The paper doesn't dispute that you can write good, high-performance servers with event-based programming. It's disputing the claim that you can't do it with threads, and then claiming that threads are an easier model for programming servers.

Here's a thought experiment: if your language of choice offered threads so lightweight as to be almost free, with proper preemptive scheduling, not limited to a single core, with the I/O using something like epoll or kqueue underneath -- would you use it?

Erlang and Haskell offer this right now. Python can give you a passable approximation. It really is as cool as it sounds.