killy971 wrote:

> By repeating this pattern, I made test files with different file
> sizes, from 1MB to 200MB, and tested the XSL transformation process
> (with an XSL file transforming the XML file to an HTML file) with the
> two libraries.

Interesting, be aware that file sizes of 200 MB are
usually too large to be processes with XSLT. It will
work in principle, but should be done offline because
it simply takes too long for interactive use.

> For little XML files (under 4MB), Intel XSLT accelerator performances
> were better than Xalan performances, but for bigger files, XSLT
> accelerator starts to be _very_ slow (exponential growth of the
> processing time).
> 
> An extract of the result of my benchmark :
> 
> 10MB XML file
> - Xalan : 4.5 seconds (processing time)
> - XSLT accelerator : 19.4 seconds
> 
> 30MB XML file
> - Xalan : 11.4 seconds
> - XSLT accelerator : 191.2 seconds
> 
> 50MB XML file
> - Xalan : 18.4 seconds
> - XSLT accelerator : 548.4 seconds (~30 times slower !)

Congratulations for your systematic analysis.
You have already found the problem:

  Xalan run-time increases as O(n), this means roughly 5-fold run-time for 5-fold size.
  XSLT accelerator run-time increases as O(n^2), this means roughly 25-fold run-time for 5-fold size.


This problem wont be solved by "optimization" in the usual sense.
The algorithm has to be changed.