Java IO Faster Than NIO?

The following article on /. piqued my interest: Java IO Faster Than NIO

I read Paul Tyma's slides intently because the subject of "how do you design a massively scalable server?" is one of my favorite topics. There are so many interesting areas to explore in this area.

I have designed several scalable servers over the years. In one case this was for some telco-class equipment that I was helping to design. On that project I ended up using a combination of threads, /dev/poll (on platforms that had this) and poll() (on other platforms). I was under a large amount of pressure to make things fast/scalable on this project, so I was always tinkering around. In the end, I came up with a fairly high-performance communications subsystem that satisfied the goals that we had. The test harnesses that I wrote to go along with this server seemed to validate this as well.

For work on another scalable server that I designed, I went with Java NIO. This is why the article referenced on /. interested me so much. When I started on this particular project, I had no idea of whether to go with a blocking/threads-based architecture or a pure NIO-based system. In the end, I guessed that NIO might be faster so I went with that. The design that I came up with for this server is based on a lot of research that I've done over the years, and this server appears to be very scalable and fast.

I do have some other random thoughts on this topic...

• One of the scalable servers that I have designed happens to interact with SSL/TLS. I have to tell you -- getting a non-blocking I/O layer to interact correctly and efficiently with a SSL/TLS layer is pretty complicated. I spent a huge amount of brainpower and time getting things right here. Given the information that is presented in Tyma's article, if I were to start designing a scalable Java-based server again, I would definitely consider just using Java IO rather than NIO. This would make the code much more simple and straightforward.
• Speaking of making a high-performance communications subsystem, what's the deal with all of the postings on the web from "programmers" who can't figure out how to use their socket API so they insert calls to sleep() in order to "guarantee" that they can read/write data? I must have seen a dozen postings with advice like this. This boggles my mind! A call to sleep() has no place in a high-performance communications subsystem, and anybody who puts one there is not producing good quality code.
• In Tyma's slides, I especially liked the story of Rob Van Behren. When I got to the sentence that read "[w]hen he was done, he sat back and realized he had written the foundation for a threading package" I felt like I had known this gentleman named Rob for a hundred years...