1. The author is Tim Deegan, who's a former Xen maintainer and currently does low-level work in a data storage startup. He holds a PhD from a reputable university (Cambridge...) and has published several papers (http://www.tjd.phlegethon.org/). This doesn't immediately make him right of course, but it does mean he's probably not an idiot and his statements merit a little consideration.
2. The bottom paragraph of the article explains his reasoning behind hoping for a recursive version in interviews: "When I actually implemented this in C, the iterative version was not only messier, but also nearly four times slower than the recursive one. ... a modern CPU will predict all the return addresses correctly. And the work of figuring out which of the two candidate disks to move next and where to is quite a bit more than just finding the spare peg. Since any Towers of Hanoi problem that will complete in less than a day should fit entirely in L1 cache, the extra instructions and (potentially mispredicted) branches to find the next move are worse than the extra memory operations for the recursion." It's not clear from some of the comments in this thread that many people read all the way down to that paragraph.
3. His test case is here: http://tjd.phlegethon.org/blog/hanoi.c ... if he has done something wrong, this would be the place to go looking for it. I don't see anything obviously wrong with his iterative version, although I'm getting a double-free error after compiling and running his code.
I've believed for years that iterative approaches gave better performance than recursive versions. This probably goes all the way back to working in 68k assembly. I'm pretty sure there's a chapter in Beautiful Code that suggests iterative is better than recursive, performance-wise. But if Tim's right, then I've learned something valuable today.
edit: I'm busy today and my C is suuuuper rusty; I think there might be something in that for loop in finish() that's screwing up the pegs[2] that's being passed to free() a couple of lines later, but I don't really have the time to debug it. Hopefully someone else is willing to try, I'd like to see what the actual results turn out to be in different environments.