HUSTERGS commented on PR #14735:
URL: https://github.com/apache/lucene/pull/14735#issuecomment-2922761475
I added a single block pre-check, so if the target locate at the first
block, we will not pay the double cost. This change seems quite useful, because
it reduces the cost of common cases. And the worse case is also mitigated. The
`AdvanceBenchmark` shows a better result (the test result was worse than the
original implemetation without this check)
```
AdvanceBenchmark.vectorUtilSearch thrpt 15 752.037 ± 33.398 ops/ms (no
expand, current implementation)
AdvanceBenchmark.vectorUtilSearch thrpt 15 625.892 ± 20.849 ops/ms
(expand 2, previous proposed implementation)
AdvanceBenchmark.vectorUtilSearch thrpt 15 802.733 ± 25.096 ops/ms
(expand 2, currently proposed implementation by this PR)
AdvanceBenchmark.vectorUtilSearch thrpt 15 893.295 ± 17.879 ops/ms
(expand 3)
AdvanceBenchmark.vectorUtilSearch thrpt 15 955.528 ± 20.036 ops/ms
(expand 4)
```
the result from luceneutil also confirmed that, `taskCountPerCat` set to 5
and concurrent search disabled:
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
CountAndHighMed 93.29 (3.5%) 90.14
(4.3%) -3.4% ( -10% - 4%) 0.086
CountOrHighMed 97.01 (2.4%) 94.42
(4.3%) -2.7% ( -9% - 4%) 0.125
FilteredAndHighHigh 13.18 (2.2%) 12.94
(1.6%) -1.8% ( -5% - 2%) 0.062
FilteredOrMany 5.10 (1.9%) 5.01
(2.8%) -1.7% ( -6% - 3%) 0.164
Prefix3 98.88 (2.1%) 97.33
(4.5%) -1.6% ( -7% - 5%) 0.371
FilteredPrefix3 92.03 (2.0%) 90.85
(3.9%) -1.3% ( -7% - 4%) 0.408
FilteredOr3Terms 59.68 (1.9%) 59.00
(2.6%) -1.1% ( -5% - 3%) 0.318
SloppyPhrase 1.54 (4.2%) 1.52
(3.3%) -1.1% ( -8% - 6%) 0.579
AndHighOrMedMed 17.85 (2.0%) 17.66
(1.6%) -1.0% ( -4% - 2%) 0.263
FilteredAndStopWords 10.36 (2.1%) 10.25
(1.6%) -1.0% ( -4% - 2%) 0.288
CombinedAndHighMed 24.68 (3.8%) 24.44
(2.9%) -1.0% ( -7% - 5%) 0.554
CombinedAndHighHigh 6.15 (2.7%) 6.10
(1.9%) -0.9% ( -5% - 3%) 0.457
FilteredOrHighHigh 18.16 (1.8%) 18.02
(1.7%) -0.8% ( -4% - 2%) 0.379
FilteredOrHighMed 54.23 (2.9%) 53.82
(2.6%) -0.8% ( -6% - 4%) 0.582
CountOrMany 6.85 (1.7%) 6.80
(3.0%) -0.7% ( -5% - 4%) 0.589
TermMonthSort 2508.14 (3.7%) 2494.67
(3.7%) -0.5% ( -7% - 7%) 0.770
PKLookup 185.70 (8.8%) 185.00
(8.7%) -0.4% ( -16% - 18%) 0.932
TermTitleSort 71.42 (3.2%) 71.20
(3.3%) -0.3% ( -6% - 6%) 0.851
TermDTSort 179.37 (4.6%) 178.92
(2.1%) -0.3% ( -6% - 6%) 0.887
TermDayOfYearSort 375.84 (1.6%) 375.20
(0.7%) -0.2% ( -2% - 2%) 0.786
OrHighRare 123.25 (4.9%) 123.10
(2.1%) -0.1% ( -6% - 7%) 0.948
FilteredOr2Terms2StopWords 69.66 (5.5%) 69.60
(5.6%) -0.1% ( -10% - 11%) 0.977
Wildcard 57.13 (3.1%) 57.12
(3.1%) -0.0% ( -6% - 6%) 0.988
CountPhrase 3.25 (4.4%) 3.25
(2.8%) -0.0% ( -6% - 7%) 0.994
FilteredTerm 88.15 (3.3%) 88.15
(4.2%) 0.0% ( -7% - 7%) 0.999
CountOrHighHigh 63.31 (2.4%) 63.34
(2.9%) 0.0% ( -5% - 5%) 0.976
FilteredOrStopWords 11.01 (3.3%) 11.03
(2.0%) 0.2% ( -4% - 5%) 0.900
FilteredPhrase 12.75 (2.4%) 12.78
(2.6%) 0.2% ( -4% - 5%) 0.880
CombinedTerm 13.47 (3.7%) 13.50
(3.9%) 0.2% ( -7% - 8%) 0.918
IntervalsOrdered 2.95 (3.3%) 2.96
(3.5%) 0.3% ( -6% - 7%) 0.871
Respell 43.81 (1.9%) 43.97
(2.1%) 0.4% ( -3% - 4%) 0.703
CountAndHighHigh 60.80 (2.5%) 61.04
(2.8%) 0.4% ( -4% - 5%) 0.763
SpanNear 3.10 (2.3%) 3.12
(1.9%) 0.6% ( -3% - 4%) 0.552
IntNRQ 48.14 (1.8%) 48.46
(2.4%) 0.7% ( -3% - 4%) 0.522
CountFilteredIntNRQ 22.06 (1.2%) 22.21
(1.5%) 0.7% ( -2% - 3%) 0.306
CountTerm 7031.70 (2.8%) 7085.76
(5.2%) 0.8% ( -7% - 9%) 0.715
CountFilteredOrHighHigh 25.07 (1.1%) 25.27
(1.4%) 0.8% ( -1% - 3%) 0.194
IntSet 391.92 (4.6%) 395.19
(5.4%) 0.8% ( -8% - 11%) 0.739
CountFilteredOrMany 6.01 (1.8%) 6.06
(2.2%) 0.9% ( -3% - 4%) 0.394
FilteredIntNRQ 47.96 (2.1%) 48.37
(2.3%) 0.9% ( -3% - 5%) 0.443
CountFilteredOrHighMed 29.55 (0.8%) 29.81
(1.7%) 0.9% ( -1% - 3%) 0.196
Phrase 9.76 (2.7%) 9.87
(2.5%) 1.1% ( -4% - 6%) 0.414
Fuzzy2 45.67 (3.3%) 46.25
(4.1%) 1.3% ( -5% - 8%) 0.495
CountFilteredPhrase 11.67 (3.4%) 11.82
(3.9%) 1.3% ( -5% - 8%) 0.483
Fuzzy1 50.95 (4.0%) 51.97
(3.5%) 2.0% ( -5% - 9%) 0.286
CombinedOrHighMed 24.10 (9.1%) 24.73
(4.2%) 2.6% ( -9% - 17%) 0.464
CombinedOrHighHigh 5.99 (9.2%) 6.18
(2.6%) 3.1% ( -7% - 16%) 0.355
FilteredAndHighMed 38.19 (8.4%) 39.39
(2.1%) 3.1% ( -6% - 14%) 0.306
AndMedOrHighHigh 19.76 (7.8%) 20.66
(2.3%) 4.6% ( -5% - 15%) 0.112
OrMany 5.44 (8.2%) 5.70
(4.3%) 4.7% ( -7% - 18%) 0.149
FilteredAnd2Terms2StopWords 66.23 (12.4%) 69.62
(5.8%) 5.1% ( -11% - 26%) 0.290
DismaxOrHighMed 63.25 (11.6%) 66.58
(3.5%) 5.3% ( -8% - 22%) 0.219
DismaxTerm 593.96 (8.5%) 625.85
(6.5%) 5.4% ( -8% - 22%) 0.157
And2Terms2StopWords 73.30 (12.9%) 77.34
(8.0%) 5.5% ( -13% - 30%) 0.305
DismaxOrHighHigh 43.52 (12.1%) 46.09
(3.1%) 5.9% ( -8% - 23%) 0.180
TermB1M 526.21 (13.7%) 561.74
(12.2%) 6.8% ( -16% - 37%) 0.298
Term10K 527.47 (13.8%) 563.17
(12.3%) 6.8% ( -16% - 38%) 0.300
FilteredAnd3Terms 80.07 (15.7%) 85.52
(2.0%) 6.8% ( -9% - 29%) 0.224
TermB1M1P 529.07 (13.9%) 565.30
(12.8%) 6.8% ( -17% - 38%) 0.305
AndStopWords 8.52 (14.5%) 9.10
(2.9%) 6.9% ( -9% - 28%) 0.190
Or2Terms2StopWords 71.07 (14.9%) 76.05
(7.5%) 7.0% ( -13% - 34%) 0.235
Term1M 526.93 (13.9%) 565.31
(11.8%) 7.3% ( -16% - 38%) 0.258
Term100 525.57 (13.3%) 563.91
(12.0%) 7.3% ( -15% - 37%) 0.249
And3Terms 78.11 (15.0%) 83.90
(1.8%) 7.4% ( -8% - 28%) 0.167
Term 526.25 (13.5%) 566.72
(12.1%) 7.7% ( -15% - 38%) 0.230
Or3Terms 72.68 (16.4%) 78.71
(1.6%) 8.3% ( -8% - 31%) 0.155
OrStopWords 8.89 (19.6%) 9.76
(3.1%) 9.7% ( -10% - 40%) 0.165
AndHighMed 61.59 (18.7%) 67.79
(2.1%) 10.1% ( -9% - 38%) 0.131
OrHighMed 80.86 (19.6%) 89.47
(2.8%) 10.6% ( -9% - 41%) 0.129
OrHighHigh 22.38 (21.7%) 24.98
(1.6%) 11.6% ( -9% - 44%) 0.131
AndHighHigh 23.37 (21.6%) 26.10
(1.7%) 11.7% ( -9% - 44%) 0.128
```
--
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
