Log in

May 16th, 2008

The End Of The World (As We Know It)!

Ok, here we go:

Event-driven non-blocking I/O isn't the way anymore for high-performance network servers, blocking I/O on a bunch of threads is better now.

Wow, I can't believe I just wrote that! Here's a post that describes some of the reasons (this is talking more about Java, but the underlying reasons apply to C++ as well, it's not just JVMs getting wackier at optimizing locking). It depends on your platform (things don't change from being true to being false just out of the blue!), and more specifically, I have NPTL-based Linux 2.6 in mind, at the very least (NPTL is needed for better futex-based synchronization, and 2.6 for the O(1) scheduler and low overhead per thread). You also want to specify the smallest stacks you can get away with, and you also want a 64-bit machine (it has a bigger address space, meaning it will explode later).

The most important thing you need is to think and not be an idiot, but that's not really new.

And when I say "bunch of threads", I really mean it! My current "ideal design" for a web server now involves not just a thread per connection, but a thread per request (of which there can be multiple requests per connection)! Basically, you want one thread reading a request from the socket, then once it's read, fork it off to let it do its work, and have the writing of the reply to the socket be done on the request thread. This allows for as much pipelining as possible.

Still, event-driven I/O is not completely useless, it is still handy in the case of protocols that have long-lived connections which stay quiet for a long time. Examples of that are IRC and LDAP servers, although it's possible that with connection keep-alive, one might want to do that with an HTTP server as well, using event notification to see that a request is arrived, then hand it back to a thread to actually process it.

I also now realize that I was thinking too hard in my previous thoughts on using multiple cores. One could simply have a "waiting strategy" (be it select() or epoll), and something else to process the events (an "executor", I think some people call that?). You could then have a simple single-threaded executor that just runs the callbacks right there and then, no more fuss (think of WvStreams' post_select()), or you could have a fancy-pants thread-poll, whatever you fancied. I was so proud of my little design, now it's all useless. Oh well, live and learn...