Avoid numactl if not available

[ChSonnabend]

Using the slc9-gpu-builder image, numactl was not availble when pulling via apptainer (apptainer pull slc9-gpu-builder.sif docker://registry.cern.ch/alisw/slc9-gpu-builder:latest). This was needed to execute the FST since docker is not yet available on the pdp-dev00 machine, but also future-proofs in non-writeable images.

[ChSonnabend]

Ping @davidrohr @ktf

[davidrohr]

But this will silently break the correct NUMA pinning of processes. Why do you want to run the FST script without numactl available?

By default it should defikitely fall with a fatal error. If you want, you can add an environment variable to allow running without numactl for debug reasons, but I do not want to allow that by default.

[ktf]

Can you in general run numactl from inside a container?

[davidrohr]

I think so, you should be able to repin processes among the cpu cores the container may use.

[ktf]

ok, then we need to have the numactl rpm added to the base RPM. @vkucera can you take care of it?

[davidrohr]

But why do we run the fst in the container at all?
This fst scripts assumes an epn setup, with 512gb of memory and 8 gpus.

[ktf]

I will let @ChSonnabend to motivate that.

[ChSonnabend]

The main motivation was to use a central script for benchmarking the performance on different systems (e.g. the dev server with a single RTX5080 GPU). As privately discussed with David, it might not be the optimal solution to use the current FST setup for it. We will iterate, for now I will close this PR.