[Iccrg] Heresy following "TCP: Train-wreck"
rbriscoe at jungle.bt.co.uk
Wed May 14 11:53:14 BST 2008
Realised I never replied to this (catching up on old emails)...
At 19:40 08/04/2008, Lachlan Andrew wrote:
>On 08/04/2008, Bob Briscoe <rbriscoe at jungle.bt.co.uk> wrote:
> > Just to be clear - we're not proposing charging for congestion like Kelly.
> > It's an important difference. Like the difference between volume charging
> > and volume capping.
>I didn't mean to misrepresent you, but to me, that's an IETF-level
>"implementation detail", rather than an IRTF-level issue. To be
>maximally useful, it should be something which affects the
>users'/applications' behaviour to avoid peak times.
>The key is that the quantity you are both sharing out (however you do
>it) is *congestion* rather than rate or volume.
True, but even we agree it should be possible to make transports
accountable for causing congestion, this consensus on the _metric_
would lead very much into a knock-on _protocol_ issue:
"However you do it" is no good if there are zero pratical ways to do it.
The IRTF can't propose a metric for networks to arbitrate between
users if networks can't see the metric (given current Internet protocols)...
The Internet was designed so only endpoints could see congestion -
it's really hard for the network to intercept congestion feedback
robustly. And if networks do use the feedback, it will just encourage
receivers not to send feedback.
ECN helps, but even them, networks can only make receivers
accountable for the congestion sent to them. Even with ECN, networks
can't see congestion up ahead when accepting packets from the sender.
But networks can't cap the amount of congestion a receiver receives.
It's too late by then. Currently networks can only charge for ECN
received (as Kelly suggested). At first sight that looks like it
creates the right incentives, but I don't believe the IRTF should
propose something that _requires_ a very specific business model.
Plus it would just lead to 'denial of funds' attacks.
Hence the need for a protocol to make _senders_ accountable for
causing congestion. That's a really hard protocol problem.
> > Finally, I don't think 'we' have to use WFQ as an interim solution -
> > fortunately 'we' don't make that choice - operators have protected a better
> > form of fairness than TCP's on the Internet with WFQ for many years.
>OK, but "we" have the choice of letting loose algorithms which will
>increase pressure to deploy it where it isn't already enabled, don't
>Lachlan Andrew Dept of Computer Science, Caltech
>1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
>Ph: +1 (626) 395-8820 Fax: +1 (626) 568-3603
Bob Briscoe, <bob.briscoe at bt.com> Networks Research Centre, BT Research
B54/77 Adastral Park,Martlesham Heath,Ipswich,IP5 3RE,UK. +44 1473 645196
More information about the Iccrg