msokolov edited a comment on pull request #2176:
URL: https://github.com/apache/lucene-solr/pull/2176#issuecomment-754153605
I ran some vector search performance tests even though this isn't probably
the most representative test, but since I have been running these for other
reasons lately and I'm all geared up for it, I thought I would fire them up
with these changes. I see a pronounced slowdown there, about 40% increase in
latency for these KNN searches with this patch (I ran both conditions w/JDK16).
Most of the workload from these searches are random-access reads of bytes from
the index, conversion to float[], and computing dot products.
I did recently discover that we were (inadvertently) reading and writing the
`float[]` vectors that these searches are based on as big-endian, so perhaps if
(when) we switch that (I have a pending patch), it would work better in
conjunction with this??, although currently on master, at the `IndexInput`
level it is just reading bytes on the heap and only afterwards converting those
to `float[]`, so it doesn't seem likely there would be any interaction there; I
just mention it since it's top of mind, and since I plan to add
`DataInput.readFloats()`, which would be needed here too.
Then I also ran luceneutil tests and see interesting mixed results:
```
TaskQPS baseline StdDevQPS candidate StdDev
Pct diff p-value
BrowseMonthTaxoFacets 2.15 (7.7%) 1.19 (2.1%)
-44.5% ( -50% - -37%) 0.000
BrowseDayOfYearTaxoFacets 2.03 (8.1%) 1.16 (2.4%)
-43.0% ( -49% - -35%) 0.000
BrowseDateTaxoFacets 2.03 (8.2%) 1.16 (2.3%)
-42.9% ( -49% - -35%) 0.000
Respell 52.33 (2.4%) 46.89 (1.7%)
-10.4% ( -14% - -6%) 0.000
BrowseMonthSSDVFacets 12.00 (13.2%) 10.81 (6.9%)
-10.0% ( -26% - 11%) 0.003
PKLookup 132.89 (2.3%) 120.09 (1.3%)
-9.6% ( -13% - -6%) 0.000
AndHighHigh 24.98 (3.0%) 22.61 (2.1%)
-9.5% ( -14% - -4%) 0.000
MedSloppyPhrase 15.66 (3.9%) 14.37 (2.0%)
-8.2% ( -13% - -2%) 0.000
LowSpanNear 96.34 (2.3%) 88.72 (1.6%)
-7.9% ( -11% - -4%) 0.000
HighSloppyPhrase 10.57 (4.6%) 9.75 (2.7%)
-7.7% ( -14% - 0%) 0.000
Prefix3 37.23 (7.2%) 34.53 (6.1%)
-7.2% ( -19% - 6%) 0.001
MedSpanNear 15.63 (1.5%) 14.52 (2.0%)
-7.1% ( -10% - -3%) 0.000
LowSloppyPhrase 15.11 (4.4%) 14.06 (2.5%)
-6.9% ( -13% - 0%) 0.000
AndHighMed 72.87 (3.4%) 67.88 (3.0%)
-6.8% ( -12% - 0%) 0.000
AndHighLow 366.41 (3.2%) 342.99 (3.6%)
-6.4% ( -12% - 0%) 0.000
OrHighMed 69.00 (2.6%) 65.09 (2.7%)
-5.7% ( -10% - 0%) 0.000
LowTerm 930.90 (5.7%) 878.74 (4.3%)
-5.6% ( -14% - 4%) 0.000
Fuzzy2 46.43 (7.2%) 43.94 (5.4%)
-5.4% ( -16% - 7%) 0.008
Fuzzy1 64.31 (6.9%) 60.98 (6.5%)
-5.2% ( -17% - 8%) 0.014
HighSpanNear 19.07 (1.7%) 18.12 (2.6%)
-5.0% ( -9% - 0%) 0.000
MedPhrase 21.45 (2.4%) 20.53 (1.9%)
-4.3% ( -8% - 0%) 0.000
HighTermTitleBDVSort 73.32 (16.1%) 70.33 (17.0%)
-4.1% ( -32% - 34%) 0.437
OrHighHigh 16.82 (1.7%) 16.19 (1.9%)
-3.7% ( -7% - 0%) 0.000
BrowseDayOfYearSSDVFacets 9.56 (8.7%) 9.24 (6.2%)
-3.4% ( -16% - 12%) 0.159
HighIntervalsOrdered 9.02 (0.6%) 8.78 (0.7%)
-2.7% ( -4% - -1%) 0.000
OrNotHighMed 324.93 (2.7%) 316.20 (3.6%)
-2.7% ( -8% - 3%) 0.007
Wildcard 129.01 (2.5%) 126.91 (2.6%)
-1.6% ( -6% - 3%) 0.046
HighTermDayOfYearSort 113.01 (11.9%) 111.68 (10.6%)
-1.2% ( -21% - 24%) 0.741
OrNotHighLow 539.19 (2.8%) 534.10 (2.8%)
-0.9% ( -6% - 4%) 0.284
OrHighLow 314.01 (5.4%) 311.23 (5.9%)
-0.9% ( -11% - 11%) 0.620
HighTermMonthSort 96.00 (9.6%) 95.48 (10.1%)
-0.5% ( -18% - 21%) 0.863
TermDTSort 100.42 (10.9%) 100.52 (13.4%)
0.1% ( -21% - 27%) 0.980
LowPhrase 259.52 (1.9%) 262.37 (2.2%)
1.1% ( -2% - 5%) 0.091
HighTerm 872.57 (3.9%) 892.07 (3.4%)
2.2% ( -4% - 9%) 0.052
MedTerm 1003.36 (4.8%) 1029.34 (4.8%)
2.6% ( -6% - 12%) 0.086
HighPhrase 220.30 (1.9%) 228.63 (3.3%)
3.8% ( -1% - 9%) 0.000
OrHighNotHigh 372.34 (3.4%) 391.32 (4.2%)
5.1% ( -2% - 13%) 0.000
OrNotHighHigh 389.61 (2.7%) 415.68 (5.5%)
6.7% ( -1% - 15%) 0.000
OrHighNotMed 409.35 (1.9%) 445.76 (6.9%)
8.9% ( 0% - 18%) 0.000
OrHighNotLow 513.64 (4.2%) 561.96 (7.9%)
9.4% ( -2% - 22%) 0.000
IntNRQ 203.68 (2.3%) 224.37 (4.7%)
10.2% ( 3% - 17%) 0.000
```
note on the setup: I had to tinker with luceneutil a bit to get it to use
JDK16 runtime and add the foreign memory access module. gradle seemed to work
OK given the environment variable setup discussed above. I did check the
command line to make sure it was in fact using JDK16, and I saw it fail once
(before I added the module import command line argument), so I'm sure it is.
----------------------------------------------------------------
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.
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
