I'm currently writing a document using DocBook, and have decided to
seperate the chapters using XInclude. Thus the main file looks like this:

<?xml version="1.0" encoding="ISO-8859-1"?>
<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
   "file:///vol/xml/share/schemas/dtds/docbook-4.2/docbookx.dtd">
  ...
  <xi:include href="chap-intro.xml"/>
  <xi:include href="chap-control.xml"/>
  ...

Each chapter is a complete XML document starting with

<?xml version="1.0" encoding="ISO-8859-1"?>
<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
   "file:///vol/xml/share/schemas/dtds/docbook-4.2/docbookx.dtd">

I have a catalog file that maps the public identifiers to appropriate
places in the filesystem, the same as used as system identifers.

Processing this file with xsltproc takes longer than necessary, because for
every x-included file, all docbook files (DTD modules, entity definitions)
are read over and over again (I've decided not to include the lenghty truss
output). I saw 260 "open" operations for 26 different files, counting only
those associated with the DocBook DTD. I have one master document and 9
chapters, therefor the 1:10 ratio. More chapters or a finer seperation of
the files would yield even worse values.

I think libxml2 and depending tools could benefit from some sort of caching
mechanism, so that each file is read only once.

Joern