[BUGS] Sysctl knobs for 'naggling' timeout (TCP_NODELAY sockopts not set)

[Jerahmy Pocott]

Hey gang.. Umm I guess the subject is a little confusing..

What I'm wondering is if/what sysctl knobs exist for changing the  
timeout on 'naggling'...

For those that don't know what the heck I'm talking about, 'naggling'  
is what happens to any INET socket you create without the TCP_NODELAY  
flag set. Instead of sending data as it's ready the tcp stack tries to  
'fill' a whole packet, this works in simple terms by appending the  
data together while you already have packets 'in-flight'. So the  
network layer either waits until a full packet is ready or until the  
other host sends an 'ACK' for all the data you sent previously.

If neither a full packet is created or an 'ACK' is returned it will  
time out and sent what ever data is ready. I'm looking for a way to  
adjust the time it waits to break a dead-lock situation.

The deadlock happens when the target host is using 'Delayed  
ACK' (sysctl net.inet.tcp.delayed_ack). Since the host will NOT 'ACK'  
on every packet, but rather wait until it gets x packets then 'ACK'.  
So on a socket that didn't set TCP_NODELAY you can end up with high  
latency. Some stuff (like samba) let you choose if you want the socket  
to use TCP_NODELAY in their config or command line options, but then  
you get hit with stuff that won't let you tell it to use the flag and  
a host on which you don't have access to disable delayed ACKing on and  
it seems the only solution would be to reduce the timeout before it  
sends the data, preferably to 0ms..

But looking at the net.inet.tcp knobs, none of them intuitively say to  
me they control this =p

So I'm not sure if there is anything I can use to control it...?