[
https://issues.apache.org/jira/browse/LUCENE-9613?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17404486#comment-17404486
]
Adrien Grand commented on LUCENE-9613:
--------------------------------------
By removing the wrapping of NumericDocValues by SortedDocValues (see attached
PR) I get even better numbers (the baseline has the above change, so this
speedup is on top of the previous one).
{noformat}
TaskQPS baseline StdDev QPS patch StdDev
Pct diff p-value
LowSloppyPhrase 138.15 (4.5%) 135.56 (4.0%)
-1.9% ( -9% - 6%) 0.163
HighSloppyPhrase 38.34 (5.5%) 37.80 (5.7%)
-1.4% ( -11% - 10%) 0.432
MedSloppyPhrase 68.13 (4.2%) 67.27 (3.6%)
-1.3% ( -8% - 6%) 0.305
MedSpanNear 97.29 (2.4%) 96.13 (2.8%)
-1.2% ( -6% - 4%) 0.150
Respell 198.32 (5.0%) 195.98 (5.0%)
-1.2% ( -10% - 9%) 0.456
HighSpanNear 26.55 (3.7%) 26.24 (6.1%)
-1.2% ( -10% - 8%) 0.462
LowSpanNear 15.53 (2.9%) 15.37 (3.6%)
-1.0% ( -7% - 5%) 0.330
HighIntervalsOrdered 20.91 (6.2%) 20.73 (5.5%)
-0.9% ( -11% - 11%) 0.633
Wildcard 248.72 (13.6%) 246.63 (14.4%)
-0.8% ( -25% - 31%) 0.849
LowIntervalsOrdered 241.69 (6.7%) 239.97 (6.4%)
-0.7% ( -12% - 13%) 0.731
LowPhrase 54.03 (3.3%) 53.67 (2.9%)
-0.7% ( -6% - 5%) 0.495
OrNotHighLow 607.98 (3.2%) 604.28 (3.1%)
-0.6% ( -6% - 5%) 0.542
MedIntervalsOrdered 33.22 (3.4%) 33.03 (2.8%)
-0.6% ( -6% - 5%) 0.562
MedPhrase 292.37 (3.8%) 290.84 (3.5%)
-0.5% ( -7% - 6%) 0.648
HighTermTitleBDVSort 21.76 (2.2%) 21.65 (3.1%)
-0.5% ( -5% - 4%) 0.564
HighTerm 2062.70 (4.0%) 2053.13 (4.2%)
-0.5% ( -8% - 8%) 0.722
OrHighLow 619.26 (2.9%) 616.86 (2.9%)
-0.4% ( -5% - 5%) 0.669
AndHighLow 922.79 (4.8%) 919.30 (4.1%)
-0.4% ( -8% - 8%) 0.788
Prefix3 409.80 (6.5%) 408.28 (6.6%)
-0.4% ( -12% - 13%) 0.857
OrHighNotMed 1354.26 (3.8%) 1349.40 (4.2%)
-0.4% ( -8% - 7%) 0.777
AndHighHigh 55.31 (4.0%) 55.14 (5.0%)
-0.3% ( -8% - 9%) 0.838
IntNRQ 190.47 (1.0%) 189.99 (0.6%)
-0.3% ( -1% - 1%) 0.351
AndHighMed 310.69 (5.0%) 310.04 (5.3%)
-0.2% ( -9% - 10%) 0.898
HighPhrase 210.32 (2.2%) 209.90 (1.9%)
-0.2% ( -4% - 4%) 0.763
TermDTSort 108.34 (3.2%) 108.15 (3.1%)
-0.2% ( -6% - 6%) 0.856
OrNotHighMed 1059.23 (2.8%) 1057.74 (3.4%)
-0.1% ( -6% - 6%) 0.887
OrHighNotLow 919.86 (3.1%) 919.37 (3.2%)
-0.1% ( -6% - 6%) 0.957
MedTerm 2131.16 (3.7%) 2140.13 (4.5%)
0.4% ( -7% - 8%) 0.747
OrNotHighHigh 1217.26 (3.1%) 1222.56 (3.9%)
0.4% ( -6% - 7%) 0.698
HighTermDayOfYearSort 91.07 (7.1%) 91.73 (7.0%)
0.7% ( -12% - 15%) 0.745
OrHighNotHigh 924.82 (3.3%) 931.81 (3.6%)
0.8% ( -5% - 7%) 0.486
Fuzzy1 66.97 (5.9%) 67.57 (7.0%)
0.9% ( -11% - 14%) 0.657
OrHighHigh 26.63 (3.1%) 26.88 (3.5%)
0.9% ( -5% - 7%) 0.373
OrHighMed 100.56 (3.3%) 101.63 (3.4%)
1.1% ( -5% - 8%) 0.315
LowTerm 3005.79 (6.2%) 3044.90 (5.7%)
1.3% ( -9% - 14%) 0.490
Fuzzy2 151.86 (10.1%) 154.03 (9.2%)
1.4% ( -16% - 23%) 0.642
BrowseDateTaxoFacets 3.12 (5.6%) 3.17 (3.8%)
1.8% ( -7% - 11%) 0.235
BrowseMonthTaxoFacets 3.44 (5.3%) 3.50 (4.3%)
1.9% ( -7% - 12%) 0.211
BrowseDayOfYearTaxoFacets 3.12 (5.6%) 3.18 (4.0%)
2.0% ( -7% - 12%) 0.202
HighTermMonthSort 70.56 (9.8%) 74.99 (10.5%)
6.3% ( -12% - 29%) 0.051
BrowseMonthSSDVFacets 14.44 (4.9%) 18.97 (34.9%)
31.4% ( -8% - 74%) 0.000
BrowseDayOfYearSSDVFacets 14.88 (7.2%) 19.77 (32.7%)
32.9% ( -6% - 78%) 0.000
{noformat}
The change might be a bit more controversial given that it requires checking
some of the numeric optimizations, which is why I didn't push it right away.
> Create blocks for ords when it helps in Lucene80DocValuesFormat
> ---------------------------------------------------------------
>
> Key: LUCENE-9613
> URL: https://issues.apache.org/jira/browse/LUCENE-9613
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Adrien Grand
> Priority: Minor
> Fix For: main (9.0)
>
> Time Spent: 1h 10m
> Remaining Estimate: 0h
>
> Currently for sorted(-set) values, we always write ords using
> log2(valueCount) bits per entry. However in several cases like when the field
> is used in the index sort, or if one value is _very_common, splitting into
> blocks like we do for numerics would help.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
