Add cgroups support for Linux targets

[seanmonstar]

Closes #80

[olix0r]

Thanks @seanmonstar !

cc @lucab

[seanmonstar]

It may be a good idea to add a CI job that run the tests inside docker with cgroups limiting to just 1 CPU.

[seanmonstar]

Well hey there, docker CI job testing with and without setting --cpus works!

[olix0r]

I can't vouch for all of the details but the overall approach looks good to me!

[lucab]

Overall looks fine to me too, thanks!

I think there is a global question about flooring vs ceiling the CPU quota.

Additionally, this only covers cgroup v1. It may need to be re-touched at some point in the future for cgroup v2.

[strohel]

This functionality seems to be perfect. In a setup with 4 physical cores and Docker container being limited to 1.0 CPU, a benchmark of a simple async actix-web based microservice resulted in 1 worker thread being run instead of 4, causing up to 50% increase in CPU efficiency:

Full benchmark report: https://storage.googleapis.com/strohel-num-cpus-containerized/bench-results.html

Workloads like this also ask for ceiling, rather than flooring, non-unit CPU soft-limits, as otherwise actix-web would not be able to saturate available resources (I can perform benchmark with e.g. 1.5 CPUs if desired).

This change can easily have high reach and noticeable impact (benchmark above being probably the extreme case), as num_cpus::get seems to be used in major async runtimes (and beyond), and it is probably not uncommon to limit CPU usage in container orchestration systems like Kubernetes. Maybe it would be worth a prominent announcement?

[seanmonstar]

Ok, got some tweaks in, along with ceiling the result (and a test that --cpus=1.5 for docker works as expected).

[seanmonstar]

@strohel mind if I use that graph in the release notes?

[strohel]

@strohel mind if I use that graph in the release notes?

Sure! An SVG version (if handy) is at https://storage.googleapis.com/strohel-num-cpus-containerized/bench-results.cpu_per_request_figure.svg

It should be probably accompanied by some context from my above post in order not to create false expectations (the same efficiency increase could be achieved by manually setting ideal number of workers; I still think this is a huge win as people probably don't do it).

[nagisa]

What happens if the cgroup limits are changed while a program is running?