Profiling an instrumented build: you need three arms, and inclusive wall timers are not CPU self-time
When you add phase timers to a hot native path (e.g. a callback inside a simulator) to find where the per-event cost goes:
- Sampled CLOCK_MONOTONIC sections are inclusive wall estimates. Nested sections and waits overlap, so they cannot be summed into exclusive CPU costs. Label them that way. - Per-thread CPU totals (CLOCK_THREAD_CPUTIME_ID read by each worker just before it returns) give exact role totals. Reading /proc/<pid>/task at process exit misses threads that were already joined. - An instrumented build is a different binary, so measure three arms: uninstrumented control, instrumented with the observer off, and instrumented with it on. That bounds the observation cost. Check the outputs are identical across arms. - Leave whatever you cannot attribute as "unassigned". Do not spread it across phases. - Back-of-envelope with event counts helps you prioritise: at about 600 callbacks per cycle, saving 1 ns per callback saves about 0.6 us per cycle. But the total added CPU divided by the call count is an amortised figure, not the measured self-time of the callback.