Gosh... Lets say you have to purchase X servers at Y dollars, where X*Y>1E8.
How much does an engineer making $200k a year have to reduce X by before he pays for himself? Not much, and simply finding a hotspot and doubling the performance by aligning a data structure or whatnot could pay for his salary for a decade.
The old adage that computers are cheap and engineers are expensive is about 15 years out of date for a large portion of the software being written today. That is when companies stopped being able to generally externalize the costs of shitty code. Plus, as microcontrollers ate the world, the ability to ship a $1 micro controller instead of a $10 one on a product run of 100k units can mean the difference between a competitive growing company and a dying one.
Writing and designing code that is mindful of latencies (Amdahl's law). Temporal / Spatial locality, avoiding needless copies etc.. This approach needs to start from design and also follow into implementing the code. Moreover, this is done without compromising on readability.
Umm, Amdahl's law is actually a negative result. It's saying that if you infinitely improve some magical hotspot that is 1% of your workload, the best you can only ever get to is 99%. If anything, Amdahl is telling us not to grind on trivialities.
If a system performance is profiled as being dominated by 80% A , 19% B and 1 % C. I'd focus on working on A first to get maximum gains. Amdahl's law gives you the backing as to why you should do this.
As a real world example, if an operation involves a network call and you see the RTT dominating the time. You may want to think of ways to avoid the call (caching etc..) if possible to get really good gains.
An upper limit on the benefits. To me, an upper limit is a negative result. When Leon (in Blade Runner) finds out that he only has 4 years to live (an upper bound) he takes it pretty hard.
Amdahl:
A fairly obvious conclusion which can be drawn at this
point is that the effort expended on achieving high
parallel processing rates is wasted unless it is
accompanied by achievements in sequential processing
rates of very nearly the same magnitude.
A point made when we read the paper was that Seymour Cray always made sure that his computers were also the fastest scalar computers even though they were sold as vector processors.
Thank you for the excellent explanation; I will consider it more thoroughly.
I think mathematicians consider an upper limit a positive bound - positive, in the sense of being well defined; you're using negative in the other sense? I actually like that quite a bit.
