Actually I didn't use epoll() for Linux::TCPServer. If I were making a little more flexible solution I would have though. For the most part, the significant speed and efficiency gains I got (I still use this module as-is for now in some other proprietary code I'm developing) came from primarily three things:

• Single-port-ness. I only needed to listen to a single tcp port, and quite frankly I think most people using such a module are in the same boat. Single port can be implemented much more efficiently than mutli-port, because you can (at least on Linux, probably most others) just block on accept() in all the children on a shared socket the parent established before the forks without locking anything. It's a big win, and could be done as easily in perl as it was in C.

• Doing this stuff in perl is just inherently an inefficient proposition, whereas doing it in C is inherently efficient if done right. One of my "lessons learned" from this experience is that apparently not many developers of perl and/or perl modules really look at the net effect of what they're doing with tools like strace. Coming from C-land, if you were to take a server like Net::Server::PreFork and run a tcp service over it and strace it, you would be shocked.

  And I'm not trying to pick on Net::Server::PreFork specifically, a lot of perl modules are like that. IO::Socket::INET is relevant and easier to pick on, although Linux::TCPServer doesn't handle replacing it (in my proprietary code, however, I did). You'd be amazed at the system call waste in IO::Socket::INET for a simple tcp connection. It does a real getpeername() system call twice every time it receives a packet (or was it sends, I can't remember now, it was a while back), for instance.

  Or the fact that it actually goes and reads /etc/protocols *two to three times in a row* on every socket object creation to figure out that TCP is protocol number 6 (which hasn't changed in like, decades, on any operating system). This includes incoming socket objects in a server off of an accept() call. If you process 50 connections a second, you're going to open /etc/protocols and scan it for "tcp" 100+ times a second. You don't notice this stuff in small simple applications, but when you're processing large volumes of network connections, it adds up. Some of this only manifests as a result of sloppy overly generic module coding combined with handling the leaky abstractions of perl, combined with attributes of the local system's C library and whatnot. But it's important to look at the big picture for major platforms, and Linux/glibc is definitely a major platform for perl.

• Efficient direct access to shared memory to track and update child state. There are times (like this one) where you know that a very efficient solution for a problem can be made by creating a real shared memory array of integers and indexing it directly from multiple threads of execution. Perl doesn't offer any clean api for this, although a semi-portable XS module could offer it with only a slight efficiency loss. "use threads" + "my @array : shared" doesn't even come close to realizing this, as anyone more familiar than me with perl internals knows. I think Net::Server::PreFork ends up communicating over a socket to it's children, for example, because that's about the best you can hope for in generic platform-abstract perl-land.

So in summary, what C had to offer over pure perl in this case was that it didn't egregiously waste system calls and disk i/o pointlessly for the purpose of ease of abstraction, and it allowed me direct access to real hardware shared memory arrays (which are available on many, many platforms, probably most that perl can run on), but epoll() had nothing to do with it really.