HUSTERGS commented on PR #15081:
URL: https://github.com/apache/lucene/pull/15081#issuecomment-3197362760
Actually I did another experiment, I was worried about the `findNextGEQ`
operation might cause a unexpected slowdown on those machine that do not have
fast SIMD instructions, so I added a new `copyWithMinDocRequired` method to
`DocAndScoreAccBuffer`, now instead of call `findNextGEQ`, I directly filter
those docs when copy, here is the result:
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
CombinedOrHighHigh 5.54 (3.2%) 5.41
(3.6%) -2.5% ( -9% - 4%) 0.023
TermTitleSort 51.50 (3.9%) 50.60
(4.1%) -1.7% ( -9% - 6%) 0.169
CombinedAndHighHigh 5.60 (2.2%) 5.53
(1.3%) -1.2% ( -4% - 2%) 0.032
DismaxOrHighMed 45.35 (7.4%) 44.93
(9.2%) -0.9% ( -16% - 16%) 0.726
SpanNear 2.51 (4.1%) 2.50
(4.1%) -0.4% ( -8% - 8%) 0.733
CombinedTerm 10.96 (4.0%) 10.91
(4.8%) -0.4% ( -8% - 8%) 0.761
AndHighOrMedMed 13.50 (2.6%) 13.47
(3.0%) -0.2% ( -5% - 5%) 0.789
FilteredAndHighHigh 10.23 (2.9%) 10.24
(2.7%) 0.0% ( -5% - 5%) 0.959
Term10K 438.89 (6.1%) 439.16
(6.6%) 0.1% ( -11% - 13%) 0.975
Term 439.07 (6.0%) 439.36
(6.6%) 0.1% ( -11% - 13%) 0.974
TermB1M 439.19 (5.9%) 439.51
(6.6%) 0.1% ( -11% - 13%) 0.971
SloppyPhrase 1.12 (3.3%) 1.12
(4.1%) 0.1% ( -7% - 7%) 0.929
CombinedOrHighMed 19.36 (5.3%) 19.40
(5.7%) 0.2% ( -10% - 11%) 0.926
FilteredAndStopWords 8.22 (3.0%) 8.23
(2.5%) 0.2% ( -5% - 5%) 0.849
Term100 438.48 (6.0%) 439.22
(6.6%) 0.2% ( -11% - 13%) 0.933
TermB1M1P 438.65 (6.0%) 439.45
(6.7%) 0.2% ( -11% - 13%) 0.927
Term1M 438.40 (6.1%) 439.36
(6.5%) 0.2% ( -11% - 13%) 0.913
CountPhrase 2.64 (3.2%) 2.64
(3.8%) 0.2% ( -6% - 7%) 0.836
CountFilteredOrHighMed 17.85 (0.7%) 17.90
(0.7%) 0.3% ( -1% - 1%) 0.186
CountFilteredOrHighHigh 15.76 (0.7%) 15.81
(0.9%) 0.3% ( -1% - 1%) 0.173
DismaxTerm 473.34 (4.1%) 475.20
(4.9%) 0.4% ( -8% - 9%) 0.782
OrHighRare 93.56 (3.0%) 93.93
(3.3%) 0.4% ( -5% - 6%) 0.683
CountFilteredIntNRQ 16.34 (1.2%) 16.42
(1.2%) 0.5% ( -1% - 2%) 0.193
IntSet 286.37 (4.4%) 287.83
(4.7%) 0.5% ( -8% - 9%) 0.723
FilteredPrefix3 68.84 (4.4%) 69.19
(3.6%) 0.5% ( -7% - 8%) 0.682
CountAndHighMed 71.34 (2.8%) 71.73
(2.6%) 0.6% ( -4% - 6%) 0.516
Wildcard 46.52 (2.5%) 46.78
(2.6%) 0.6% ( -4% - 5%) 0.476
IntervalsOrdered 2.43 (2.9%) 2.44
(3.1%) 0.6% ( -5% - 6%) 0.528
FilteredOrStopWords 7.87 (2.0%) 7.93
(2.1%) 0.7% ( -3% - 4%) 0.251
Prefix3 73.58 (4.8%) 74.13
(3.7%) 0.8% ( -7% - 9%) 0.578
CountFilteredOrMany 4.29 (2.5%) 4.33
(2.6%) 0.9% ( -4% - 6%) 0.278
CountAndHighHigh 48.27 (1.6%) 48.70
(1.5%) 0.9% ( -2% - 4%) 0.072
CountOrHighHigh 49.25 (1.9%) 49.71
(1.8%) 0.9% ( -2% - 4%) 0.101
OrStopWords 8.91 (8.4%) 9.00
(12.2%) 1.0% ( -18% - 23%) 0.771
TermDayOfYearSort 252.19 (1.9%) 254.74
(1.9%) 1.0% ( -2% - 4%) 0.093
Phrase 7.36 (2.4%) 7.43
(2.1%) 1.0% ( -3% - 5%) 0.145
FilteredTerm 60.82 (2.3%) 61.48
(2.6%) 1.1% ( -3% - 6%) 0.165
CountOrHighMed 71.77 (2.2%) 72.56
(1.9%) 1.1% ( -2% - 5%) 0.094
FilteredPhrase 9.30 (2.9%) 9.41
(2.7%) 1.2% ( -4% - 6%) 0.188
CombinedAndHighMed 19.58 (5.1%) 19.81
(5.0%) 1.2% ( -8% - 11%) 0.460
CountOrMany 4.81 (2.3%) 4.87
(2.6%) 1.2% ( -3% - 6%) 0.132
OrMany 4.17 (6.2%) 4.22
(5.3%) 1.2% ( -9% - 13%) 0.509
FilteredOrMany 3.87 (3.0%) 3.92
(2.9%) 1.2% ( -4% - 7%) 0.196
IntNRQ 42.46 (2.9%) 42.99
(3.0%) 1.3% ( -4% - 7%) 0.179
FilteredOrHighHigh 12.37 (2.8%) 12.54
(2.9%) 1.3% ( -4% - 7%) 0.134
Respell 34.57 (2.7%) 35.05
(3.5%) 1.4% ( -4% - 7%) 0.162
OrHighHigh 20.69 (10.5%) 20.98
(14.8%) 1.4% ( -21% - 29%) 0.730
CountFilteredPhrase 8.57 (3.3%) 8.69
(3.3%) 1.4% ( -5% - 8%) 0.181
TermMonthSort 2048.69 (2.8%) 2078.78
(1.8%) 1.5% ( -3% - 6%) 0.049
CountTerm 5432.73 (2.5%) 5516.44
(2.7%) 1.5% ( -3% - 6%) 0.063
Fuzzy2 33.85 (3.7%) 34.38
(3.7%) 1.6% ( -5% - 9%) 0.180
FilteredIntNRQ 42.05 (2.9%) 42.75
(3.1%) 1.7% ( -4% - 7%) 0.080
DismaxOrHighHigh 33.48 (5.4%) 34.03
(7.0%) 1.7% ( -10% - 14%) 0.400
FilteredAnd3Terms 99.67 (3.3%) 101.33
(3.2%) 1.7% ( -4% - 8%) 0.104
TermDTSort 136.90 (3.5%) 139.43
(4.0%) 1.9% ( -5% - 9%) 0.118
FilteredOrHighMed 36.11 (4.1%) 36.80
(4.3%) 1.9% ( -6% - 10%) 0.146
Fuzzy1 37.33 (4.3%) 38.06
(4.1%) 2.0% ( -6% - 10%) 0.136
FilteredOr3Terms 41.00 (3.8%) 41.84
(4.2%) 2.0% ( -5% - 10%) 0.106
FilteredAndHighMed 30.09 (4.2%) 30.78
(3.6%) 2.3% ( -5% - 10%) 0.061
FilteredOr2Terms2StopWords 45.64 (4.9%) 46.83
(5.3%) 2.6% ( -7% - 13%) 0.105
Or3Terms 59.85 (9.7%) 61.50
(12.9%) 2.8% ( -18% - 28%) 0.444
OrHighMed 60.34 (11.4%) 62.06
(15.6%) 2.9% ( -21% - 33%) 0.507
Or2Terms2StopWords 53.33 (8.4%) 55.96
(10.6%) 4.9% ( -12% - 26%) 0.103
FilteredAnd2Terms2StopWords 54.99 (5.3%) 58.02
(5.9%) 5.5% ( -5% - 17%) 0.002
AndHighHigh 21.28 (10.6%) 22.68
(10.7%) 6.6% ( -13% - 31%) 0.051
And3Terms 65.77 (9.0%) 70.21
(8.8%) 6.8% ( -10% - 27%) 0.017
AndHighMed 49.53 (10.6%) 52.99
(10.4%) 7.0% ( -12% - 31%) 0.036
And2Terms2StopWords 51.74 (8.1%) 55.41
(8.8%) 7.1% ( -9% - 26%) 0.008
AndMedOrHighHigh 14.90 (4.2%) 16.03
(4.0%) 7.6% ( 0% - 16%) 0.000
AndStopWords 8.18 (6.7%) 8.86
(7.4%) 8.3% ( -5% - 24%) 0.000
```
BTW, This benchmark contains #15039, which is different from previous
benchmark setup. The affected tasks also changed, e.g. OrHighHigh no longer got
a speedup.
I'd push the newest code latter, maybe you can help run the benchmark to
verify it ?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
