For the sake of curiosity, I happen to have recently done a benchmark of a "hello world" rack app (literally just responds "hello world" to every request) on a number of Ruby servers (mostly JRuby, but also Puma on MRI).

They were all run in production mode with logging disabled, etc.

http://polycrystal.org/~pat/scratch/microbenchmark.png

Note that a difference from 10k requests per second seems huge compared to 3k, but if you invert it, you get 100 and 333 micro seconds per request, respectively. In a real, non-"hello world" app, these differences are going to be negligible.

Though perhaps it would be more interesting if instead of just responding "hello, world", the app parsed some query parameters or something. But I was mostly interested in the overhead of different JRuby servers, not comparing different app servers (i.e. overhead from Sinatra should be more or less identical whether you're on Puma, Trinidad, or whatever).