[
https://issues.apache.org/jira/browse/LUCENE-7101?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael McCandless updated LUCENE-7101:
---------------------------------------
Attachment: LUCENE-7101.patch
Patch, adding a trivial "log merge policy" to {{OfflineSorter}}, and changing
its default merge factor from 128 to 10.
I also fixed the 2B points tests to not "cheat" by giving BKD more heap than
the defaults, and improved {{BKDWriter}}'s temp file naming
I'm running {{Test2BBKDPoints.test2D}} now ...
> OfflineSorter's merging is O(N^2) cost for large sorts
> ------------------------------------------------------
>
> Key: LUCENE-7101
> URL: https://issues.apache.org/jira/browse/LUCENE-7101
> Project: Lucene - Core
> Issue Type: Bug
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Priority: Blocker
> Fix For: master, 6.0
>
> Attachments: LUCENE-7101.patch
>
>
> Our {{OfflineSorter}} acts just like Lucene, writing small initial
> segments of sorted values (from what it was able to sort at once in
> heap), periodically merging them when there are too many, and doing a
> {{forceMerge(1)}} in the end.
> But the merge logic is too simplistic today, resulting in O(N^2)
> cost. Smallish sorts usually won't hit it, because the default 128
> merge factor is so high, but e.g. the new 2B points tests do hit the
> N^2 behavior. I suspect the high merge factor hurts performance (OS
> struggles to use what free RAM it has to read-ahead on 128 files), and
> also risks file descriptor exhaustion.
> I think we should implement a simple log merge policy for it, and drop
> its default merge factor to 10.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
