msokolov commented on PR #14874:
URL: https://github.com/apache/lucene/pull/14874#issuecomment-3058163612
OK I discovered the loss of recall was due to a silly bug. After fixing
that, these are the results I'm seeing with the addition of a separate code
path for `dotProduct(byte[], byte[])`:
## candidate (reindex)
```
recall latency(ms) netCPU avgCpuCount nDoc topK fanout maxConn
beamWidth quantized visited index(s) num_segments
0.979 7.794 7.795 1.000 500000 100 50 32
250 no 25766 497.73 3
0.979 13.193 13.194 1.000 500000 100 50 32
250 no 25766 0.00 3
0.979 13.487 13.489 1.000 500000 100 50 32
250 no 25766 0.00 3
```
## baseline
```
0.979 24.282 24.280 1.000 500000 100 50 32
250 no 25766 0.00 3
0.979 25.850 25.848 1.000 500000 100 50 32
250 no 25766 0.00 3
```
## baseline (reindex)
```
0.986 16.265 16.265 1.000 500000 100 50 32
250 no 25702 441.32 3
0.986 24.319 24.316 1.000 500000 100 50 32
250 no 25702 0.00 3
0.986 24.319 24.316 1.000 500000 100 50 32
250 no 25702 0.00 3
```
## candidate
```
0.986 13.406 13.407 1.000 500000 100 50 32
250 no 25702 0.00 3
0.986 13.088 13.089 1.000 500000 100 50 32
250 no 25702 0.00 3
0.986 13.319 13.320 1.000 500000 100 50 32
250 no 25702 0.00 3
```
so now the situation is reversed and we see *faster* searching after
indexing in the same JVM.
--
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
