lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Erick Erickson" <erickerick...@gmail.com>
Subject Re: Reading Performance
Date Sat, 09 Dec 2006 00:29:20 GMT
I'm stumped. It seems like it might be time to haul a profiler out. I'm
particularly surprised because I put together a test system that fired a
bunch of threads at a searcher and saw nothing like you're seeing even up to
the 30 simultaneous requests running.

Do you know whether you're I/O bound, CPU bound, or???

It really sounds like there is some sort of synchronization issue, but
whether it's in Lucene or your code or some other system resource
contention, I have no idea.

I'd probably write a small program that spawned off several search threads
simultaneously to try to narrow it down (do something creative so you don't
just re-execute the same query and, perhaps, get skewed results due to
caching). That way, you could run it all on an isolated box and really pin
where the issue lives rather than having to wade through all the other stuff
that's surely running on your server. I'd really, really, really start with
*only* the lucene search, then add in other parts of my system until it
broke. Of course, if it broke with just the Lucene you wouldn't have added
much <G>. This approach should also give you a much easier environment to
run a profiler on than, say, a server.

It should be fairly simple to do just the Lucene part, but the part about
"adding other parts of your system" may be prohibitively expensive. But it's
a thought.....

Of course, Andrew's suggestion might make it all unnecessary...

Not a huge help I know, but it's the best I can do.....

Erick

On 12/8/06, Aigner, Thomas <TAigner@wescodist.com> wrote:
>
> I have tried the HitsCollector and the time has improved ~ 3/4 a second
> on the searching.  I still get really bad times when two or more people
> ask for data at the same time.  The problem doesn't seem to be in
> writing the files, it's in getting data from the index when two or more
> people ask for large recordsets back (I can take all the I/O statements
> out and still see the performance bottleneck)
>
> Tom
>
>
> hc = new HitCollector() {
>         BufferedWriter fw2 = null;
>         public void collect(int id, float score) {
>                     try
>                       {
>
>                            Document doc = is.doc(id);
>                            if (fw2 == null)
>                             {
>                                 fw2 = new BufferedWriter( new
> FileWriter( "WHERETOWRITEFILE"), 8196 );
>                             }
>                             fw2.write(doc.get("field1") + "\n");
>                             fw2.flush();
>                             }
>                      catch (Exception ex) {
>
>                         }
>                    }
> };
> is.search(query, hc);
>
>
> -----Original Message-----
> From: Aigner, Thomas [mailto:TAigner@WescoDist.com]
> Sent: Thursday, December 07, 2006 1:36 PM
> To: java-user@lucene.apache.org
> Subject: RE: Reading Performance
>
> Thanks Grant and Erik for your suggestions.  I will try both of them and
> let you know if I see a marked increase in speed.
>
> Tom
>
>
> -----Original Message-----
> From: Grant Ingersoll [mailto:grant.ingersoll@gmail.com]
> Sent: Thursday, December 07, 2006 1:24 PM
> To: java-user@lucene.apache.org
> Subject: Re: Reading Performance
>
> Have you done any profiling to identify hotspots in Lucene versus
> your application?
>
> You might look into the FieldSelector code (used in IndexReader) in
> the Trunk version of Lucene could be used to only load the fields you
> are interested when getting the document from disk.  This can be
> useful if you have large fields that are being loaded that you don't
> necessarily need (thus skipping them).
>
> Also, do you need the BufferedWriter construction and check in side
> the loop?  Probably small in comparison to loading, but  It seems
> like it is only created once, why have it in the loop?
>
>
>
> On Dec 7, 2006, at 1:14 PM, Aigner, Thomas wrote:
>
> >
> >
> >
> >
> > Howdy all,
> >
> >
> >
> >       I have a question on reading many documents and time to do this.
> > I have a loop on the hits object reading a record, then writing it
> > to a
> > file.  When there is only 1 user on the Index Searcher, this
> > process to
> > read say 100,000 takes around 3 seconds.  This is slow, but can be
> > acceptable.  When a few more users do searchers, this time to just
> > read
> > from the hits object becomes well over 10 seconds, sometimes even 30+
> > seconds.  Is there a better way to read through and do something with
> > the hits information?  And yes, I have to read all of them to do this
> > particular task.
> >
> >
> >
> > for (int i = 0;(i <= hits.length() - 1); i++)
> >
> > {
> >
> >
> >
> >       if (fw == null)
> >
> >       {
> >
> >             fw = new BufferedWriter( new FileWriter
> > ( searchWriteSpec ),
> > 8196) ;
> >
> >       }
> >
> >
> >
> >       //Write Out records
> >
> >       String tmpHold = "";
> >
> > tmpHold = hits.doc(i).get("somefield1") + hits.doc(i).get
> > ("somefield2");
> >
> >
> >
> >       fw.write(tmpHold + "\n" );
> >
> >
> >
> > }
> >
> >
> >
> > Any ideas on how to speed this up especially with multiple users?
> > Each
> > user gets their own class which has the above code in it.
> >
> >
> >
> > Thanks,
> >
> > Tom
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
>
> ------------------------------------------------------
> Grant Ingersoll
> http://www.grantingersoll.com/
>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message