Conversion of cgroup v1 CPU shares to v2 CPU weight causes workloads to have low CPU priority

[mwek]

Description

In cgroup v1, the default CPU shares is 1024.
In cgroup v2, the default CPU weight is 100.

When creating a container with 1024 CPU shares, the current formula translates the value to 39, resulting in ~39% of the actual CPU weight it should receive.

To preserve the default values, we should use a different formula: CPUWeight = 1 + ((CPUShares – 2) * 99) / (1024 - 2)

Reference:

• Conversion of cgroup v1 CPU shares to v2 CPU weight causes workloads to have low CPU priority kubernetes/kubernetes#131216
• change conversion from cgroup v1 to cgroup v2 containers/crun#1721

Steps to reproduce the issue

1. container_id=$(docker run -d --cpu-shares 1024 ubuntu bash -c 'sleep infinity')
2. cat /sys/fs/cgroup/system.slice/docker-${container_id}.scope/cpu.weight (prints 39)

Describe the results you received and expected

The CPU weight is returned as 39, where I expect it to return 100.

What version of runc are you using?

1.2.4

Host OS information

No response

Host kernel information

No response

[lifubang]

One question, I think it's useful for the default value of cpu.shares and cpu.weight. But maybe unfair to other values of cpu.shares? For example:
container_1: docker run -d --cpu-shares 102400 ubuntu bash -c 'sleep infinity'
container_2: docker run -d --cpu-shares 204800 ubuntu bash -c 'sleep infinity'

[mwek]

One question, I think it's useful for the default value of cpu.shares and cpu.weight. But maybe unfair to other value of cpu.shares?

I think we should keep the following invariants:

1. cpu_shares are evaluated linearly, i.e. a container with 2x the cpu_shares should have the relevant CPU weight being 2x
2. Default cpu_shares = 1024 translates to default cpu.weight = 100.
3. Minimum cpu_shares = 2 translates to minimum cpu.weight = 1.

Unfortunately, cgroup v2 weight is more restrictive -- it allows for values of 1-10000, with max weight is 100 * default, where cgroup v1 cpu_shares were accepting 2-262144, with max shares being 256 * default. systemd is doing value clamping here, I would vote for consistency.

[kolyshkin]

Yes, we're aware of the issue are are working to fix this. Here's a PR: opencontainers/cgroups#20

[kolyshkin]

@mwek feel free to review and test the above and join the discussion in kubernetes/kubernetes#131216