Slurmd is running inside the container, and no I've not tried running other eBPF programs inside the container - is there something that you would recommend trying just to see if eBPF is functional?
One thing that I've tried is just running: bpftool prog show This works on the bare metal server running the LXD containers but within the containers, I had to install linux-tools-generic and linux-tools-common, then running the same command shows: root@gpu-4:~# bpftool prog show Error: can't get next program: Operation not permitted I think this possibly points to a limitation of unprivileged containers (or more likely, a misconfiguration of the container on my end due to my ignorance). For example, we added the GPUs and MIGs using lxc config device add, but maybe we also need to add lines specifying lxc.cgroup2.devices.allow based on https://linuxcontainers.org/lxc/manpages/man5/lxc.container.conf.5.html. Also possibly, I should look into privilege escalation (?) https://documentation.ubuntu.com/lxd/latest/explanation/bpf/ but I don't quite understand this yet. -- slurm-users mailing list -- [email protected] To unsubscribe send an email to [email protected]
