Amdahl''''s law and self-time: the ceiling on every speedup you can ship

A team spends three weeks rewriting a hash function for a 10x local speedup. The deployed service is 1.05x faster. Amdahl’s law would have told them the answer before they started — if they had checked the profile first.

Amdahl’s law: the ceiling on every speedup

If a section of code takes fraction p of total execution time, and you make that section s times faster, the total application speedup is:

total_speedup = 1 / ((1 − p) + p/s)

The ceiling case — making the section infinitely fast — gives:

max_speedup = 1 / (1 − p)

A function that is 10% of total time can never give more than 1.11x total speedup, even if you replace it with a no-op. A function that is 80% of total time gives 5x if you halve it.

This law converts “should I optimise this?” into a quantitative question. Profile, see that function X is 12% of execution, and accept up-front that even a perfect rewrite caps your win at 13.6%. If the ceiling is below your bar, do not start. If function Y is 70% of execution, a 2x speedup of Y gives 1.54x total — worth the engineering time.

Wall-clock vs CPU time vs self-time

A profile reports several different time measurements and confusing them is the most common reading error.

• Wall-clock time — what the user feels: how long the request took from start to end. A request at 500 ms wall-clock that only used 50 ms of CPU is 90% waiting (on disk, network, locks, GC).
• CPU time — how much of wall-clock was spent actually computing.
• Self-time — time spent in a function’s own code, not including time spent in functions it called.
• Cumulative time (cum-time) — includes time in callees.

Reading a profile: if cum-time is large but self-time is small, the function is a routing/dispatch layer — the slow work is in something it calls; walk down. If self-time is large, the function is itself doing the work; that is where the fix lives.

Confusing the two leads to optimising a wrapper while the real cost sits in a callee — or vice versa.