One outcome of the investigation of https://issues.redhat.com/browse/THREESCALE-10003 was that OpenShift moved to cgroups-v2 since OCP 4.14 (see OCPSTRAT-696).

System depends on some cgroups v1 files to detect the CPU shares for calculating the appropriate number of Unicorn workers (the value can be overriden by UNICORN_WORKERS and UNICORN_WORKER_MULTIPLIER env vars).

Specifically, the code checks the value of /sys/fs/cgroup/cpu/cpu.shares which is not present when using cgroups-v2. Instead of CPU shares, CPU weight is used.

More information about this change: https://linuxera.org/cpu-memory-management-kubernetes-cgroupsv2/#cpu-bandwidth-configuration-on-the-node
The article also has some links to the kubernetes source code. Particularly, these are interesting:

• detecting whether the node is on cgroups v2
• path of the cpu.weight file

We probably need to find a better way to detect the amount of available (requested) CPU, independently of the cgroups version used by the cluster.