[
https://issues.apache.org/jira/browse/LUCENE-9254?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17049266#comment-17049266
]
Bruno Roustant commented on LUCENE-9254:
----------------------------------------
Benchmark to compare UniformSplit FST on-heap to UniformSplit FST off-heap
(wikimediumall).
Conclusion: no noticeable difference in query throughput.
TaskQPS UniformSplitOnHeap StdDevQPS UniformSplitOffHeap StdDev
Pct diff
HighIntervalsOrdered 14.54 (2.8%) 13.81 (3.9%)
-5.0% ( -11% - 1%)
HighTermDayOfYearSort 24.38 (12.3%) 23.87 (6.8%)
-2.1% ( -18% - 19%)
BrowseMonthTaxoFacets 1.07 (2.5%) 1.05 (5.6%)
-1.8% ( -9% - 6%)
BrowseDateTaxoFacets 0.99 (2.4%) 0.97 (5.9%)
-1.7% ( -9% - 6%)
BrowseDayOfYearTaxoFacets 0.98 (2.3%) 0.97 (5.9%)
-1.6% ( -9% - 6%)
OrHighMed 59.94 (2.4%) 59.16 (3.8%)
-1.3% ( -7% - 5%)
HighSpanNear 6.59 (2.4%) 6.51 (3.8%)
-1.1% ( -7% - 5%)
AndHighHigh 29.17 (4.3%) 28.91 (4.2%)
-0.9% ( -8% - 7%)
OrHighHigh 8.92 (2.2%) 8.85 (3.5%)
-0.9% ( -6% - 5%)
MedSpanNear 18.14 (2.5%) 18.00 (3.6%)
-0.8% ( -6% - 5%)
HighSloppyPhrase 6.49 (5.0%) 6.45 (4.6%)
-0.7% ( -9% - 9%)
MedSloppyPhrase 8.89 (4.6%) 8.82 (4.3%)
-0.7% ( -9% - 8%)
LowSloppyPhrase 43.10 (4.4%) 42.79 (4.3%)
-0.7% ( -9% - 8%)
Respell 21.36 (3.2%) 21.23 (3.7%)
-0.6% ( -7% - 6%)
BrowseDayOfYearSSDVFacets 3.90 (2.5%) 3.87 (3.4%)
-0.6% ( -6% - 5%)
BrowseMonthSSDVFacets 4.36 (2.5%) 4.33 (3.5%)
-0.5% ( -6% - 5%)
AndHighMed 102.06 (4.6%) 101.57 (5.0%)
-0.5% ( -9% - 9%)
Fuzzy1 34.33 (4.6%) 34.20 (5.5%)
-0.4% ( -10% - 10%)
AndHighLow 426.06 (4.3%) 424.57 (6.1%)
-0.4% ( -10% - 10%)
OrHighLow 274.31 (4.7%) 273.36 (4.1%)
-0.3% ( -8% - 8%)
LowSpanNear 101.07 (2.6%) 100.81 (3.8%)
-0.3% ( -6% - 6%)
LowPhrase 23.46 (4.0%) 23.44 (3.8%)
-0.1% ( -7% - 7%)
PKLookup 176.41 (4.0%) 176.83 (4.0%)
0.2% ( -7% - 8%)
IntNRQ 24.21 (21.1%) 24.27 (21.8%)
0.3% ( -35% - 54%)
OrNotHighLow 449.31 (3.3%) 450.87 (6.2%)
0.3% ( -8% - 10%)
Wildcard 73.49 (5.6%) 73.76 (4.7%)
0.4% ( -9% - 11%)
Fuzzy2 27.01 (3.4%) 27.17 (5.8%)
0.6% ( -8% - 10%)
OrHighNotHigh 571.05 (7.4%) 575.98 (6.9%)
0.9% ( -12% - 16%)
HighTerm 1240.74 (6.2%) 1252.63 (4.2%)
1.0% ( -8% - 12%)
OrNotHighMed 470.36 (6.3%) 475.41 (7.7%)
1.1% ( -12% - 16%)
MedPhrase 120.67 (5.1%) 121.97 (5.1%)
1.1% ( -8% - 11%)
OrNotHighHigh 649.46 (5.9%) 656.70 (7.0%)
1.1% ( -11% - 14%)
OrHighNotMed 525.03 (6.6%) 531.55 (7.7%)
1.2% ( -12% - 16%)
OrHighNotLow 675.24 (8.9%) 687.58 (8.3%)
1.8% ( -14% - 20%)
Prefix3 86.14 (6.8%) 87.81 (7.4%)
1.9% ( -11% - 17%)
MedTerm 1187.58 (5.8%) 1213.76 (5.9%)
2.2% ( -8% - 14%)
HighPhrase 224.41 (3.7%) 229.86 (5.7%)
2.4% ( -6% - 12%)
LowTerm 1336.79 (5.8%) 1375.16 (5.7%)
2.9% ( -8% - 15%)
HighTermMonthSort 67.74 (15.9%) 76.73 (12.2%)
13.3% ( -12% - 49%)
> FST off-heap support for UniformSplit
> -------------------------------------
>
> Key: LUCENE-9254
> URL: https://issues.apache.org/jira/browse/LUCENE-9254
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Bruno Roustant
> Assignee: Bruno Roustant
> Priority: Major
> Time Spent: 40m
> Remaining Estimate: 0h
>
> UniformSplit should support off-heap FST the same way BlockTree postings
> format does, with same AUTO mode, and with same ability to force on/off heap
> with reader attributes.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
