easyice commented on PR #12954:
URL: https://github.com/apache/lucene/pull/12954#issuecomment-1864749851
I took several hours to confirm the results, the benchmark shows it became
faster, this exceeded my expectation, we think the speedup is due to remove the
loop that initializes the `freqBuffer` to 1 in `reset()` like below:
```
if (indexHasFreq == false || needsFreq == false) {
for (int i = 0; i < ForUtil.BLOCK_SIZE; ++i) {
freqBuffer[i] = 1;
}
}
```
Since if we always allocate the 128-size `freqBuffer` for this PR, the
benchmark shows it still has a speedup. therefore, performance improvement has
no relevance to reducing memory allocation. so maybe we can consider the other
approach: try to avoid the for-loop in `reset()` if the instance can be reused.
thanks for the suggestions from @gf2121 when i investigating the cause of the
performance speedup.
Benchmark output for the PR(using `wikimediumall`):
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
HighSloppyPhrase 0.38 (5.3%) 0.37
(4.6%) -1.6% ( -10% - 8%) 0.323
MedTerm 226.03 (4.8%) 223.08
(5.2%) -1.3% ( -10% - 9%) 0.409
HighIntervalsOrdered 2.19 (6.4%) 2.17
(6.5%) -0.9% ( -12% - 12%) 0.676
MedSloppyPhrase 18.52 (2.7%) 18.39
(2.5%) -0.7% ( -5% - 4%) 0.402
Fuzzy2 36.10 (1.8%) 35.86
(1.6%) -0.7% ( -3% - 2%) 0.219
Fuzzy1 43.40 (1.6%) 43.16
(1.7%) -0.6% ( -3% - 2%) 0.276
Respell 21.69 (1.8%) 21.58
(1.8%) -0.5% ( -4% - 3%) 0.375
LowTerm 232.03 (3.0%) 231.08
(2.9%) -0.4% ( -6% - 5%) 0.659
LowSloppyPhrase 18.26 (2.0%) 18.20
(2.1%) -0.3% ( -4% - 3%) 0.660
HighTerm 267.11 (5.2%) 266.50
(5.6%) -0.2% ( -10% - 11%) 0.893
HighSpanNear 1.85 (5.7%) 1.84
(6.7%) -0.2% ( -11% - 12%) 0.935
OrHighNotLow 167.52 (5.7%) 167.26
(5.6%) -0.2% ( -10% - 11%) 0.931
HighTermTitleBDVSort 1.90 (3.7%) 1.90
(4.5%) -0.1% ( -7% - 8%) 0.915
MedIntervalsOrdered 7.07 (3.4%) 7.06
(3.8%) -0.1% ( -7% - 7%) 0.910
MedSpanNear 24.97 (2.1%) 24.94
(2.7%) -0.1% ( -4% - 4%) 0.874
HighPhrase 10.67 (6.0%) 10.66
(5.7%) -0.1% ( -11% - 12%) 0.950
LowPhrase 4.70 (4.0%) 4.70
(3.8%) -0.0% ( -7% - 8%) 0.979
OrHighNotMed 130.98 (6.2%) 131.01
(6.1%) 0.0% ( -11% - 13%) 0.989
OrNotHighHigh 171.61 (5.4%) 171.67
(5.3%) 0.0% ( -10% - 11%) 0.984
LowIntervalsOrdered 28.65 (4.3%) 28.68
(4.3%) 0.1% ( -8% - 9%) 0.947
OrHighHigh 18.94 (2.9%) 19.00
(3.6%) 0.3% ( -5% - 6%) 0.766
OrHighNotHigh 125.97 (5.5%) 126.41
(6.0%) 0.3% ( -10% - 12%) 0.848
OrNotHighMed 181.48 (4.0%) 182.38
(3.6%) 0.5% ( -6% - 8%) 0.679
LowSpanNear 6.89 (2.5%) 6.93
(3.2%) 0.6% ( -5% - 6%) 0.516
MedPhrase 110.79 (2.8%) 111.45
(2.9%) 0.6% ( -5% - 6%) 0.515
OrHighMed 38.51 (2.4%) 38.79
(2.1%) 0.7% ( -3% - 5%) 0.311
AndHighMed 40.73 (2.4%) 41.06
(2.5%) 0.8% ( -4% - 5%) 0.304
TermDTSort 74.72 (4.1%) 75.32
(2.6%) 0.8% ( -5% - 7%) 0.460
AndHighHigh 10.24 (5.6%) 10.33
(3.9%) 0.8% ( -8% - 11%) 0.600
HighTermMonthSort 1071.18 (2.9%) 1079.84
(4.5%) 0.8% ( -6% - 8%) 0.499
AndHighLow 167.91 (5.2%) 170.10
(5.6%) 1.3% ( -9% - 12%) 0.446
IntNRQ 13.84 (4.0%) 14.05
(3.3%) 1.5% ( -5% - 9%) 0.208
OrNotHighLow 241.06 (4.5%) 244.91
(4.8%) 1.6% ( -7% - 11%) 0.276
OrHighLow 175.82 (4.0%) 178.70
(3.5%) 1.6% ( -5% - 9%) 0.174
HighTermDayOfYearSort 188.33 (3.3%) 191.56
(3.7%) 1.7% ( -5% - 8%) 0.121
Prefix3 391.46 (4.0%) 404.16
(3.5%) 3.2% ( -4% - 11%) 0.006
HighTermTitleSort 87.08 (3.9%) 91.84
(4.1%) 5.5% ( -2% - 14%) 0.000
Wildcard 33.16 (3.3%) 37.56
(3.3%) 13.3% ( 6% - 20%) 0.000
PKLookup 92.19 (1.8%) 104.90
(2.8%) 13.8% ( 9% - 18%) 0.000
```
--
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
