Date:	Fri, 7 Nov 2008 14:22:18 +0200 (EET)
From:	"Ilpo Järvinen" <ilpo.jarvinen@...sinki.fi>
To:	Mikael Abrahamsson <swmike@....pp.se>
cc:	David Miller <davem@...emloft.net>, daniel.blueman@...il.com,
	LKML <linux-kernel@...r.kernel.org>,
	Netdev <netdev@...r.kernel.org>, linux-net@...r.kernel.org
Subject: Re: time for TCP ECN defaulting to on?

On Fri, 7 Nov 2008, Mikael Abrahamsson wrote:

> On Fri, 7 Nov 2008, Ilpo Järvinen wrote:
> 
> > I think you partially miss the point here. In many cases not every single
> > router has to _support_ ECN to get its benefits, not-supporting is not the
> > problem in itself (though it would be nice to get that "fixed" as well) but
> > breaking ecn-enabled connections. I suppose you didn't check that aspect?
> > I'd guess those mentioned devices will interoperate just fine since one can
> > mostly connect ok with ecn too besides rare exceptions rather than things
> > being vice-versa.
> 
> I don't understand. My point is that most of the ISP core equipment out there
> doesn't act on ECN rendering it mostly useless. The N in ECN renders useless
> because there is no device doing the *notification*. They'll just pass the
> traffic without acting on it differently regardless if ECN is on or off.

Likewise, not enabling ecn renders any device doing notification useless.

One alternative to full enable would be enable it for the listening 
allowing client end to decide but that is effectively same as not enabling 
it at all (using the same logic as above).

> > The most crucial components are anyway the points of congestion, I don't
> > know enough isp topologies but I suppose those core routers are not the ones
> > where towards subscribers device traffic congests?
> 
> There can be congestion anywhere in the network, best would be if all routers
> supported it.

I agree. But...

> My problem with ECN is that the most advanced routers do not
> support it, it's useless with L2/L3 switches (as they have very small buffers,
> there is "nothing" to do WRED on), so that leaves potential implementation by
> either DSLAM/BRAS vendors (where Cisco BRAS does support it but it needs to be
> enabled by the ISP) or the SOHO devices which run Linux and might implement
> it, but I'd rather see them do active queue management at all (fair-queue for
> instance) before asking them to do ECN. Of course, if users start to ask for
> ECN and we get fair-queue at the same time, all the better. One very common
> congestion point is definitely the upstream connection of someones cable or
> DSL modem.

...I'd assume that marking on end(s) of the cable/dsl link which is often 
congested is the most low hanging fruit, and like you seem to say (if I 
understood you correctly), there exists some support already which could 
then incrementally be turned on. Just realized though that it might, after 
all, not be what isps like doing since just getting higher bw 
subscriptions is perhaps more rewarding from their perspective (please 
don't take this as an offense, it's not meant to be one :-))...

Imho, the end-users end has less gain since those packets usually travel 
just locally before getting dropped but ecn-aware streaming might change 
that as well.

> > I doubt it any worse than with eg. timestamps.
> 
> According to <http://www.imperialviolet.org/binary/ecntest.pdf> it's 0.5% of
> hosts that drop packets when ECN is enabled. It's a substantial part of the
> Internet. Yes, not doing blackhole detection might get these hosts fixed
> faster, but at the expense of more end user hurt.

Did you notice:

However, the failing hosts are not distributed randomly. The
7,627 hosts span only 4,613 /24 subnets. The top twenty
subnets account for 778 failing hosts (10%). WHOIS[1] information
for 18 of those 20 subnets suggests that they are located in China.

Just some bigger device(s) perhaps?


-- 
 i.

Powered by blists - more mailing lists