Through forking + gevent and then sleeping in each request handler. Of course it measures nothing other than a whole bunch of while loops running in one fork per CPU waiting for just about nothing. In other words, I'm benchmarking "how much memory do I have", which is pointless. But it sure does scale!