/EasyBuild

However, I am reinstalling software compiled for different microarchitectures 
in a directory structure which looks like the following:

   /nfs/easybuild/arch
   ├── generic
   └── x86_64
       ├── amd
       │   └── zen3
       └── intel
           ├── cascadelake
           └── skylake_avx512

At the moment I am essentially using something like

   eb --prefix=/nfs/easybuild/arch/x86_64/intel/cascadelake --force ...

to rebuild modules.  However, at some point I will want to avoid using 
'--force', but IIUC that will mean that I have to 'unuse' the original module 
path

   /nfs/easybuild/software/EasyBuild/modules

in order to build modules which already exist in the non-architecture-specific 
path.  This in turn, since EasyBuild module itself is in that path, would 
deactivate the EasyBuild module.

Does that mean I have to reinstall EasyBuild under

   /nfs/easybuild/arch/generic

which I guess would be the most consistent solution, or is there an alternative?

Cheers,

Loris

--
Dr. Loris Bennett (Herr/Mr)
ZEDAT, Freie Universität Berlin

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

Reply via email to