WCET (worst-case execution time)

Averages are of no interest when an airbag needs to fire. Schedulability analysis such as rate-monotonic needs an upper bound per task, and that bound must be safe, not typical. Measurement only gives you the longest time you happened to observe; the true worst path may never have run during testing.

Hence static tools such as AbsInt's aiT, which analyse the machine code and model the processor: loop bounds, cache misses, pipeline stalls. That is where it gets hard. Caches, speculative execution and multiple cores sharing a bus make a modern CPU fast on average and nearly impossible to bound tightly. It's one reason avionics and brake controllers favour simple cores without caches — slow, but predictable.