[
https://issues.apache.org/jira/browse/LUCENE-4366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13455940#comment-13455940
]
Michael McCandless commented on LUCENE-4366:
--------------------------------------------
I agree it'd be nice to not score MUST_NOT clauses for both BS and
BS2. Really crazy that we do that!
Separately I think BS should handle MUST clauses in some cases...
There are several optimizations:
* When collecting first clause per-chunk, don't bother checking
whether the bucket is stale since it will always be (saves an if
per hit). I also sort by smallest first-docID first (proxy for
highest docFreq) so this saved-if-per-hit has the most impact.
* Use int[] instead of linked list to record filled buckets.
* Don't call .score for prohibited hits.
* Don't enroll prohibited hits into the "live" buckets list.
* Don't call .score for a hit that was already prohibited due to a
previous clause (I sort prohibited clauses first for this
reason).
* Use "boolean prohibited" instead of int bitmask (not sure this
matters but it was confusing to use a bitmask).
I suspect the saved if per collect is most of the gains (the results
above didn't test MUST_NOT clauses ... still TODO). But ... I'm getting
erratic results when performance testing (and can't reproduce the above
results on my current patch...). Not sure what's up...
> Small speedups for BooleanScorer
> --------------------------------
>
> Key: LUCENE-4366
> URL: https://issues.apache.org/jira/browse/LUCENE-4366
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Attachments: LUCENE-4366.patch, LUCENE-4366.patch
>
>
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
