On Mon, 12 Dec 2005, Zubin Dittia wrote:
I'm trying to find out what kind of bounds the ephemeral ports allocation method in the kernel imposes on the maximum number of connections out of a given machine. Specifically, if there are 32K ephemeral ports configured, then does that mean: (a) I can have 32K connections open total, regardless of server (b) I can have 32K connections open per server (to which we connect) or (c) I can have 32K connections open per <server,port> (to which we connect)?
(a) including connections in TIME_WAIT.
I know that theoretically a connection is identified by its 4-tuple <SA,SP,DA,DP>, so in theory I should be able to establish a very large number of outgoing connections, much higher than the ephemeral port range would suggest, but I'm not sure how this has been implemented in practice.
Correct, but the simplest possible path is currently taken and all ephemeral port assignments works the same way (looks for a completely unused port in the range).
I did try to look into making a patch for improving this on connect() to look for a free tuple rather than just port, but got a bit tangled up in the different layers where port and IP is assigned. It is defenitely technically doable, it's just that the kernel won't do it for you automatically yet..
Regards Henrik - : send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
