lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael McCandless <luc...@mikemccandless.com>
Subject Re: Speed up payload loading?
Date Tue, 03 May 2011 10:31:01 GMT
On Tue, May 3, 2011 at 5:35 AM, Chris Bamford
<chris.bamford@talktalk.net> wrote:
> Hi,
>
> I have been experimenting with using a int payload as a unique identifier, one per Document.
 I have successfully loaded them in using the TermPositions API with something like:
>
>    public static void loadPayloadIntArray(IndexReader reader, Term term, int[] intArray,
int from, int to) throws Exception {
>        TermPositions tp = null;
>        byte[] dataBuffer = new byte[4];
>        int intVal = -1;
>
>        //System.out.println("loadPayloadIntArray(" + from + " -> " + to + ")");
>        try {
>            tp = reader.termPositions(term);
>            tp.skipTo(from);
>            do {
>                tp.nextPosition();
>                tp.getPayload(dataBuffer, 0);
>                intVal = PayloadHelper.decodeInt(dataBuffer, 0);
>                intArray[from] = intVal;
>            } while (++from <= to && tp.next());
>        }
>        finally {
>            if (tp != null) {
>                tp.close();
>            }
>        }
>    }
>
>
> This works fine, it even allows me to use several threads in parallel to each load a
part of the array, thus reducing the time taken.
> However, performance is not as great as I would like - with indexes of ~4m docs it can
take over 3 seconds.  (These are cold start times, no reader cacheing and OS read caches
completely flushed).   Increasing the number of loader threads (calling the method shown)
doesn't help much, which makes me suspect there is some synchronized stuff going on inside
TermPositions.

Since it's cold, likely your bottleneck is the IO system, and that's
why adding threads doesn't help?

> 1) Because this approach is still faster than using FieldCache, my belief is that payloads
are stored sequentially in the index - is this true?

They are... they are encoded into the positions (*.prx) file.

> 2) If so, is there some raw, lower level I/O I can call which will allow me to load much
bigger chunks of payload rather than a single int at a time?

Not really... I mean you could try to parse the bytes yourself from
the prx file, but Lucene doesn't have existing APIs to bulk load from
here.

In trunk we are working towards faster bulk-load APIs, but, how to
make payloads work "well" there is an open question...

Furthermore, there is a new feature branch (docvalues), which aims to
handle your use case ("store X per document") well.  Maybe check out
that branch?

Mike

---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message