This is pretty old, right? These days 'Fast Servers' refers to bare-metal network interface without a kernel which can achieve far higher throughput than passing it through a kernel/epoll.
The effect basically disappears when you're looking at webpages that take 10-1000ms to generate.
The article is talking specifically about "simple one-page performant servers that easily get into the 100k requests/second territory", which can see very large throughput increases with direct NIC access.
I have considered writing a userspace TCP HTTP server like the one you suggest, but it would be a demonstration of how to write a state machine and not how to better use the facilities that the operating system provides for writing a server.