[
https://issues.apache.org/jira/browse/LUCENE-888?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12499280
]
Marvin Humphrey commented on LUCENE-888:
----------------------------------------
I have some auxiliary data points to report after experimenting with buffer
size in KS today on three different systems: OS X 10.4.9, FreeBSD 5.3, and an
old RedHat 9 box.
The FS i/o classes in KinoSearch use a FILE* and
fopen/fwrite/fread/fseek/ftell, rather than file descriptors and the POSIX
family of functions. Theoretically, this is wasteful because FILE* stream i/o
is buffered, so there's double buffering happening. I've meant to change that
for some time. However, when I've used setvbuf(self->fhandle, NULL, _IONBF)
to eliminate the buffer for the underlying FILE* object, performance tanks --
indexing time doubles. I still don't understand exactly why, but I know a
little more now.
* Swapping out the FILE* for a descriptor and switching all the I/O calls to
POSIX variants has no measurable impact on any of these systems.
* Changing the KS buffer size from 1024 to 4096 has no measurable impact on
any of these systems.
* Using setvbuf to eliminate the buffering at output turns out to have no
impact on indexing performance. It's only killing off the read mode FILE*
buffer that causes the problem.
So, it seems that the only change I can make moves the numbers in the wrong
direction.
The results are somewhat puzzling because I would ordinarily have blamed
sub-optimal flush/refill scheduling in my app for the degraded performance
with setvbuf() on read mode. However, the POSIX i/o calls are unbuffered, so
that's not it. My best guess is that disabling buffering for read mode
disables an fseek/ftell optimization.
> Improve indexing performance by increasing internal buffer sizes
> ----------------------------------------------------------------
>
> Key: LUCENE-888
> URL: https://issues.apache.org/jira/browse/LUCENE-888
> Project: Lucene - Java
> Issue Type: Improvement
> Components: Index
> Affects Versions: 2.1
> Reporter: Michael McCandless
> Assigned To: Michael McCandless
> Priority: Minor
> Attachments: LUCENE-888.patch, LUCENE-888.take2.patch
>
>
> In working on LUCENE-843, I noticed that two buffer sizes have a
> substantial impact on overall indexing performance.
> First is BufferedIndexOutput.BUFFER_SIZE (also used by
> BufferedIndexInput). Second is CompoundFileWriter's buffer used to
> actually build the compound file. Both are now 1 KB (1024 bytes).
> I ran the same indexing test I'm using for LUCENE-843. I'm indexing
> ~5,500 byte plain text docs derived from the Europarl corpus
> (English). I index 200,000 docs with compound file enabled and term
> vector positions & offsets stored plus stored fields. I flush
> documents at 16 MB RAM usage, and I set maxBufferedDocs carefully to
> not hit LUCENE-845. The resulting index is 1.7 GB. The index is not
> optimized in the end and I left mergeFactor @ 10.
> I ran the tests on a quad-core OS X 10 machine with 4-drive RAID 0 IO
> system.
> At 1 KB (current Lucene trunk) it takes 622 sec to build the index; if
> I increase both buffers to 8 KB it takes 554 sec to build the index,
> which is an 11% overall gain!
> I will run more tests to see if there is a natural knee in the curve
> (buffer size above which we don't really gain much more performance).
> I'm guessing we should leave BufferedIndexInput's default BUFFER_SIZE
> at 1024, at least for now. During searching there can be quite a few
> of this class instantiated, and likely a larger buffer size for the
> freq/prox streams could actually hurt search performance for those
> searches that use skipping.
> The CompoundFileWriter buffer is created only briefly, so I think we
> can use a fairly large (32 KB?) buffer there. And there should not be
> too many BufferedIndexOutputs alive at once so I think a large-ish
> buffer (16 KB?) should be OK.
--
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: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
