Add Concurrent.cpu_requests that is cgroups aware.
#1058
Conversation
|
Could you summarize in one/a few sentences why one should use cpu requests instead of cpu limits? |
|
@eregon CPU limits can lead to throttling and reduced performance under high load. So, as Tim Hockin suggested, I've set only requests in my application at company. I think this is a quiet common pattern and it'll be nice if we can provide method to get cpu requests. |
|
I've read quickly through the blog posts, so it seems the |
|
@byroot @casperisfine Could you review this too? |
| if RbConfig::CONFIG["target_os"].include?("linux") | ||
| if cgroups_v2? | ||
| # Ref: https://github.com/kubernetes/enhancements/tree/master/keps/sig-node/2254-cgroup-v2#phase-1-convert-from-cgroups-v1-settings-to-v2 | ||
| weight = File.read('/sys/fs/cgroup/cpu.weight').to_f |
There was a problem hiding this comment.
It's probably fine, but in compute_cpu_quota we only ever read files we tested for existence prior.
Here we check /sys/fs/cgroup/cpu.max and assume it means /sys/fs/cgroup/cpu.weight exists? Maybe that's a correct assertion, but I personally don't know if it's true.
I think ultimately it would be simpler to not extract cgroups_v1? / cgroups_v2? and check if /sys/fs/cgroup/cpu.weight exists here.
| weight = File.read('/sys/fs/cgroup/cpu.weight').to_f | ||
| ((((weight - 1) * 262142) / 9999) + 2) / 1024 | ||
| elsif cgroups_v1? | ||
| File.read('/sys/fs/cgroup/cpu/cpu.shares').to_f / 1024 |
In K8s environments, cpu requests are often more useful than limits, but concurrent-ruby currently lacks a related method, so I've added it. Below are links to articles about the usefulness of cpu requests.