`container_spec_cpu_shares` will use wrong cgroups v2 conversion formula for runc v1.3.2+

[CatMachina]

cadvisor uses the linear conversion metric when converting from cgroups v2 cpu.weight to the correct cpu limit share value (ref).

Although kubelet has this conversion, it's not being used anywhere, so presumably the cpu resource request is being converted into cgroup v2 cpu.weight in the containerization layer. runc 1.3.2 will use a different conversion formula, when they bump their cgroups dependency from 0.0.1 to 0.0.4 (ref), as the new conversion is done in 0.0.3 (ref).

This would result in container_spec_cpu_shares to become inaccurate as the container runtime layer would use the new conversion to convert shares to weight, and cadvisor the old conversion from weight to shares. I did some calculation to see what the discrepancy would be, and it seems quite big (feel free to double check my math on this):
https://www.desmos.com/calculator/jwnh1p4fov

Is there a plan in place to have a stable version with this new formula from runc?

[SergeyKanzhelev]

cc: @bitoku @iholder101

This PR supports both: kubernetes/kubernetes#132791 in tests now. And previous PR kubernetes/kubernetes#124216 eliminated the use of the formula from k/k.

Clean up: kubernetes/kubernetes#134409

What is the question specifically? There is a switch to the new formula happening. Please comment here: kubernetes/enhancements#5246

[iholder101]

I was raising this in the discussion here.

TLDR: I envisioned this change to be around the CRI opting in/out of the new formula, this way it would have been possible to ensure the formula is enabled on both runtimes on the same k8s version. However, it ended up being changed as an OCI-level implementation detail (that's out of k8s' control).

With the current situation, I guess the KEP can be closed (or GA for documentation purposes?), and cadvisor's conversion method needs to be updated.

cc: @giuseppe