(I've added comp.os.linux.networking to the mix.)

John Temples wrote:

> I'm writing a server that needs to support 100,000 concurrent, mostly
> idle TCP connections. Once this server receives around 65,400
> connections, it will no longer accept new connections -- client calls
> to connect() time out. Client connections are coming from different
> IP addresses, so I'm not running out of TCP port numbers.
> getrlimit(RLIMIT_NOFILE) and sysconf(_SC_OPEN_MAX) both return
> 250,000.
>
> I can run clients on the server machine and have ~130,000 system-wide
> file descriptors open under the same UID, so it seems like a
> per-process limit rather than a system or per-user limit. Is this
> something that can be configured?
>
> $ cat /proc/sys/fs/file-max
> 296295
>
> $ ulimit -n
> 250000
>
> $ uname -rvi
> 2.6.22.18-0.2-default #1 SMP 2008-06-09 13:53:20 +0200 i386

I don't know the answer to your specific question, but you might find
something in the following article.

http://www.kegel.com/c10k.html