perf: add threading capability to perfbdiff
Since we are releasing the GIL during diffing, it is interesting to see how a
thread pool would perform on diffing. We add a new `--threads` argument to
commands. Synchronizing the thread pool is a bit complex because we want to be
able to reuse it from one run to another.
On my computer (i7 with 4 cores + hyperthreading), I get the following data for
about 12000 revisions:
threads wall comb wall gain comb overhead
none 31.596715 31.59 0.00% 0.00%
1 31.621228 31.62 -0.08% 0.09%
2 16.406202 32.8 48.08% 3.83%
3 11.598334 34.76 63.29% 10.03%
4 9.205421 36.77 70.87% 16.40%
5 8.517604 42.51 73.04% 34.57%
6 7.94645 47.58 74.85% 50.62%
7 7.434972 51.92 76.47% 64.36%
8 7.070638 55.34 77.62% 75.18%
Compared to the feature disabled (threads=0), the overhead is negligible with
the threading code (threads=1), and the gain is already 48% with two threads.
#require no-pure
A script to generate nasty diff worst-case scenarios:
$ cat > s.py <<EOF
> import random
> for x in xrange(100000):
> print
> if random.randint(0, 100) >= 50:
> x += 1
> print(hex(x))
> EOF
$ hg init a
$ cd a
Check in a big file:
$ $PYTHON ../s.py > a
$ hg ci -qAm0
Modify it:
$ $PYTHON ../s.py > a
Time a check-in, should never take more than 10 seconds user time:
$ hg ci --time -m1
time: real .* secs .user [0-9][.].* sys .* (re)