cut: optimizations

[cbjadwani]

• Use buffered stdout to reduce write sys calls.

  This simple change yielded the biggest performance gain.

• Use for_byte_record_with_terminator from the bstr crate.

  This is to minimize the per line copying needed by BufReader::read_until. The cut_fields and cut_fields_delimiter functions used read_until to iterate over lines. That required copying each input line to the line buffer. With for_byte_record_with_terminator copying is minimized as it calls our closure with a reference to BufReader's buffer most of the time. It needs to copy (internally) only to process any incomplete lines at the end of the buffer.

• Re-write Searcher to use memchr.

  Switch from the naive implementation to one that uses memchr.

• Rewrite cut_bytes almost entirely.

  This was already well optimized. The performance gain in this case is not from avoiding copying. In fact, it needed zero copying whereas new implementation introduces some copying similar to cut_fields described above. But the occasional copying cost is more than offset by the use of the very fast memchr inside for_byte_record_with_terminator. This change also simplifies the code significantly. Removed the buffer module.

Benchmarks

• $ hyperfine -w3 "./target/release/cut -c2-4,8 test-large.txt" "cut -c2-4,8 test-large.txt"

Before

Benchmark #1: ./target/release/cut -c2-4,8 test-large.txt
  Time (mean ± σ):     159.6 ms ±   2.4 ms    [User: 112.5 ms, System: 46.9 ms]
  Range (min … max):   156.3 ms … 163.7 ms    18 runs
 
Benchmark #2: cut -c2-4,8 test-large.txt
  Time (mean ± σ):      29.5 ms ±   1.4 ms    [User: 26.7 ms, System: 2.8 ms]
  Range (min … max):    28.3 ms …  34.2 ms    99 runs
 
Summary
  'cut -c2-4,8 test-large.txt' ran
    5.41 ± 0.27 times faster than './target/release/cut -c2-4,8 test-large.txt'

After

Benchmark #1: ./target/release/cut -c2-4,8 test-large.txt
  Time (mean ± σ):      13.9 ms ±   1.4 ms    [User: 11.0 ms, System: 2.9 ms]
  Range (min … max):    12.7 ms …  19.4 ms    208 runs
 
Benchmark #2: cut -c2-4,8 test-large.txt
  Time (mean ± σ):      29.8 ms ±   1.6 ms    [User: 26.8 ms, System: 3.0 ms]
  Range (min … max):    28.1 ms …  34.1 ms    99 runs
 
Summary
  './target/release/cut -c2-4,8 test-large.txt' ran
    2.14 ± 0.25 times faster than 'cut -c2-4,8 test-large.txt'

• $ hyperfine -w3 "./target/release/cut -f2-4,8 -d' ' test-large.txt" "cut -f2-4,8 -d' ' test-large.txt"

Before

Benchmark #1: ./target/release/cut -f2-4,8 -d' ' test-large.txt
  Time (mean ± σ):     165.4 ms ±   2.4 ms    [User: 116.6 ms, System: 48.6 ms]
  Range (min … max):   161.6 ms … 170.6 ms    17 runs
 
Benchmark #2: cut -f2-4,8 -d' ' test-large.txt
  Time (mean ± σ):      64.0 ms ±   1.5 ms    [User: 61.1 ms, System: 2.8 ms]
  Range (min … max):    62.2 ms …  68.8 ms    44 runs
 
Summary
  'cut -f2-4,8 -d' ' test-large.txt' ran
    2.59 ± 0.07 times faster than './target/release/cut -f2-4,8 -d' ' test-large.txt'

After

Benchmark #1: ./target/release/cut -f2-4,8 -d' ' test-large.txt
  Time (mean ± σ):      26.8 ms ±   1.6 ms    [User: 23.7 ms, System: 3.1 ms]
  Range (min … max):    24.9 ms …  31.1 ms    103 runs
 
Benchmark #2: cut -f2-4,8 -d' ' test-large.txt
  Time (mean ± σ):      63.9 ms ±   1.6 ms    [User: 60.5 ms, System: 3.4 ms]
  Range (min … max):    62.3 ms …  69.1 ms    46 runs
 
Summary
  './target/release/cut -f2-4,8 -d' ' test-large.txt' ran
    2.38 ± 0.16 times faster than 'cut -f2-4,8 -d' ' test-large.txt'

• hyperfine -w3 "./target/release/cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt" "cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt"

Before

Benchmark #1: ./target/release/cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt
  Time (mean ± σ):     190.6 ms ±   6.1 ms    [User: 140.5 ms, System: 49.9 ms]
  Range (min … max):   185.5 ms … 210.5 ms    15 runs
 
Benchmark #2: cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt
  Time (mean ± σ):      63.8 ms ±   1.8 ms    [User: 60.9 ms, System: 2.8 ms]
  Range (min … max):    62.2 ms …  69.5 ms    46 runs
 
Summary
  'cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt' ran
    2.99 ± 0.13 times faster than './target/release/cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt'

After

Benchmark #1: ./target/release/cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt
  Time (mean ± σ):      34.2 ms ±   1.8 ms    [User: 31.4 ms, System: 2.8 ms]
  Range (min … max):    32.2 ms …  39.1 ms    87 runs
 
Benchmark #2: cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt
  Time (mean ± σ):      63.4 ms ±   1.3 ms    [User: 60.6 ms, System: 2.8 ms]
  Range (min … max):    62.3 ms …  68.0 ms    46 runs
 
Summary
  './target/release/cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt' ran
    1.85 ± 0.10 times faster than 'cut -f2-4,8 -d' ' --output-delimiter=- test-large.txt'

[sylvestre]

Well done for the huge wins!

Could you please create a benchmark files like:
https://github.com/uutils/coreutils/blob/master/src/uu/sort/BENCHMARKING.md
to be able to store way to benchmark this program? thanks

[cbjadwani]

Added benchmarking doc. I think it would be more accessible if instead of adding all details for each utility separately common details can be extracted and kept at project root level.

[sylvestre]

Great doc, thanks