[Planetlab-users] limit on number of connections to a port

[Vivek Pai]

Ioan Raicu wrote:

> My "fundemental limit" of 64K connections was refering to 1 single
> machine with 1 single network interface, and having all 64K connections
> simultaneously active.  For every single network interface, you get

This is true only if the machine is acting as the client. If the
machine is acting as the server, it's not the case.

> another 64K connections, that is true.  You say that flows are composed
> of a 5-tuple, but I was refering to raw TCP connections that need a port
> to communicate over, and since there are only 64K possible ports defined
> in the TCP header, you cannot have more than 64K concurent and active
> TCP connections on 1 single interface in 1 single machine.  The bottom

again, if you're the server, it doesn't matter.

> line is that in practice, managing this many concurent TCP conenctions
> is not trivial... so you either have a powerful machine, or you have to
> manage the TCP connections very smart, and it either takes some powerful
> hardware, or

There's been a lot of work on scalable event delivery mechanisms,
and even Linux has it now. If you configure your kernel properly,
you can do this on a $400 machine.

> If you want to explain how the 5-tuple can get you an unlimmited number
> of concurent TCP connections on a single machine and a single network
> interface, I'd be very curious to hear how :)

Run a server, open up a port to accept connections. Have any many client
machines as you want connect to that same port, with as many simultaneous
connections as you want from each client. It's really not that difficult.

However, this is now very far removed from the typical setup necessary
on PlanetLab, so I suggest we move this discussion to private e-mail if
you'd like to continue it further.

-Vivek