[
https://issues.apache.org/jira/browse/LUCENE-2723?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Yonik Seeley updated LUCENE-2723:
---------------------------------
Attachment: LUCENE-2723_facetPerSeg.patch
Here's a patch that changes just one pace in faceting to per-segment bulk.
Same 10M doc index, testing w/ cache.minDf=maxdoc only (no use of filterCache
since I haven't changed that to per-seg yet). Time in ms.
|unique values in field|trunk per-seg|branch per-seg|speedup|
|10|161|173|-7%|
|100|217|218|-0%|
|1000|267|262|2%|
|10000|465|325|43%|
|100000|2025|678|199%|
|10000000|21061|4393|379%|
Now, facet.method=enum wasn't even designed for many unique values in a field,
but this more efficient per-segment bulk code certainly expands the range where
it's feasible. My guess is that the speedup is due to us dropping to
per-segment quicker with this patch (trunk get's a bulk enum, and then drops to
per-segment).
The drop in performance for the high df field (each value will match ~1M docs)
is curious. Seems like this should be a more efficient inner loop, but I guess
hotspot just optimized it differently.
Based on these results, I'll re-convert the rest of the code to go per-segment
too.
> Speed up Lucene's low level bulk postings read API
> --------------------------------------------------
>
> Key: LUCENE-2723
> URL: https://issues.apache.org/jira/browse/LUCENE-2723
> Project: Lucene - Java
> Issue Type: Improvement
> Components: Index
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Fix For: 4.0
>
> Attachments: LUCENE-2723-termscorer.patch,
> LUCENE-2723-termscorer.patch, LUCENE-2723-termscorer.patch,
> LUCENE-2723.patch, LUCENE-2723.patch, LUCENE-2723.patch, LUCENE-2723.patch,
> LUCENE-2723.patch, LUCENE-2723_facetPerSeg.patch, LUCENE-2723_openEnum.patch,
> LUCENE-2723_termscorer.patch, LUCENE-2723_wastedint.patch
>
>
> Spinoff from LUCENE-1410.
> The flex DocsEnum has a simple bulk-read API that reads the next chunk
> of docs/freqs. But it's a poor fit for intblock codecs like FOR/PFOR
> (from LUCENE-1410). This is not unlike sucking coffee through those
> tiny plastic coffee stirrers they hand out airplanes that,
> surprisingly, also happen to function as a straw.
> As a result we see no perf gain from using FOR/PFOR.
> I had hacked up a fix for this, described at in my blog post at
> http://chbits.blogspot.com/2010/08/lucene-performance-with-pfordelta-codec.html
> I'm opening this issue to get that work to a committable point.
> So... I've worked out a new bulk-read API to address performance
> bottleneck. It has some big changes over the current bulk-read API:
> * You can now also bulk-read positions (but not payloads), but, I
> have yet to cutover positional queries.
> * The buffer contains doc deltas, not absolute values, for docIDs
> and positions (freqs are absolute).
> * Deleted docs are not filtered out.
> * The doc & freq buffers need not be "aligned". For fixed intblock
> codecs (FOR/PFOR) they will be, but for varint codecs (Simple9/16,
> Group varint, etc.) they won't be.
> It's still a work in progress...
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
