How to build a server that accepts many thousands of persistent network connections?

Question

I am about build a server on linux (I get to pick programming language) that accepts many TCP/IP socket persistent connections from a desktop software. How can this be done cheaply and efficiently? A machine cannot have more than 60000 ports, so if I have to support 600k connections then I will need 10 linux boxes?

Since the computation needed for each connection is extremely small (idle 95% of the time) one linux box can already handle 600k I don't want to waste money and resource running 10 boxes just to get around port limitations.

Any ideas?

Answer 1

A connection is identified by the following tuple: [Server IP;Server Port; Client IP; Client Port], hence, more than 60,000 ports are available per machine.

Erlang effectively handles 10's of 1000s of connections. Have a look at this comparison between Apache and Erlang Yaws.

Answer 2

You do NOT need a port for each connection. An HTTP server will just use one port (typically 80).

Answer 3

Since any relatively widely used language has a reasonably efficient socket library I would go for a language that:

1. Is efficient in whatever the processing you need to do with the data from the desktops
2. You are familiar enough (read; most familiar) with to avoid the most common pitfalls / design patterns

Answer 4

You should be able to do this easily with any framework which is decent. We use Twisted which is good (Assuming you understand python and asynchronous programming), if a little weird to understand.

Asynchronous programming is pretty much a given, because you will probably NOT want to start 600k threads.