You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@lucene.apache.org by Michael McCandless <lu...@mikemccandless.com> on 2007/03/23 20:52:20 UTC
Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then
IndexWriter may over-merge
"Grant Ingersoll" <gs...@apache.org> wrote:
> Your timing is ironic. I was just running some benchmarks for
> ApacheCon (using contrib/benchmarker) and noticed what I think are
> similar happenings, so maybe you can validate my assumptions. I'm
> not sure if it is because I'm hitting RAM issues or not.
>
> Below is the algorithm file for use w/ benchmarker. To run it, save
> the file, cd into contrib/benchamarker (make sure you get the lastest
> commits) and run
> ant run-task -Dtask.mem=XXXXm -Dtask.alg=<path to file>
>
> The basic idea is, there are ~21580 docs in the Reuters, so I wanted
> to run some experiments around them with different merge factors and
> max.buffered. Granted, some of the factors are ridiculous, but I
> wanted to look at these a bit b/c you see people on the user list
> from time to time talking/asking about setting really high numbers
> for mergeFactor and maxBufferedDocs.
>
> The sweet spot on my machine seems to be mergeFactor == 100,
> maxBD=1000. I ran with -Dtask.mem=1024M on a machine with 2gb of
> RAM. If I am understanding the numbers correctly, and what you are
> arguing, this sweet spot happens to coincide approximately with the
> amount of memory I gave the process. I probably could play a little
> bit more with options to reach the inflection point. So, to some
> extent, I think your approach for RAM based modeling is worth pursuing.
Interesting results! Because an even higher maxBufferedDocs (10000 =
299.1 rec/s and 21580 = 271.9 rec/s, @ mergeFactor=100) gave you worse
performance even though they were able to complete (meaning you had
enough RAM to buffer all those docs). Perhaps this is because GC had
to work harder? So it seems like at some point the benefits of giving
more RAM taper off.
Also, one caveat: whenever #docs (21578 for Reuters) divided by
maxBuffered docs is less than mergeFactor, you will have no merges
take place during your runs. This greatly skews the results.
I'm also struggling with this on LUCENE-843 because it makes it far
harder to do apples to apples comparison. With the patch for
LUCENE-843, many more docs can be buffered into a given fixed MB RAM.
So then it flushes less often and may hit no merges (when the baseline
Lucene trunk does hit merges), or the opposite: it may hit a massive
large merge close to the end when the baseline Lucene trunk did a few
small merges. We sort of need some metric that can normalize for how
much "merge servicing" took place during a run, or something.
Mike
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
RE: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then IndexWriter may over-merge
Posted by Steven Parkes <st...@esseff.org>.
> Very long documents are useful for testing for anomalies, but they're
> not so useful as retrieved documents, nor typical of applications.
That's what I thought, too. I'm kinda curious to see how gutenburg
compares to wikipedia for things like merge policy, in particular,
by-docs vs. by-bytes. But while a curiosity, I don't see that it has all
much practical value, so I'm not sure I can get it to the top of the
to-do list.
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then
IndexWriter may over-merge
Posted by Doug Cutting <cu...@apache.org>.
Steven Parkes wrote:
> And what about Project Gutenburg?
>
> Wikipedia is going to have relatively short text, Gutenburg very long.
Very long documents are useful for testing for anomalies, but they're
not so useful as retrieved documents, nor typical of applications. Very
long hits are awkward for users. Book search engines usually operate
best either by breaking texts into small units (chapters, pages,
overlapping windows, etc.) and searching those rather than the entire
work, perhaps merging multiple hits from the same work in displayed
results. (See, e.g., California Digital Library's XTF system, built by
Kirk Hastings using Lucene. http://www.cdlib.org/inside/projects/xtf/)
I think Wikipedia is a much more typical use of Lucene.
Doug
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
RE: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then IndexWriter may over-merge
Posted by Steven Parkes <st...@esseff.org>.
And what about Project Gutenburg?
Wikipedia is going to have relatively short text, Gutenburg very long.
-----Original Message-----
From: Steven Parkes [mailto:steven_parkes@esseff.org]
Sent: Friday, March 23, 2007 2:37 PM
To: java-dev@lucene.apache.org
Subject: RE: [jira] Commented: (LUCENE-845) If you "flush by RAM usage"
then IndexWriter may over-merge
Well, since I want to look at the impact of merge policy, I'll look into
this.
Wikipedia is easy to download (bandwidth notwithdstanding). The bz2'd of
the current English pages is 2.1G. That's certainly a lot of data. It
looks like the English is about 1.8M docs. All languages is something
like 21M now.
I was also thinking of the TREC data but that seems hard to come by?
-----Original Message-----
From: Grant Ingersoll [mailto:gsingers@apache.org]
Sent: Friday, March 23, 2007 1:09 PM
To: java-dev@lucene.apache.org
Subject: Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage"
then IndexWriter may over-merge
Yeah, I didn't play yet with millions of documents. We will need a
bigger test collection, I think! Although the benchmarker can add as
many as you want from the same source, index compression will effect
the results possibly more than a bigger collection with all unique docs.
Maybe it is time to look at adding Wikipedia as a test collection. I
think there are something like 18+ million docs in it.
On Mar 23, 2007, at 4:01 PM, Doug Cutting wrote:
> Michael McCandless wrote:
>> Also, one caveat: whenever #docs (21578 for Reuters) divided by
>> maxBuffered docs is less than mergeFactor, you will have no merges
>> take place during your runs. This greatly skews the results.
>
> Also, my guess is that this index fits entirely in the buffer
> cache. Things behave quite differently when segments are larger
> than available memory and merging requires lots of disk i/o.
>
> Doug
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-dev-help@lucene.apache.org
>
--------------------------
Grant Ingersoll
Center for Natural Language Processing
http://www.cnlp.org
Read the Lucene Java FAQ at http://wiki.apache.org/jakarta-lucene/
LuceneFAQ
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
RE: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then IndexWriter may over-merge
Posted by Steven Parkes <st...@esseff.org>.
Well, since I want to look at the impact of merge policy, I'll look into
this.
Wikipedia is easy to download (bandwidth notwithdstanding). The bz2'd of
the current English pages is 2.1G. That's certainly a lot of data. It
looks like the English is about 1.8M docs. All languages is something
like 21M now.
I was also thinking of the TREC data but that seems hard to come by?
-----Original Message-----
From: Grant Ingersoll [mailto:gsingers@apache.org]
Sent: Friday, March 23, 2007 1:09 PM
To: java-dev@lucene.apache.org
Subject: Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage"
then IndexWriter may over-merge
Yeah, I didn't play yet with millions of documents. We will need a
bigger test collection, I think! Although the benchmarker can add as
many as you want from the same source, index compression will effect
the results possibly more than a bigger collection with all unique docs.
Maybe it is time to look at adding Wikipedia as a test collection. I
think there are something like 18+ million docs in it.
On Mar 23, 2007, at 4:01 PM, Doug Cutting wrote:
> Michael McCandless wrote:
>> Also, one caveat: whenever #docs (21578 for Reuters) divided by
>> maxBuffered docs is less than mergeFactor, you will have no merges
>> take place during your runs. This greatly skews the results.
>
> Also, my guess is that this index fits entirely in the buffer
> cache. Things behave quite differently when segments are larger
> than available memory and merging requires lots of disk i/o.
>
> Doug
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-dev-help@lucene.apache.org
>
--------------------------
Grant Ingersoll
Center for Natural Language Processing
http://www.cnlp.org
Read the Lucene Java FAQ at http://wiki.apache.org/jakarta-lucene/
LuceneFAQ
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then IndexWriter may over-merge
Posted by Grant Ingersoll <gs...@apache.org>.
Yeah, I didn't play yet with millions of documents. We will need a
bigger test collection, I think! Although the benchmarker can add as
many as you want from the same source, index compression will effect
the results possibly more than a bigger collection with all unique docs.
Maybe it is time to look at adding Wikipedia as a test collection. I
think there are something like 18+ million docs in it.
On Mar 23, 2007, at 4:01 PM, Doug Cutting wrote:
> Michael McCandless wrote:
>> Also, one caveat: whenever #docs (21578 for Reuters) divided by
>> maxBuffered docs is less than mergeFactor, you will have no merges
>> take place during your runs. This greatly skews the results.
>
> Also, my guess is that this index fits entirely in the buffer
> cache. Things behave quite differently when segments are larger
> than available memory and merging requires lots of disk i/o.
>
> Doug
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-dev-help@lucene.apache.org
>
--------------------------
Grant Ingersoll
Center for Natural Language Processing
http://www.cnlp.org
Read the Lucene Java FAQ at http://wiki.apache.org/jakarta-lucene/
LuceneFAQ
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org
Re: [jira] Commented: (LUCENE-845) If you "flush by RAM usage" then
IndexWriter may over-merge
Posted by Doug Cutting <cu...@apache.org>.
Michael McCandless wrote:
> Also, one caveat: whenever #docs (21578 for Reuters) divided by
> maxBuffered docs is less than mergeFactor, you will have no merges
> take place during your runs. This greatly skews the results.
Also, my guess is that this index fits entirely in the buffer cache.
Things behave quite differently when segments are larger than available
memory and merging requires lots of disk i/o.
Doug
---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org