Tornado failing at high levels was a secondary point. The main point was the performance. I have serious doubts as to whether epoll would make a significant difference there, but I'd love to see someone making the claim prove it.
In fact epoll was created specifically because select does not scale well. That is, if you have lots of open connections, and an event happens on one of them you have to iterate all the file descriptors to see which one was active.
With epoll, only those descriptors are returned which were active. With 8000/s (as in the tests) this really makes a significant difference.
