[
https://issues.apache.org/jira/browse/LUCENE-9125?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17017978#comment-17017978
]
Bruno Roustant commented on LUCENE-9125:
----------------------------------------
In the benchmark above I used by error wikimedium10k (I edited to mention that).
Here is the benchmark for wikimediumall:
Task QPS trunk StdDev QPS patch StdDev
Pct diff
OrHighNotHigh 769.84 (4.8%) 756.84 (5.0%)
-1.7% ( -10% - 8%)
OrNotHighLow 664.03 (4.2%) 653.64 (3.4%)
-1.6% ( -8% - 6%)
OrNotHighMed 574.56 (3.0%) 566.90 (2.5%)
-1.3% ( -6% - 4%)
MedTerm 1373.80 (3.9%) 1359.30 (5.1%)
-1.1% ( -9% - 8%)
AndHighHigh 19.84 (3.6%) 19.67 (2.9%)
-0.9% ( -7% - 5%)
AndHighLow 474.49 (2.9%) 470.36 (3.6%)
-0.9% ( -7% - 5%)
Fuzzy1 69.27 (10.7%) 68.75 (11.0%)
-0.7% ( -20% - 23%)
OrNotHighHigh 569.30 (3.4%) 565.26 (5.0%)
-0.7% ( -8% - 7%)
MedPhrase 36.97 (2.4%) 36.76 (2.7%)
-0.6% ( -5% - 4%)
HighTerm 1133.65 (4.2%) 1128.30 (4.3%)
-0.5% ( -8% - 8%)
OrHighLow 227.08 (2.9%) 226.24 (3.3%)
-0.4% ( -6% - 6%)
OrHighHigh 24.17 (2.6%) 24.08 (2.4%)
-0.4% ( -5% - 4%)
Prefix3 25.30 (3.8%) 25.22 (3.7%)
-0.3% ( -7% - 7%)
OrHighMed 48.26 (3.1%) 48.11 (3.1%)
-0.3% ( -6% - 6%)
LowTerm 1087.75 (3.4%) 1084.44 (3.3%)
-0.3% ( -6% - 6%)
AndHighMed 69.62 (3.9%) 69.44 (4.1%)
-0.3% ( -7% - 7%)
HighSloppyPhrase 15.11 (2.6%) 15.08 (2.6%)
-0.2% ( -5% - 5%)
Respell 43.34 (2.0%) 43.28 (2.3%)
-0.1% ( -4% - 4%)
OrHighNotLow 666.79 (3.4%) 665.98 (4.9%)
-0.1% ( -8% - 8%)
HighSpanNear 8.21 (1.8%) 8.20 (2.0%)
-0.1% ( -3% - 3%)
HighIntervalsOrdered 14.46 (1.2%) 14.45 (1.4%)
-0.1% ( -2% - 2%)
HighPhrase 333.99 (3.3%) 333.74 (3.9%)
-0.1% ( -7% - 7%)
MedSpanNear 12.08 (1.8%) 12.07 (2.0%)
-0.1% ( -3% - 3%)
LowPhrase 481.10 (2.5%) 481.14 (3.4%)
0.0% ( -5% - 6%)
MedSloppyPhrase 6.78 (2.9%) 6.78 (2.9%)
0.0% ( -5% - 6%)
PKLookup 157.80 (2.5%) 157.83 (2.5%)
0.0% ( -4% - 5%)
LowSpanNear 21.48 (2.1%) 21.48 (2.3%)
0.0% ( -4% - 4%)
OrHighNotMed 590.59 (3.9%) 591.21 (3.8%)
0.1% ( -7% - 8%)
BrowseMonthTaxoFacets 1.06 (1.1%) 1.06 (0.9%)
0.1% ( -1% - 2%)
LowSloppyPhrase 40.57 (2.1%) 40.63 (2.2%)
0.1% ( -4% - 4%)
IntNRQ 124.31 (4.2%) 124.53 (4.9%)
0.2% ( -8% - 9%)
BrowseDateTaxoFacets 1.00 (1.0%) 1.00 (0.7%)
0.2% ( -1% - 1%)
BrowseDayOfYearTaxoFacets 0.99 (0.9%) 1.00 (0.7%)
0.2% ( -1% - 1%)
HighTermDayOfYearSort 18.57 (6.2%) 18.62 (6.0%)
0.3% ( -11% - 13%)
BrowseMonthSSDVFacets 4.38 (1.0%) 4.40 (0.9%)
0.4% ( -1% - 2%)
BrowseDayOfYearSSDVFacets 3.92 (0.7%) 3.94 (0.7%)
0.5% ( 0% - 1%)
Wildcard 52.17 (4.0%) 52.47 (5.0%)
0.6% ( -8% - 9%)
Fuzzy2 57.57 (9.5%) 58.32 (9.3%)
1.3% ( -16% - 22%)
HighTermMonthSort 40.51 (14.2%) 41.47 (13.9%)
2.4% ( -22% - 35%)
{quote}There's an option for lucene-util to format the output for JIRA
{quote}
Last time I used this option Jira interpreted some tags and the resulting
display was not better than this basic one.
{quote}Looking at the results you posted, the optimization seems fairly
invisible
{quote}
Yes. The change optimizes the construction only of the CompiledAutomaton, so
this is a tiny part of the fuzzy query execution.
{quote}that's 4.7% of "noise"
{quote}
Yes, there is noise. I tried baseline vs baseline and got the same noise. Maybe
with wikimediumall this time there is less noise.
> Improve Automaton.step() with binary search and introduce Automaton.next()
> --------------------------------------------------------------------------
>
> Key: LUCENE-9125
> URL: https://issues.apache.org/jira/browse/LUCENE-9125
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Bruno Roustant
> Assignee: Bruno Roustant
> Priority: Major
> Fix For: 8.5
>
> Time Spent: 40m
> Remaining Estimate: 0h
>
> Implement the existing todo in Automaton.step() (lookup a transition from a
> source state depending on a given label) to use binary search since the
> transitions are sorted.
> Introduce new method Automaton.next() to optimize iteration & lookup over all
> the transitions of a state. This will be used in RunAutomaton constructor and
> in MinimizationOperations.minimize().
--
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
