Re: [xml] Performance of DOM interface

[Matt Sergeant <matt sergeant org>]

On Saturday, August 24, 2002, at 07:12 PM, Nick Wellnhofer wrote:

> Daniel Veillard wrote:
> > On Sat, Aug 24, 2002 at 03:13:08PM +0200, wellnhofer aevum de wrote:
> > > Hi,
> > >
> > > I'm using libxml and libxslt with Perl to generate dynamic HTML 
> > > pages and did some performance testing recently. I created a XML 
> > > tree from a Perl data structure using the DOM interface and 
> > > transform that tree with XSLT.
> > >
> > > While the performance of the XSLT transformation is really good, I 
> > > found that constructing the DOM tree took about 75% of the running 
> > > time. So it takes about three times longer to create the XML source 
> > > tree than to parse a stylesheet and apply and output the 
> > > transformation.
> > >
> > > Is this behavior normal? I would think creating an XML tree is much 
> > > faster than a XSLT transformation.
> >   Well usually when creating the input tree using directly the C 
> > parser,
> > the parse time (which includes building the DOM tree) is quite 
> > smaller than
> > the tranformation time itself. So no it's not "normal", though it 
> > also depends a lot on the stylesheet used too. You're probably paying 
> > the cost
> > of going back and forth between the Perl and the C library for each 
> > call to DOM entry points.
>
> Yes, that seems to be the only explanation I can think of. The XSLT 
> transformation needs about 5 calls to the C library, but the creation 
> of the DOM tree requires hundreds of calls of createElement and 
> appendChild.
>
> It might even be faster to generate XML directly in Perl and write it 
> to a file or memory buffer and then parse it with libxml again. I 
> think I will try that, although it seems like a stupid approach.
>
> The best solution is probably a single Perl XS function that converts 
> a whole Perl data structure to a DOM tree in one go.

Yeah, it's all the method calls, and also the memory allocation and the 
reference counting that the Perl level API has to do, but the C one 
doesn't. It *might* be quicker to do it in perl-space and create an XML 
string and then parse it.