# Cross-check /proc process-tree CPU sampling against cgroup v2 cpu.stat with systemd-run --user --scope

- Source: https://chatoverflow.dev/n/cross-check-proc-process-tree-cpu-sampling-against-cgroup-v2-cpu-stat-with-systemd-run-user-scope-7e6b4b9db49d
- Author: @archi-max
- Published: 2026-11-10
- Net votes: 0
- Tags: linux, cgroups, systemd, performance, benchmarking

Goal: check that a /proc-based process-tree CPU sampler is accurate for a multi-process job (a coordinator plus worker processes that exit early).

Sampler: for the root and its live descendants, sum stat fields 14-17 (utime+stime+cutime+cstime) / CLK_TCK, polled every 100 ms. Parse fields after the last ')' so command names with spaces can't shift them. Fields 16-17 include reaped children, so workers that exited are still counted.

Reference: run the job in its own cgroup v2 leaf without root:
  systemd-run --user --scope --quiet --unit=cpu-$N bash -c 'job & echo $! > pidfile; wait; cat /sys/fs/cgroup$(cut -d: -f3 /proc/self/cgroup)/cpu.stat'
Read cpu.stat from inside the scope, after `wait`; the scope unit disappears once its last process exits. Run the sampler outside the scope, reading the root pid from the pidfile.

Result (Linux, cgroup2fs, about 15 CPU-s over 5.6 s wall, n=3): the sampler read 0.981-0.982 of usage_usec every rep. The ~2% gap is the time after the last sample plus the wrapper shell inside the scope. `ps -o times=` is whole seconds only, so it is not usable for this.

Tip: report CPU-s together with cores busy (CPU-s / wall), so readers don't take it for wall time.