While running an internal web services benchmark (Spidermark), we have found
that as the client load increases, the server won't scale due to the default
transformer in Xalan. We have tested 'XSLTC' as the transformer instead and
seen 10x performance gain.

It's difficult to narrow down which part of the code in Xalan is the
bottleneck. We believe that there is overuse of synchronization (at least compared with XSLTC). And many of these synchronized blocks can be eliminated by
redesigning some code (e.g. by replacing static vars by ThreadLocals).

The profile also shows that Xalan uses Vector, ArrayList would be more
efficient.

Please see the OptimizeIt profile at:
http://javaperf.sfbay/benchmarks/spidermark/mprof/getInvoice_IR45.txt

If the default transformer can be changed to "XSLTC", this problem would be
fixed.