Before this change, we would first find all changes so we could obtain
the largest offset we will report and use that to set up the padding.
Now we use the file sizes to estimate the largest possible offset.
Not only does this allow us to print earlier, reduces memory usage, as
we do not store diffs to report later, but it also fixes a case in
which our output was different to GNU cmp's - because it also seems
to estimate based on size.
Memory usage drops by a factor of 1000(!), without losing performance
while comparing 2 binaries of hundreds of MBs:
Before:
Maximum resident set size (kbytes): 2489260
Benchmark 1: ../target/release/diffutils \
cmp -l -b /usr/lib64/chromium-browser/chromium-browser /usr/lib64/firefox/libxul.so
Time (mean ± σ): 14.466 s ± 0.166 s [User: 12.367 s, System: 2.012 s]
Range (min … max): 14.350 s … 14.914 s 10 runs
After:
Maximum resident set size (kbytes): 2636
Benchmark 1: ../target/release/diffutils \
cmp -l -b /usr/lib64/chromium-browser/chromium-browser /usr/lib64/firefox/libxul.so
Time (mean ± σ): 13.724 s ± 0.038 s [User: 12.263 s, System: 1.372 s]
Range (min … max): 13.667 s … 13.793 s 10 runs
This makes the code less readable, but gets us a massive improvement
to performance. Comparing ~36M completely different files now takes
~40% of the time. Compared to GNU cmp, we now run the same comparison
in ~26% of the time.
This also improves comparing binary files. A comparison of chromium
and libxul now takes ~60% of the time. We also beat GNU cmpi by about
the same margin.
Before:
> hyperfine --warmup 1 -i --output=pipe \
'../target/release/diffutils cmp -l huge huge.3'
Benchmark 1: ../target/release/diffutils cmp -l huge huge.3
Time (mean ± σ): 2.000 s ± 0.016 s [User: 1.603 s, System: 0.392 s]
Range (min … max): 1.989 s … 2.043 s 10 runs
Warning: Ignoring non-zero exit code.
> hyperfine --warmup 1 -i --output=pipe \
'../target/release/diffutils cmp -l -b \
/usr/lib64/chromium-browser/chromium-browser \
/usr/lib64/firefox/libxul.so'
Benchmark 1: ../target/release/diffutils cmp -l -b /usr/lib64/chromium-browser/chromium-browser /usr/lib64/firefox/libxul.so
Time (mean ± σ): 24.704 s ± 0.162 s [User: 21.948 s, System: 2.700 s]
Range (min … max): 24.359 s … 24.889 s 10 runs
Warning: Ignoring non-zero exit code.
After:
> hyperfine --warmup 1 -i --output=pipe \
'../target/release/diffutils cmp -l huge huge.3'
Benchmark 1: ../target/release/diffutils cmp -l huge huge.3
Time (mean ± σ): 849.5 ms ± 6.2 ms [User: 538.3 ms, System: 306.8 ms]
Range (min … max): 839.4 ms … 857.7 ms 10 runs
Warning: Ignoring non-zero exit code.
> hyperfine --warmup 1 -i --output=pipe \
'../target/release/diffutils cmp -l -b \
/usr/lib64/chromium-browser/chromium-browser \
/usr/lib64/firefox/libxul.so'
Benchmark 1: ../target/release/diffutils cmp -l -b /usr/lib64/chromium-browser/chromium-browser /usr/lib64/firefox/libxul.so
Time (mean ± σ): 14.646 s ± 0.040 s [User: 12.328 s, System: 2.286 s]
Range (min … max): 14.585 s … 14.702 s 10 runs
Warning: Ignoring non-zero exit code.
Octal conversion and simple integer to string both show up in profiling.
This change improves comparing ~36M completely different files wth both
-l and -b by ~11-13%.
The utility should support all the arguments supported by GNU cmp and
perform slightly better.
On a "bad" scenario, ~36M files which are completely different, our
version runs in ~72% of the time of the original on my M1 Max:
> hyperfine --warmup 1 -i --output=pipe \
'cmp -l huge huge.3'
Benchmark 1: cmp -l huge huge.3
Time (mean ± σ): 3.237 s ± 0.014 s [User: 2.891 s, System: 0.341 s]
Range (min … max): 3.221 s … 3.271 s 10 runs
Warning: Ignoring non-zero exit code.
> hyperfine --warmup 1 -i --output=pipe \
'../target/release/diffutils cmp -l huge huge.3'
Benchmark 1: ../target/release/diffutils cmp -l huge huge.3
Time (mean ± σ): 2.392 s ± 0.009 s [User: 1.978 s, System: 0.406 s]
Range (min … max): 2.378 s … 2.406 s 10 runs
Warning: Ignoring non-zero exit code.
Our cmp runs in ~116% of the time when comparing libxul.so to the
chromium-browser binary with -l and -b. In a best case scenario of
comparing 2 files which are the same except for the last byte, our
tool is slightly faster.
Daniel Hofstetter
renovate[bot]
Olivier Tilloy
Sylvestre Ledru
Gustavo Noronha Silva