GeoServer developers are busy writing and improving code, so have scant time to run lots of performance and scalability tests. We test on our own, to make sure that it does scale and is quick, but formalizing the tests and publishing the results is a bit hard for us. So we're hoping that users will contribute their various findings to this page. Please post any kind of testing that you do with geoserver, and be sure to include the details of OS, servlet container, java version, size of dataset, processor speed, RAM, and anything else that might be useful. Comparisons in particular are quite nice. One thing we would love to have benchmarked is the JAI rendering vs. MapServer, as to our eyes the latest improvements have it performing quite nicely, but we don't have any time to do a bunch of tests.
From Greg Cockroft (these were done before JAI improvements, I expect png and jpeg should be faster now):
Redhat 7.3 on dual 2 Ghz processor
400x400 pixel area zoom level 8
800x800 pixel area zoom level 16