How to improve the build speed with saxon 6.X / docbook

[Sylvestre Ledru <sylvestre.ledru@inria.fr>]

Hello,

I am currently trying to improve the build time of the documentation of
a free scientific software (Scilab).

There are almost 1800 XML files. The size of these files is between 1 k
to 10 k.
Before calling saxon, some processing is done (mathml => png through
jeuclid, etc) and finally merged all of them into a single file [1].
This file is processed against chunk.xsl or javahelp.xsl from
docbook-xsl. Both are taking a long time (pretty much the same).

However, the build time is way too long (between 30m to 60m on a
powerfull computer to hours on a small CPU). Especially for some small
architectures like s390 or armel... For example, Debian compilation
chains are killing the process since it is taking more than 150 minutes,
just to "load" the XML.

Therefor, I am trying to improve the speed of the process.
I wonder if there are any tricks to improve the speed. Some people told me that the merge of all xml files
is not necessary but I haven't been able to find how to do it.

I was wondering if there is a better way to structure the XML document.

For now, it is (mainly) structured the following way (by merge of all xml files):
<book>

<part>
<title>title of the chapter 1</title>
<refentry>
Details about the function
[...]
</refentry>
<refentry>
Details about the function 2
[...]
</refentry>
</part>

<part>
<title>title of the chapter 2</title>
<refentry> 
[...]
</refentry>
</part>
</book>

Some rare refentry are also stored in some <chapter> <section>.

There are quite many links between all the refentry (especially coming
from the "see also" section). 

Does anybody know how to improve this ?

Note that the PDF or PS generation is very fast and based on the same
master xml file.

Many thanks,
Sylvestre
PS: I sent this email on the saxon mailing list. They told me that this is most probably due to docbook and not saxon.
[1]
http://www.scilab.org/team/sylvestre.ledru/master_en_US_help-processed.xml.gz