[e2e] admission control vs congestion control
lisongxu2 at gmail.com
Wed Apr 19 07:25:50 PDT 2006
You are right that no one would pay for it. But if it is free, I guess
it is still very attractive compared to the current best-effort
On 4/19/06, Fred Baker <fred at cisco.com> wrote:
> Would you find that acceptable as a service? Would you pay for it? Or
> would you say "this crappy service is unreliable; I would rather go
> to the corner store and rent the movie than wonder whether the rental
> from my ISP is going to actually stay up or maybe have little squares
> all over it half the time"?
> On Apr 18, 2006, at 11:00 PM, Dah Ming Chiu wrote:
> > How about a "quick-brobing delayed blocking" approach?
> > In other words, you only probe for tolerable "post-dial delay" and
> > admit if okay or uncertain. When things get bad, the call may be
> > dropped in the middle. Hopefully with sufficient over-provisioning,
> > the "delayed blocking" happens very rarely. This may be what is
> > already "implemented" in the sense that the "quick probing" part is
> > "no
> > probing" and the "delayed blocking" part is done by the human user.
> > Dah Ming
> > ----- Original Message ----- From: "Henning Schulzrinne"
> > <hgs at cs.columbia.edu>
> > To: "Lisong Xu" <lisongxu2 at gmail.com>
> > Cc: "Fred Baker" <fred at cisco.com>; <end2end-interest at postel.org>
> > Sent: Wednesday, April 19, 2006 4:52 AM
> > Subject: Re: [e2e] admission control vs congestion control
> >> - Delay until admission decision makes many such techniques
> >> unsuitable for applications where humans are waiting ("post-dial
> >> delay")
> >> - Overhead of probing, particularly if the probe has to be
> >> repeated after a failure
> >> - Guarantees tend to be weak, i.e., you may get a positive answer
> >> and still suffer packet loss, particularly if the number of flows
> >> is small and flows are bursty or on-off (such as voice) if the
> >> probe gets "lucky"
> >>> I agree with you that "people don't want to build large amounts of
> >>> state into the network." But there are also admission control
> >>> methods
> >>> that do not build any state into the network, such as probing-based
> >>> methods. Why these methods have not been widely accepted and
> >>> implemented? I guess the tcp friendliness is one of the reasons, are
> >>> there any other fundamental reasons?
> >>> Thanks
> >>> Lisong
More information about the end2end-interest