Add a statically dispatched version of make_comparator

[adriangb]

Working on apache/datafusion#18449 it came up that the dynamic dispatch here can hurt performance for simple types. This version seems to perform better in benchmarks there (admittedly run alongside other changes).

[Dandandan]

Would be nice to see the sorting benchmarks on this change! My feeling is it might significantly reduce the overhead of the lexical sorting kernels (compared to row format).

[adriangb]

Or it might make it worse… it’s a trade off of indirection vs branch prediction. As we found out in apache/datafusion#18449 the same change can be 70% faster on ARM MacBooks and 70% slower on x86.

So needless to say this should not be merged without benchmarks

[adriangb]

I added benchmarks and comparing to main this seems overall slower:

make_comparator i32 100 rows
                        time:   [3.8138 µs 3.8190 µs 3.8246 µs]
                        change: [+1.2144% +1.3230% +1.4364%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 6 outliers among 100 measurements (6.00%)
  5 (5.00%) high mild
  1 (1.00%) high severe

make_comparator i32 1M rows
                        time:   [3.8120 µs 3.8133 µs 3.8147 µs]
                        change: [+0.8398% +0.9598% +1.0570%] (p = 0.00 < 0.05)
                        Change within noise threshold.
Found 5 outliers among 100 measurements (5.00%)
  2 (2.00%) low mild
  3 (3.00%) high mild

make_comparator i32 nulls 50% 100 rows
                        time:   [4.7215 µs 4.7240 µs 4.7272 µs]
                        change: [+15.169% +15.252% +15.353%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 13 outliers among 100 measurements (13.00%)
  4 (4.00%) high mild
  9 (9.00%) high severe

make_comparator i32 nulls 100% 100 rows
                        time:   [4.7141 µs 4.7168 µs 4.7204 µs]
                        change: [+18.554% +18.690% +18.819%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 12 outliers among 100 measurements (12.00%)
  4 (4.00%) high mild
  8 (8.00%) high severe

make_comparator f64 100 rows
                        time:   [4.2620 µs 4.2640 µs 4.2664 µs]
                        change: [+4.1768% +4.3310% +4.4556%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 11 outliers among 100 measurements (11.00%)
  1 (1.00%) low mild
  1 (1.00%) high mild
  9 (9.00%) high severe

make_comparator f64 1M rows
                        time:   [4.2628 µs 4.2637 µs 4.2647 µs]
                        change: [+5.1286% +5.3869% +5.6529%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 11 outliers among 100 measurements (11.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild
  8 (8.00%) high severe

make_comparator utf8 100 rows
                        time:   [7.0815 µs 7.0846 µs 7.0880 µs]
                        change: [+13.332% +13.648% +13.927%] (p = 0.00 < 0.05)
                        Performance has regressed.

make_comparator utf8 1M rows
                        time:   [7.6914 µs 7.6936 µs 7.6959 µs]
                        change: [+8.4761% +8.8173% +9.1074%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 10 outliers among 100 measurements (10.00%)
  9 (9.00%) high mild
  1 (1.00%) high severe

make_comparator utf8view 100 rows
                        time:   [8.0346 µs 8.0509 µs 8.0688 µs]
                        change: [+59.938% +60.188% +60.453%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 14 outliers among 100 measurements (14.00%)
  4 (4.00%) high mild
  10 (10.00%) high severe

make_comparator utf8view 1M rows
                        time:   [10.060 µs 10.066 µs 10.073 µs]
                        change: [+60.032% +60.127% +60.223%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 4 outliers among 100 measurements (4.00%)
  4 (4.00%) high mild

make_comparator dict(u32,utf8) 100 rows
                        time:   [11.618 µs 11.623 µs 11.628 µs]
                        change: [+43.742% +43.911% +44.091%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 14 outliers among 100 measurements (14.00%)
  2 (2.00%) high mild
  12 (12.00%) high severe

make_comparator dict(u32,utf8) 1M rows
                        time:   [11.899 µs 11.907 µs 11.918 µs]
                        change: [+51.329% +51.453% +51.603%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 10 outliers among 100 measurements (10.00%)
  10 (10.00%) high severe

make_comparator list(i32) 100 rows
                        time:   [11.389 µs 11.400 µs 11.415 µs]
                        change: [+69.504% +70.082% +70.679%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 16 outliers among 100 measurements (16.00%)
  4 (4.00%) high mild
  12 (12.00%) high severe

make_comparator list(i32) 1M rows
                        time:   [11.955 µs 11.969 µs 11.981 µs]
                        change: [+68.186% +68.422% +68.643%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 21 outliers among 100 measurements (21.00%)
  21 (21.00%) high mild

So I'm going to close this for now as a failed experiment.