iprithv opened a new pull request, #16079:
URL: https://github.com/apache/lucene/pull/16079

   there was a TODO in MergingHnswGraphBuilder.build() where we were scanning 
all nodes even if most were already initialized, this change adds nextUnsetBit 
and uses that instead, so we only go over nodes that are not set, instead of 
checking every node, we just jump to the next unset one. 
   
   so earlier it goes through all nodes and check each one, now jump directly 
to next unset bit.
   this is faster when most bits are already set, which is common during merges
   
   benchmark (bitset iteration)
   
   | size      | set % | scan loop (us) | nextUnsetBit (us) | speedup |
   |-----------|-------|----------------|-------------------|---------|
   | 1,000     | 0%    | 8.76           | 11.89             | slower  |
   | 1,000     | 75%   | 2.27           | 1.25              | ~1.8x   |
   | 1,000     | 90%   | 2.30           | 0.54              | ~4.2x   |
   | 1,000     | 99%   | 2.24           | 0.06              | ~35x    |
   | 100,000   | 75%   | 217            | 82.3              | ~2.6x   |
   | 100,000   | 90%   | 218            | 33.4              | ~6.5x   |
   | 100,000   | 99%   | 215            | 4.69              | ~45x    |
   | 1,000,000 | 99%   | 2157           | 43.3              | ~49x    |
   
   
   luceneutil (real run)
   
   |              | baseline | candidate | result |
   |--------------|----------|-----------|--------|
   | force merge  | 220.11s  | 218.68s   | same   |
   | index time   | 224.73s  | 229.25s   | same   |
   | recall       | 0.988    | 0.988     | same   |


-- 
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]

Reply via email to