Re: initial congestion window for connections in the listen queue

["Ilpo Järvinen" <ilpo.jarvinen@helsinki.fi>]

On Mon, 13 Apr 2009, David Miller wrote:

> From: Octavian Purdila <opurdila@ixiacom.com>
> Date: Mon, 13 Apr 2009 23:46:15 +0300
> 
> > 
> > A question for the TCP wizards:
> > 
> >>struct sock *tcp_create_openreq_child(struct sock *sk, 
> >>        struct request_sock >*req, struct sk_buff *skb)
> >>{
> >>...
> >>         /* So many TCP implementations out there (incorrectly) count the
> >>          * initial SYN frame in their delayed-ACK and congestion control
> >>          * algorithms that we must have the following bandaid to talk 
> >>          * efficiently to them.  -DaveM
> >>          */
> >>          newtp->snd_cwnd = 2;
> > 
> > Shouldn't the same logic from tcp_init_cwnd() be used here?
> > 
> >>From my traces, this seems to prevent TSO from helping short lived 
> > connections.
> 
> On any standard ethernet MTU or larger, you should be getting
> an initial CWND of 3 or 4 because of the logic in tcp_init_cwnd().
> 
> Don't just guesstimate what initial ->snd_cwnd value the
> kernel is using by looking casually at tcpdump traces.  Add
> some kernel debugging printk's and find out for sure.

A long-standing feature in tcp_init_metrics() is such that any of its goto 
reset prevents call to tcp_init_cwnd(). I never remembered to fix that.

-- 
 i.