Hacker News Reader

Comment by joshu | Hacker News Reader

welcome to amdahl's law

Hey what's amdahl's law. Can you explain

The Wikipedia article is pretty good, but tldr: "the overall performance improvement gained by optimizing a single part of a system is limited by the fraction of time that the improved part is actually used."

https://en.wikipedia.org/wiki/Amdahl%27s_law

Usually used in the parallel computing context, but works here too.

So if you have something that takes 1 second out of 10 seconds, you can't improve the overall system more than 1 second (by removing it entirely).

That's not true, because in addition to taking 1 second it could be emptying all your caches. Performance isn't just adding up wallclock times.

Same thing happens when driving - your trip duration usually depends on the slow bits not the faster parts e.g. in NZ driving through small towns will often be the longest part of your trip time.

Something about 5% of things can’t be parallelized which then ends up dominating performance after the 95% of things that can be parallelized are.

In hardware engineering this is called critical path.

That is about latency, but Amdahl's law is about throughput I think.

Amdahl's law is not about latency or throughput, it is about optimizing one part of a program having diminishing returns relative to the part that isn't optimized. It is not really a law, it is more common sense that most people would never assume needs to be explicitly said.

Amdahl's law is common sense, but also the one of the reasons premature optimization is the root of all evil.

A small speedup on the longest (by time) part of the program is usually better than infinite speedup on a short part of the program.

Reply on news.ycombinator.com