I can't offer to do a detailed performance analysis for you. That's consultancy which I would charge for. If the XSLT code isn't too long, however, then it's worth posting the code, we can have a quick glance at it to see if there's anything glaring.
I wouldn't normally start by measuring throughput, I would focus first on latency (elapsed time for each transformation). However, always measure multiple transformations to ensure that you eliminate Java VM warmup time. Also, make sure you aren't compiling the stylesheet each time you run it. The throughput targets you cite are quite attainable but of course it depends on the size and complexity of the transformations.
Author, XSLT 2.0 and XPath 2.0 Programmer\'s Reference