Pitfall in Amdahl's law

Question

Amdahl's law is a formula used to find the maximum improvement possible by improving a particular part of a system.

Could someone explain me what does it mean by "pitfall" in Amdahl's law?

Answer 1

PROLOGUE:

Dr. Gene AMDAHL has formulated this principle, later called an Amdahl's law, for a general, abstracted, organisation of process-flow.

Computer Science started to use his arguments, and as @PeterCordes has noted already, on lowest level ( as deep as on the silicon resources level, ILPx tricks, cache pollution, memory-I/O limits and/or other non-local NUMA side-effects ) there often start surprises of its own kind.

---

OBSERVATIONS:

Yet,
the main criticism of a rather "naively deployed" original formulation of the Amdahl's law in the fields of finding a parallel-computing principal speedup ceiling, is the absence of adding all software-related add-on costs, that become added if an original, pure- [SERIAL] code, started to be re-decorated, with add-on tricks, so as to be able to run in a "just"- [CONCURRENT] or a true- [PARALLEL] fashion.

Not adding all of these add-on costs ( that are not present is a pure- [SERIAL] original code ) into the comparison principally skews the results of using the "overhead-naive" original Amdahl's law formulation "mechanically".

Similarly, there are cases, where the original formulation does not correctly account for cases, when there are positively some amount of free-resources that do not participate on the speedup, as the computed job/task contains some "atomic"-section, that can be computed but in block. Thus the free resources ( CPU ) cannot "participate" on any further acceleration of the "last" such block, even when they are free and ready to serve, right because of the said sort of "atomicity"-of-work.

---

SOLUTION:

For more details, feel free to read about this "criticism" and find the added formulae for both the overhead-aware and atomicity-respecting forms of the adapted Amdahl's law .