I make no claim as to who first assembled that particular table of data, but Norvig's article is dated 2001.

EDIT: I stand corrected.

The numbers seem to have been evolving and the original source seems to be that page.

http://wayback.archive.org/web/*/http://norvig.com/21-days.h...

This is shameless self promotion (well, me and Carl promotion) but we were measuring these sorts of things in the late 80's and wrote a paper about it that got best paper at Usenix in 95. I think we had most of those numbers, not in the same format.

Pissed off the BSD folks because it made them look bad. Oh, well.

Helped make Linux better, largely because while the BSD guys refused to engage, Linus did. He and I spent many many hours discussing what was the right thing to measure and what should not be measured. We both felt that lmbench would influence OS design (and it's influenced processor design, see all the cache prefetch stuff, I'm pretty convinced that's because all the processor people used lmbench). Linus was already on the "OS should be cheap path" but lmbench helped him make the case to other people who wanted to add overhead because of their pet project.

The cool part about working with Linus was he was never about making Linux look better, he was about measuring the right things. If Linux sucked, oh, well, he'd fix it or get someone else to fix it. Awesome attitude, I feel the same way.

The only published work that might predate lmbench for these sorts of numbers is Hennessy and Patterson computer architecture. They talked about memory latency but so far as I recall, didn't have a benchmark. That said, that book is friggin awesome and anyone who cares about this sort of thing and hasn't carefully read that book is missing out.