msokolov commented on pull request #315:
URL: https://github.com/apache/lucene/pull/315#issuecomment-926762372
I did run a quick benchmark on an internal data-set (using KnnGraphTester)
and saw some improvement to both latency and recall, and a reduction in nodes
visited from this change:
condition|recall| latency| nDoc | fanout| maxConn| beamWidth|
visited| index ms
------------|-------|-----------|--------------|------------|---------------|---------------|------------|----------------
main |0.815| 0.91| 1000000| 50| 32|
64| 2590| 484529
change |0.836| 0.86 | 1000000 | 50| 32|
64| 2331| 462900
and I think this test is actually unfair to HNSW because of the fanout which
incurs some additional cost in the HNSW case for no benefit (it just increases
top K by that amount and then throws away the extras).
--
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]
