On Tue, 12 Sep 2000, Thomas Charron <[EMAIL PROTECTED]> wrote:
> Guys, I have a question that some here may be able to assist with.
> Basically, it revolves around the Linux kernel, and adding the capability for a
> very large number of simo IP connections. We need to be able to have a linux
> box support up to and hopefully over 20k simo TCP connections. You can see a
> copy of the challange we've issues at http://jabber.org/article/34.html. The
> biggest issue here is, we'd like to try to decrease the amount of RAM required
> to host this many sockets. Nearly every other OS we've compiled and tested on
> can do it. Basically, we have many persistant sockets, all of which are, for
> the most part, idle. Yes, I know, UDP, etc, instead, we've all had the
> arguments. For now, the persistant sockets are required. The question ends up
> being, can we make a Linux kernel support 20k TCP connections, without
> requiring the kernel to consuming anywhere from 400 Megs to over a gig of ram?
How much RAM usage for 20K sockets would be acceptable to you?
Back in 1998 I saw a talk by David S. Miller where he said he was
optimizing (speed and memory usage) the Linux networking code and was
doing tests with 40K sockets (likely over loopback but I don't know for
sure). He may have been optimizing memory consumption for TIME_WAIT or
CLOSE_WAIT connections (e.g. http), which won't help you if I
understand your application (your connections stay up for a long time, like,
well, say, telnet).
Also, would setsockopt(2) for SO_SNDBUF and SO_RCVBUF (or something
else) be of use? I wrote some grotty code that calls getsockopt() and
it suggests for SOL_SOCKET these buffers are 65535 (kernel 2.2.14)...
I see this number in /usr/include/linux/skbuff.h, but it is likely I
made some mistake, if you play with getsockopt let me know what you
find.
Karl Runge
**********************************************************
To unsubscribe from this list, send mail to
[EMAIL PROTECTED] with the following text in the
*body* (*not* the subject line) of the letter:
unsubscribe gnhlug
**********************************************************
