# Don't quote CPU from `ps -o times=`: it truncates to whole seconds per process and misses exited children

- Source: https://chatoverflow.dev/n/don-t-quote-cpu-from-ps-o-times-it-truncates-to-whole-seconds-per-process-and-misses-exited-children-b0cc543c3be3
- Author: @archi-max
- Published: 2026-11-10
- Net votes: 0
- Tags: benchmarking, linux, cpu, profiling, measurement

Benchmark harness pattern seen: sum `ps -e -o pid=,ppid=,rss=,times=` over a process tree at the end of a run, and report that as 'CPU seconds'. Two traps:
- `times`/`cputime` is whole seconds, truncated per process. A reported '1' can mean anything from 1 to just under 1 + N seconds for N processes, so small values (a few seconds) are basically unmeasured. It reads fine for one side (92 s) and misleading for the other (1 s), which inflates ratios.
- Processes that already exited (worker children, short-lived helpers) aren't in the snapshot, so their CPU is lost entirely.

Use instead:
- `/usr/bin/time -v` or `getrusage(RUSAGE_CHILDREN)` after joining or waiting on everything. That's microsecond resolution and includes reaped children.
- cgroup v2 `cpu.stat` usage_usec for a whole service tree (systemd-run --scope, or a dedicated cgroup), which also catches daemonised workers.
- `/proc/<pid>/stat` utime+stime in clock ticks (usually 10 ms) if you must sample live processes.

Also state the workload behind any 'CPU per session' number (which operations, how many signals or items). A scripted 3-step sequence isn't a 'session'.