On Wed, 17 Apr 2002, Mikael Olsson wrote:
:Fwiw, I tried convincing the openssh people that a build switch or
:config file option for disabling nagle is a good idea, but they
:just would NOT believe that disabling nagle can actually reduce
:latency.  Rather, they were talking about removing all the NODELAY
:code.  This is where I tried to explain that "no, NODELAY doesn't
:help for a _single_ packet, but things get iffy on packets 2 and 3,
:especially with smart TCP stacks on the other side that delay ACKs",
:but that didn't help either. *sigh*

i don't recall suggesting TCP_NODELAY never be used. also, we don't want
to have a nagle tunable for every possible TCP endpoint and usage
variation.

as i posted earlier, we now disable nagle for TCP and X11 forwarding
endpoints.

for transport my current thinking is:

for the ssh transport connection, we use nagle by default only if it is
a session channel with a pty and shell or exec channel (and agent
forward?)

other channel types would cause connection to be nodelay on both sides
(as determined and set by ssh and sshd)

nodelay yes/no ssh and sshd configuration option would be available to
change default for pty,shell,exec only behavior.

_______________________________________________
Firewalls mailing list
[EMAIL PROTECTED]
http://lists.gnac.net/mailman/listinfo/firewalls

Reply via email to