Re: [docbook-apps] Re: Support for callout extensions in xsltproc

[Daniel Veillard <veillard@redhat.com>]

On Thu, Aug 07, 2003 at 10:36:24AM -0400, Jeff Beal wrote:
> > though.  It seems
> > > that Saxon is able to write files faster on Windows then 
> > XSLTProc, so even
> > > though xsltproc is "transforming" faster, Saxon is able to 
> > get the file onto
> > > the hard drive faster.  When I'm going to FO (where there's 
> > only one file
> > > write operation), xsltproc is considerably faster.
> > 
> >   Okay, do a Request For Enhancement on libxslt bugzilla about this.
> > Maybe I can find why writing is slower than expected, I don't remember
> > doing any performance analysis for chunking i.e. 
> > exslt:document extension
> > maybe there is something wrong.
> 
> I'll go ahead and file that.  If you need a large test-doc, I'll figure
> something out.  Our doc is proprietary, so I can't just send it to you
> as-is, but I can run it through an XSLT transform to obfuscate the text.

  Hum, done some preliminary testing but I could not find some clear 
performance problem. On the other hand it's clear that version 1.56.1 of
the stylesheets are way slower than 1.40 which I use for my regression tests.
On my relatively small test document what used to take 3 seconds now
takes 5 seconds (using html/chunk.xsl).
  The HTML serialization cost I initially suspected is neglectable.
XSLT Number formatting primitive used in 1.40 to be the most costly
part and could have been optimized easilly but apparently are not used 
in 1.56.1 and are replaced by XSLT code which seems quite more expensive
to process. I also note that one of the basic steps of XPath evaluation
used to be called 62859 times is now called 183694 times i.e. nearly
3 times more. In geneal is seems there is an increase of 2.5 for the
average number of XPath operations requested to process the same stylesheet.
The template chunk-filename used to be the most costly, with gentext.template
and gentext being second and third, now gentext.template takes the most CPU
time apparently while still called as frequently, l10n.language takes the
second place, and lookup.key the third, all 3 consuming more CPU each than
gentext.template in 1.40
  Seems to me that the stylesheet have become increasingly more complex,
maybe there is a good reason, but it costs a lot.

Daniel

-- 
Daniel Veillard      | Red Hat Network https://rhn.redhat.com/
veillard@redhat.com  | libxml GNOME XML XSLT toolkit  http://xmlsoft.org/
http://veillard.com/ | Rpmfind RPM search engine http://rpmfind.net/