Saxon diaries :: Push and Pull parsing

Previous:	Thoughts on schema-awareness
Next:	Comparability

Push and Pull parsing

by Michael Kay on Sat 11 Feb 2006 21:03 GMT | Permanent Link | Cosmos

Saxon has supported the StAX interface for pull parsing for a while now, but I've seen little evidence that people are using it; and in fact generally the scene seems to have gone rather quiet. There's talk that a future JAXP release will require support for StAX as well as SAX input, and on .NET the standard System.Xml parser is a pull parser, so I thought I'd take a look at relative performance - perhaps the time might even come when it's appropriate to use a StAX parser as the default.

I wrote a little program to build a Saxon tree using various parsers: two pull and two push. These were the results, all timings in milliseconds:

Input size (Mb)    Xerces  Piccolo  SJSXP  Woodstox
     6.8            1200     1320    1270     1030
    14.0	    2400     2670    2640     2033

Xerces and Piccolo are SAX (push) parsers; SJSXP and Woodstox are StAX (pull) parsers. So there's certainly no reason from these figures (or from any other considerations really) to believe that pull is intrinsically faster than push. But it's interesting (a) that Woodstock does appear to have a demonstrable edge (20%) on all the others at least on this sample -- one would have to test a variety of other XML documents to be sure that it's sustainable, and (b) that Piccolo doesn't beat Xerces in this test, though they claim to be twice as fast in other tests.

I'm not recommending that everyone switches over immediately to Woodstox - as often happens, they seem to have achieved some of the speed by taking short cuts on conformance, especially checking the detailed rules for valid names and valid characters. But if the 20% matters to you and you know the XML is valid already, it may be worth considering.

Posted to:

Main Page

Comments

Post a comment

Re: Push and Pull parsing

by Cowtowncoder on Mon 13 Feb 2006 17:59 GMT | Profile | Permanent Link

Hi Michael! Thank you for your results -- although results between the parsers in questions varies quite a bit depending on types of documents, your results are well within range I have seen. I was also initially surprised by the good performance of Xerces, but it has been consistently amongst fastest implementations, esp. for bigger documents. Sjsxp is also based on Xerces internals (I think; using XNI interface to low-level parsing aspects), which would indicate that sjsxp authors considered it well-performing.

As to conformance, Xerces still has the best overall xml conformance, but from my testing with xmltest/saxtest, Woodstox appears at this point more conformant than Piccolo (75% pass rate vs. 50% pass rate; Xerces had ~90%). There is also on-going work to really close down remaining uncatched well-formedness issues (esp. regarding less commonly encountered cases -- basic WFC as nesting constraints and namespace collisions have been checked since 1.0), so I am hopeful that the performance difference is not just based on omitting checks that Xerces does (esp. checking of names should be fully xml 1.1 compliant, but not yet 1.0 compliant -- upgrading those should not significantly degrade performance.

Anyway, I hope you will found it useful to use pull parsers for certain things -- I do know that for some tasks SAX is perfectly adequate (building full DOM[-like] trees, and by extension activities that rely on in-memory trees, which is probably the case for many XSLT tasks), but for others pull parsing (data binding, serialization) has its benefits.

Main Page

Photos

RSS Newsfeeds

Main Page RSS