[
https://issues.apache.org/jira/browse/LUCENE-5416?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13889410#comment-13889410
]
Paul Elschot commented on LUCENE-5416:
--------------------------------------
Recent comments at LUCENE-5425 indicate that performance of OpenBitSetIterator
is critical for facets.
Would it be correct to conclude from that and from the measurements here that
faceting involves nextDoc() on high bit densities?
That would also mean that the FixedBitSetDBI here could be better for as a
general filter with a lower bit density,
and that could have implications for LUCENE-5293.
For even lower bit densities a more compressed version is preferable, and we
already have WAH8DocIdSet for that.
For the DocBlockIterator at LUCENE-5092 I'll stick to the FixedBitSetDBI for
now, but since it does a prevDoc() under the hood,
it might be a good idea to use the same technique there as in
OpenBitSetIterator, only backwards.
Would someone have an idea how to merge the Long.numberOfTrailingZeros() used
here for advance() into OpenBitSetIterator?
Or would it be better to always choose a DocIdSet implementation based on bit
density?
> Performance of a FixedBitSet variant that uses Long.numberOfTrailingZeros()
> ---------------------------------------------------------------------------
>
> Key: LUCENE-5416
> URL: https://issues.apache.org/jira/browse/LUCENE-5416
> Project: Lucene - Core
> Issue Type: Bug
> Components: core/search
> Affects Versions: 5.0
> Reporter: Paul Elschot
> Priority: Minor
> Fix For: 5.0
>
>
> On my machine the current byte index used in OpenBitSetIterator is slower
> than Long.numberOfTrailingZeros() for advance().
> The pull request contains the code for benchmarking this taken from an early
> stage of DocBlocksIterator.
> In case the benchmark shows improvements on more machines, well, we know what
> to do...
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
