Re: [PATCH] (x)memcoll: performance improvement when input is known to be NUL delimited.

[Bruno Haible]

Hi Chen,

> ... it was consistently (as in every run for 20 runs) faster, and thus I've
> included it in the patch. 

Thanks. I'll wait for the FSF notification that the legal papers have arrived.

> This is not what I expected at all, and I'm having a hard time coming up with
> a reason why this is

15 years ago, on a Linux/x86 machine, I could measure CPU performance with
about 1% precision (with 3 or 5 runs of a program). At the same time, on
Windows, I got only about 5% precision.

On SPARC CPUs, sometimes the same program was 30% slower than the previous day.
So, such platforms were unusable for benchmarking.

Today, x86 CPUs have additional complexity: multiple CPU cores that fight over
the memory bus; hyperthreads. The Linux memory management has become more
complex as well (the kernel actively swapping out some pages). Sometimes you're
even running inside a virtual machine. I usually can get only about 2% of
precision nowadays.

Bruno