[
https://issues.apache.org/jira/browse/LUCENE-3233?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Robert Muir updated LUCENE-3233:
--------------------------------
Attachment: LUCENE-3233.patch
here is a patch with a little microbenchmark... so we have some tuning to do.
the benchmark analyzes a short string a million times, that doesn't match any
synonyms (actually hte solr default)
||impl||ms||
|SynonymsFilter|1692|
|FST with array arcs|2794|
|FST with no array arcs|8823|
so, disabling the array arcs is a pretty crucial hit here. but we could do
other options to speed up this common case, e.g. with daciuk we could build a
charrunautomaton of the K-prefixes of the synonyms, this would be really fast
to reject these terms that don't match any syns.
or we could explicitly put our bytesref output in a byte[], and use long
pointers as outputs.
or we could speed up FST! But i think its interesting to see how important this
parameter is.
> HuperDuperSynonymsFilter™
> -------------------------
>
> Key: LUCENE-3233
> URL: https://issues.apache.org/jira/browse/LUCENE-3233
> Project: Lucene - Java
> Issue Type: Improvement
> Reporter: Robert Muir
> Attachments: LUCENE-3223.patch, LUCENE-3233.patch, LUCENE-3233.patch,
> LUCENE-3233.patch, LUCENE-3233.patch, LUCENE-3233.patch, LUCENE-3233.patch,
> LUCENE-3233.patch, LUCENE-3233.patch, LUCENE-3233.patch, synonyms.zip
>
>
> The current synonymsfilter uses a lot of ram and cpu, especially at build
> time.
> I think yesterday I heard about "huge synonyms files" three times.
> So, I think we should use an FST-based structure, sharing the inputs and
> outputs.
> And we should be more efficient with the tokenStream api, e.g. using
> save/restoreState instead of cloneAttributes()
--
This message is automatically generated by JIRA.
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]
