zhaih commented on PR #15208:
URL: https://github.com/apache/lucene/pull/15208#issuecomment-3858491870
I tried to play with the beamWidth and realized the recall is actually quite
hard to improve:
```
### Candidate
recall latency(ms) netCPU avgCpuCount nDoc topK fanout maxConn
beamWidth quantized visited index(s) index_docs/s force_merge(s)
num_segments index_size(MB) vec_disk(MB) vec_RAM(MB) indexType
0.700 0.272 0.264 0.971 5000000 100 50 48
128 no 1405 99.00 50506.58 113.60
1 2165.19 1907.349 1907.349 HNSW
0.750 0.276 0.271 0.982 5000000 100 50 48
256 no 1643 182.84 27346.76 165.61
1 2224.54 1907.349 1907.349 HNSW
0.771 0.294 0.290 0.986 5000000 100 50 48
512 no 1744 338.41 14774.93 239.39
1 2247.09 1907.349 1907.349 HNSW
0.779 0.294 0.289 0.983 5000000 100 50 48
768 no 1794 487.29 10260.75 261.00
1 2257.86 1907.349 1907.349 HNSW
### Baseline
recall latency(ms) netCPU avgCpuCount nDoc topK fanout maxConn
beamWidth quantized visited index(s) index_docs/s force_merge(s)
num_segments index_size(MB) vec_disk(MB) vec_RAM(MB) indexType
0.742 0.277 0.272 0.982 5000000 100 50 48
128 no 1604 107.16 46658.33 104.35
1 2211.86 1907.349 1907.349 HNSW
0.782 0.306 0.301 0.984 5000000 100 50 48
256 no 1852 207.64 24080.49 193.89
1 2273.83 1907.349 1907.349 HNSW
0.808 0.343 0.338 0.985 5000000 100 50 48
512 no 2123 375.13 13328.82 380.45
1 2336.24 1907.349 1907.349 HNSW
0.825 0.362 0.357 0.986 5000000 100 50 48
768 no 2294 582.56 8582.85 622.42
1 2372.60 1907.349 1907.349 HNSW
```
If we look at the 0.779 recall in candidate vs. 0.782 in baseline I think
the baseline definitely wins.. So for now I'll just close this PR. Tho I think
some ideas here might be able to improve this overall performance so someone
else/me probably can take a look in the future and grab some ideas.
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
