lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Michael McCandless (JIRA)" <>
Subject [jira] [Commented] (LUCENE-3200) Cleanup MMapDirectory to use only one MMapIndexInput impl with mapping sized of powers of 2
Date Tue, 14 Jun 2011 10:49:47 GMT


Michael McCandless commented on LUCENE-3200:

+1 to commit!

In my stress NRT test (runs optimize on a full Wiki index with ongoing
indexing / reopening), without this patch, I see performance drop
substantially (like 180 QPS down to 140 QPS) when the JVM cuts over to
the optimized segment.  With the patch I see it jump up a bit after
the optimize completes!  So this seems to make hotspot's job

> Cleanup MMapDirectory to use only one MMapIndexInput impl with mapping sized of powers
of 2
> -------------------------------------------------------------------------------------------
>                 Key: LUCENE-3200
>                 URL:
>             Project: Lucene - Java
>          Issue Type: Improvement
>            Reporter: Uwe Schindler
>            Assignee: Uwe Schindler
>         Attachments: LUCENE-3200.patch, LUCENE-3200.patch, LUCENE-3200.patch, LUCENE-3200.patch,
LUCENE-3200.patch, LUCENE-3200_tests.patch
> Robert and me discussed a little bit after Mike's investigations, that using SingleMMapIndexinput
together with MultiMMapIndexInput leads to hotspot slowdowns sometimes.
> We had the following ideas:
> - MultiMMapIndexInput is almost as fast as SingleMMapIndexInput, as the switching between
buffer boundaries is done in exception catch blocks. So normal code path is always the same
like for Single*
> - Only the seek method uses strange calculations (the modulo is totally bogus, it could
be simply: int bufOffset = (int) (pos % maxBufSize); - very strange way of calculating modulo
in the original code)
> - Because of speed we suggest to no longer use arbitrary buffer sizes. We should pass
only the power of 2 to the indexinput as size. All calculations in seek and anywhere else
would be simple bit shifts and AND operations (the and masks for the modulo can be calculated
in the ctor like NumericUtils does when calculating precisionSteps).
> - the maximum buffer size will now be 2^30, not 2^31-1. But thats not an issue at all.
In my opinion, a buffer size of 2^31-1 is stupid in all cases, as it will no longer fit page
boundaries and mmapping gets harder for the O/S.
> We will provide a patch with those cleanups.

This message is automatically generated by JIRA.
For more information on JIRA, see:


To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message