[tcpm] [Iccrg] Re: ECN feedback discussion

[Emmanuel Lochin]

On 9 November 2012 18:56, Mikael Abrahamsson <swmike at swm.pp.se> wrote:
> On Fri, 9 Nov 2012, Emmanuel Lochin wrote:
>
>> Ok, ECN might be not used inside the whole network but who cares? All
>> my core devices are ECN compliant so why not using inside my own
>> network?
>
> What core devices are that? When I last looked at platforms such as Cisco
> CRS-1 / ASR9k, Juniper MX/P-series, Huawei 5000, none of them supported ECN.
> This was a few years ago, so I'd be happy to be enlightened about progress!
> The only platform I have encountered so far that does ECN are Cisco
> processor based platforms such as the 7200.

Hi Mikael

Well ... I should precise that I use only GNU/Linux or *BSD. For ECN
marking I use tc on GNU/Linux.

>> We know, following "The power of explicit congestion notification", A.
>> Kuzmanovic, Sigcomm05, that even a partial deployment of ECN brings
>> out benefit for the end-user. My GNU/Linux stack does not react to
>> ECN? Not a problem ... my access network does.
>
>
> What devices are those?

All my routers are Linux or BSD boxes (HP Proliant blades or simple
PC), obvioulsy I need to tune both BSD or Linux kernels to be able to
correctly manage the trafic.

>
>> On my side, I use specific mechanisms conjointly with ECN inherited
>> from the QoS epoch. For instance, I can shape ECN feedbacks to slow
>> down connexions or to give a delay penalty to a specific flow.
>> Basically I use ECN like an "alert".
>
>
> But how many flows do you have that are ECN enabled? What host systems uses
> ECN by default?

All are GNU/Linux boxes except some *BSD. Most of them are used by
students/researchers that access to simulation servers or educational
ressources.
All my network is ECN-enabled. I have also used to enable an ECN path
over a link between Reunion Island university and Paris
VI university in 2004 and paced the traffic following the marking with
a penalty box.

> I have this on my linux boxes:
>
> $ grep -i ecn /etc/sysctl.conf
> net.ipv4.tcp_ecn = 1
> net.ipv6.tcp_ecn = 1
>
> (I believe the ipv6 line is redundant, just put it in there to be future
> proof). I do not believe this is the standard setting? Please correct me if
> I'm wrong.

This is the standard setting. Today kernels enable CUBIC and ECN. But
as far as I know, CUBIC does not react to ECN.

Manu

>
>
> --
> Mikael Abrahamsson    email: swmike at swm.pp.se