Jim Meyering wrote:
I also tried with two copies of the output of "seq 9999999" on a
tmpfs file system, with the same result: no discernible difference.
There is something weird going on, as I can't reproduce my earlier
results. Perhaps I built one version of 'diff' without optimization and
the other one with it, by accident. Sorry about sending you down a wild
goose chase.
I'm still seeing a significant performance improvement due to the
change, though not as dramatic as what I earlier reported. Here's the
benchmark:
$ seq 100000000 >0
$ cp 0 1
$ time ./diff-old 0 1
real 0m2.540s
user 0m1.055s
sys 0m1.464s
$ time ./diff-new 0 1
real 0m1.734s
user 0m0.256s
sys 0m1.463s
where 'diff-old' and 'diff-new' are the old
(b6e691277288c4e8d53b1d2577137d265008d13e) and current
(df3af29627a92495a740da13cb8bb0d4fcc1bf84) versions of diffutils, both
compiled with plain 'configure; make' on the same Fedora 20 x86-64
platform I mentioned earlier. This is on an ext4 file system that is
built atop a mirrored hard-disk subsystem, and the locale is en_US.utf8
(dunno if any of this matters).
This benchmark is dominated by system CPU time, so the new version is
only about 45% faster than the old if one looks at real time, but it's
still clearly a win as the user CPU time about 4x faster.
