netdev
[Top] [All Lists]

Re: ECN timeout support in Linux TCP?

To: <kuznet@xxxxxxxxxxxxx>
Subject: Re: ECN timeout support in Linux TCP?
From: jamal <hadi@xxxxxxxxxx>
Date: Sun, 29 Apr 2001 17:54:12 -0400 (EDT)
Cc: Bartok Istvan <bartoki@xxxxxxxxxx>, <netdev@xxxxxxxxxxx>
In-reply-to: <200104291709.VAA03686@ms2.inr.ac.ru>
Sender: owner-netdev@xxxxxxxxxxx

On Sun, 29 Apr 2001 kuznet@xxxxxxxxxxxxx wrote:

> Hello!
>
> > The short question: Does the Linux TCP implementation support the ECN
> > timeout suggested in rfc2481?
>
> No.
>
>
> > Without this mechanism, one should need an (MSS * num_of_connections)
> > size txbuffer on a bottleneck link to totally avoid packet losses even
> > when marking every packet with ECN CE. :(
>
> Grr... goal of ECN is not "avoiding of losses", its goal is preservation
> of ACK clock in presence of congestion. Artificial timeout has nil impact
> comparing to normal timeout.
>

The MSS * num_of_connections arguement is equivalent to the "if you need
every flow through the bottleneck link to use their fair share of the
bandwidth then you preallocate BW * RTT * num_of_connections buffers"
Reality is: loss avoidance is gravy for ECN and that at the level of one
packet cwnds, with so many competing flows, it's a lost cause already.

>
> Note also one more thing (which makes threshold twice worse and, probably,
> should be changed unlike above): we do not allow to drop cwnd under 2 segments
> without real losses. The reason is the same: dropping cwnd under 2 segments
> also breaks ACK clock due to delacks and therefore it is unacceptable.

Aye, aye.
I dont think you passed this feedback to the RFC which is going
proposed standard next week ;-> But we dont wanna delay the publication,
one could argue it is a implementation detail which Linux ignores.

cheers,
jamal


<Prev in Thread] Current Thread [Next in Thread>