XOM API for Processing XML with Java

[Wolfgang Hoschek <wolfgang.hoschek AT mac.com>]

Hi Galip,

That's a bit odd indeed. Here are some ideas:

Are you sure you are using woodstox-3.0 (rather than some other STAX  
impl)? Did you also try the SAX input route, using the same XML  
parser that's used for FI?

Perhaps wrapping the input stream into a reasonably sized  
BufferedInputStream helps (e.g. 8 KB)? This isn't necessary for bnux  
because it does its own internal buffering, and it's not necessary  
for xerces/SAX either, and I thought it isn't necessary for woodstox  
either, but it can't hurt to try experimenting with that.

How are you measuring the SAX <--> FI route?

Does the driver run a single JVM instance for more than 30 secs  
(ensuring warmup to compiler and VM steady state)? Does the file  
system cache play a role? Try removing actual disk and network I/O to  
focus on the interesting portions.

What kind of input data is being used? Does it use DTDs? Could you  
share the docs with us so we could play with it?

Finally, what's the use case you are working on? For best efficiency,  
perhaps you could try to eliminate reading and writing standard  
textual XML entirelly, using a binary route throughout. In other  
words, if the the sender directly generates binary and the receiver  
directly consumes binary the expensive intermediate file conversion  
step from and to textual XML can be avoided altogether.

Regards,
Wolfgang.

On Jul 20, 2006, at 3:28 PM, Galip Aydin wrote:

> Wolfgang,
> I've been doing some tests to compare BNUX and Fast Infoset  
> performance for various document sizes, my findings are;
> - Encoding: FI (with SAX) is almost twice as fast as BNUX (with  
> stax) ,
> - Decoding: BNUX is 2 to 3 times faster than FI.
>
> So I have no problem with decoding but somehow I can't get good  
> results with encoding, and was wondering if I am doing anything  
> wrong. The fastest way I could measure for xml2bnux among your  
> previous suggestions is this:
>
> serializer = codec.createStreamingSerializer(out, 0);
> redirector = XOMUtil.getRedirectingNodeFactory(serializer);
> StaxUtil.createBuilder(null, redirector).build(in);
>
>
> The sizes of the XML documents are between 500KB to 5MB.
> Any ideas what might be happening?
> Thanks,
> Galip.
>
>
> ----- Original Message ----- From: "Wolfgang Hoschek"  
> <wolfgang.hoschek AT mac.com>
> To: "Galip Aydin" <gaydin AT grids.ucs.indiana.edu>
> Cc: <xom-interest AT lists.ibiblio.org>
> Sent: Tuesday, July 18, 2006 6:57 PM
> Subject: Re: [XOM-interest] BNUX Question
>
>
> > On Jul 18, 2006, at 3:42 PM, Wolfgang Hoschek wrote:
> >
> > > On Jul 18, 2006, at 3:12 PM, Galip Aydin wrote:
> > >
> > > > Hi All;
> > > > What is the fastest way to convert an XML document to bnux and  
> > > > vice-
> > > > versa,
> > > > and does anybody have any examples?
> > > > I see that the last version of bnux is faster than Fast Infoset but
> > > > I want
> > > > to make sure that I am using the right methods.
> > > >
> > > > Here is what I use:
> > > >      Document doc = new Builder().build(new File(XMLpath));
> > > >      BinaryXMLCodec codec = new BinaryXMLCodec();
> > > >      byte[] bnuxDoc = codec.serialize(doc, 0);
> > >
> > > The given code materializes the document as a XOM main memory tree,
> > > requiring memory proportional to the size of the input document. It
> > > is convenient but can be expensive or even prohibitive for large
> > > documents. A more efficient approach is using streaming to convert
> > > arbitrarily large XML documents to and from bnux, requiring only
> > > memory proportional to the maximum depth of the XML tree, i.e.
> > > constant memory for all practical purposes. This is outlined by the
> > > example in section "Streaming conversion of standard textual XML to
> > > and from binary format"
> > > http://dsd.lbl.gov/nux/#Streaming%20conversion%20of%20standard%
> > > 20textual%20XML%20to%20and%20from%20binary%20format
> > >
> > >    // streaming conversion of standard textual XML to bnux binary  
> > > XML:
> > >    InputStream  in  = new FileInputStream("samples/data/ 
> > > weblog.xml");
> > >    OutputStream out = new FileOutputStream("/tmp/weblog.xml.bnux");
> > >    StreamingSerializerFactory factory = new  
> > > StreamingSerializerFactory
> > > ();
> > >    StreamingSerializer serializer =  
> > > factory.createBinaryXMLSerializer
> > > (out, 0);
> > >    NodeFactory redirector = XOMUtil.getRedirectingNodeFactory
> > > (serializer);
> > >
> > >    new Builder(redirector).build(in); // performs streaming  
> > > conversion
> > >
> > >    in.close();
> > >    out.close();
> > >
> > >
> > >
> > >    // streaming conversion of bnux binary XML to standard textual  
> > > XML:
> > >    InputStream  in  = new FileInputStream("/tmp/weblog.xml.bnux");
> > >    OutputStream out = new FileOutputStream("/tmp/weblog.xml");
> > >    StreamingSerializerFactory factory = new  
> > > StreamingSerializerFactory
> > > ();
> > >    StreamingSerializer serializer = factory.createXMLSerializer(out,
> > > "UTF-8");
> > >    NodeFactory redirector = XOMUtil.getRedirectingNodeFactory
> > > (serializer);
> > >
> > >    new BinaryXMLCodec().deserialize(in, redirector); // performs
> > > streaming conversion
> > >
> > >    in.close();
> > >    out.close();
> > >
> > > There's also a small command line test tool that's essentially a
> > > glorified version of said streaming example, allowing to convert
> > > arbitrarily large XML documents to and from bnux. Example:
> > >
> > > # bin/fire-bnux < samples/data/teams.xml > /tmp/teams.xml.bnux
> > > # bin/fire-bnux < /tmp/teams.xml.bnux > /tmp/teams.xml
> > >
> > > See class BinaryXMLConverter for the source code. Previously
> > > mentioned approach uses SAX. Alternatively, you can also use StAX by
> > > replacing the line
> > >
> > > new Builder(redirector).build(in) // performs streaming conversion
> > >
> > > with
> > >
> > > StaxUtil.createBuilder(null, redirector).build(in) // performs
> > > streaming conversion
> > >
> > > Using StAX with woodstox-3.0 is more efficient than SAX with
> > > xerces-2.8.0, in particular for very small documents. For detailed
> > > numbers, see http://dsd.lbl.gov/~hoschek/tmp/plots-stax2.pdf (test
> > > documents are sorted by file size, left to right).
> > >
> > > For reading or writing large numbers of very small documents (e.g.
> > > 200 bytes), it is more efficient to reuse the Builder and
> > > BinaryXMLCodec objects, minimizing setup time.
> > >
> > > Wolfgang.
> >
> > I should add that when writing standard XML it is more efficient  
> > to  use StAX as well (see the plots). In the example above this  
> > can be  done by replacing the line
> >
> > // use XOM serializer:
> > StreamingSerializer serializer = factory.createXMLSerializer(out,
> > "UTF-8");
> >
> > with
> >
> > // use StAX serializer (woodstox-3.0 recommended):
> > XMLStreamWriter writer = XMLOutputFactory.newInstance  
> > ().createXMLStreamWriter(out, "UTF-8");
> > StreamingSerializer serializer = factory.createStaxSerializer 
> > (writer);
> >
> > Again, when using many small documents, reusing the StAX   
> > XMLOutputFactory instance increases efficiency significantly.
> >
> > Similarly, such object reuse on the XML --> bnux route can be   
> > achieved by replacing
> >
> > StreamingSerializer serializer = factory.createBinaryXMLSerializer  
> > (out, 0);
> >
> > with
> >
> > StreamingSerializer serializer = codec.createStreamingSerializer  
> > (out, 0);
> >
> > where "codec" is a BinaryXMLCodec instance that's beeing reused N  
> > times.
> >
> > Wolfgang.