[
https://issues.apache.org/jira/browse/LUCENE-9049?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16979422#comment-16979422
]
Jack Conradson commented on LUCENE-9049:
----------------------------------------
[~bruno.roustant] I have uploaded a patch. It passes both precommit and test
locally for me. I used luceneutil to run benchmarks using wikimedium10m for
this change to ensure that direct-addressing did indeed make cachedRootArcs
redundant. From what I can tell the numbers look reasonable --
TaskQPS baseline StdDevQPS candidate StdDev Pct diff
LowTerm 1682.82 (3.9%) 1657.84 (5.3%)
-1.5% ( -10% - 7%)
MedTerm 1339.95 (4.0%) 1321.72 (3.4%)
-1.4% ( -8% - 6%)
Fuzzy2 40.25 (7.4%) 39.80 (7.3%)
-1.1% ( -14% - 14%)
HighTermDayOfYearSort 68.44 (10.7%) 67.87 (10.6%)
-0.8% ( -20% - 22%)
OrNotHighMed 553.72 (5.2%) 550.82 (5.1%)
-0.5% ( -10% - 10%)
OrNotHighLow 670.08 (5.2%) 668.20 (5.2%)
-0.3% ( -10% - 10%)
HighPhrase 383.50 (1.8%) 382.51 (2.1%)
-0.3% ( -4% - 3%)
AndHighLow 748.51 (4.1%) 746.82 (5.0%)
-0.2% ( -8% - 9%)
Prefix3 128.89 (4.3%) 128.66 (4.6%)
-0.2% ( -8% - 9%)
PKLookup 190.04 (1.5%) 189.89 (1.6%)
-0.1% ( -3% - 3%)
Wildcard 125.05 (4.1%) 125.02 (3.9%)
-0.0% ( -7% - 8%)
OrNotHighHigh 642.63 (6.1%) 642.55 (5.3%)
-0.0% ( -10% - 12%)
OrHighNotLow 683.21 (7.1%) 683.29 (6.1%)
0.0% ( -12% - 14%)
HighSpanNear 30.22 (2.5%) 30.23 (2.5%)
0.0% ( -4% - 5%)
BrowseDayOfYearTaxoFacets 10190.78 (2.3%) 10194.55 (1.7%)
0.0% ( -3% - 4%)
AndHighHigh 77.77 (2.0%) 77.81 (1.9%)
0.0% ( -3% - 4%)
BrowseDateTaxoFacets 3.61 (1.4%) 3.62 (1.4%)
0.1% ( -2% - 2%)
OrHighMed 78.14 (2.5%) 78.25 (2.8%)
0.1% ( -4% - 5%)
AndHighMed 206.06 (2.4%) 206.45 (2.6%)
0.2% ( -4% - 5%)
OrHighLow 560.24 (5.1%) 561.72 (5.2%)
0.3% ( -9% - 11%)
MedSpanNear 37.18 (2.2%) 37.31 (1.9%)
0.4% ( -3% - 4%)
IntNRQ 133.13 (1.5%) 133.65 (1.6%)
0.4% ( -2% - 3%)
HighIntervalsOrdered 53.77 (2.0%) 53.98 (1.8%)
0.4% ( -3% - 4%)
LowSloppyPhrase 91.00 (2.8%) 91.36 (3.1%)
0.4% ( -5% - 6%)
HighTerm 1675.67 (5.1%) 1683.33 (4.9%)
0.5% ( -9% - 11%)
BrowseMonthTaxoFacets 10228.66 (2.0%) 10277.31 (1.3%)
0.5% ( -2% - 3%)
OrHighNotHigh 650.24 (6.0%) 653.55 (6.4%)
0.5% ( -11% - 13%)
Respell 93.92 (1.8%) 94.48 (2.1%)
0.6% ( -3% - 4%)
MedSloppyPhrase 41.38 (2.7%) 41.63 (3.0%)
0.6% ( -5% - 6%)
BrowseDayOfYearSSDVFacets 14.25 (7.4%) 14.34 (7.5%)
0.6% ( -13% - 16%)
LowSpanNear 39.78 (2.3%) 40.03 (1.7%)
0.6% ( -3% - 4%)
OrHighHigh 37.28 (2.4%) 37.53 (2.6%)
0.7% ( -4% - 5%)
HighSloppyPhrase 40.95 (4.6%) 41.25 (4.7%)
0.7% ( -8% - 10%)
MedPhrase 68.73 (2.9%) 69.28 (2.8%)
0.8% ( -4% - 6%)
LowPhrase 404.79 (3.0%) 409.52 (2.8%)
1.2% ( -4% - 7%)
OrHighNotMed 674.63 (5.3%) 683.79 (5.7%)
1.4% ( -9% - 13%)
BrowseMonthSSDVFacets 15.97 (5.5%) 16.19 (3.3%)
1.4% ( -6% - 10%)
Fuzzy1 102.56 (11.7%) 104.14 (11.2%)
1.5% ( -19% - 27%)
HighTermMonthSort 209.12 (15.9%) 212.36 (13.5%)
1.6% ( -23% - 36%)
> Remove FST cachedRootArcs now redundant with direct-addressing
> --------------------------------------------------------------
>
> Key: LUCENE-9049
> URL: https://issues.apache.org/jira/browse/LUCENE-9049
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Bruno Roustant
> Priority: Major
> Attachments: LUCENE-9049.patch
>
>
> With LUCENE-8920 FST most often encodes top level nodes with
> direct-addressing (instead of array for binary search). This probably made
> the cachedRootArcs redundant. So they should be removed, and this will reduce
> the code.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
