An aggregation of GPU utilization (GPU and VRAM) per cgroup?

[gleventhal]

From what I can tell, the G (Cgroupv2) view and the E (GPU) views are mutually exclusive, as are the G and e views, such that (FWICT) there is not any aggregation for GPU utilization per cgroup (v2).

I understand that there are no control files (AFAIK) in cgroupsv2 for GPU, so atop would need to perform the calculations/aggregations, I think.

I think it could be useful, for instance: a multi-tenancy research cluster with many GPUs, to show which jobs are responsible for what GPU utilization on a cgroup level (assuming that a cgroup name is more meaningful to the admins than a PID for some cases).