[e2e] admission control vs congestion control
fred at cisco.com
Mon Apr 17 19:25:44 PDT 2006
To be very honest, it is because people don't think objectively.
Objectively, TCP-like approaches work well for TCP-like traffic,
which is to say traffic that will adjust itself to available
bandwidth, and people are familiar with this model. However, real-
time streams (which include voice, video, and telemetry applications)
send data at a fixed rate or within a fixed range, and as a result
has limited ability if any to adapt to competing traffic or available
capacity. As you say, if you really want to control real-time traffic
congestion in a world of finite bandwidth - however large that
bandwidth might be - admission at bottlenecks becomes critical.
It turns out that there are requirements also that networks be able
to make decisions according to a stated policy. Policies being
discussed today include such questions as "is there enough
bandwidth", "is this guy my subscriber", "is this a subscriber with a
special contract that lets him in when others aren't allowed in", and
"is this a subscriber with a very extremely special contract that
calls for me to push someone else out to let him in?". In that last
case, "and whom do I push out, how do I push out only the ones I need
to, and how do I tell them I have decided to do so" becomes pretty
important. TCP-like approaches provide no such possibility.
There is the ECN thing being proposed by BT and Nortel in the IETF.
There are a number of interesting assumptions built into that along
these lines: that new calls come in at a rate less than some
threshold, above which matters become far less deterministic, that
the probe traffic is not seen as competition to existing calls to a
sufficient extent that existing calls come down, and that edge
systems cooperate with the network in certain ways. Network operators
of my acquaintance don't trust the edge...
The usual reason given for avoiding admission models is that people
don't want to build large amounts of state into the network. Having
said that, network operators then build MPLS or other circuit-switch
infrastructures, and perhaps engineer those routes to maximize the
traffic they can send over them or to maximize their ability to
recover cleanly from failures. This involves a *lot* of state in the
network, much more than bandwidth admission techniques call for.
People avoid bandwidth admission techniques, and love to hate them,
because they don't think objectively. If they thought objectively,
they would find they were the simplest approach to the problem,
involve the least state, the least heuristics, and where heuristics
are required, the most predictable ones.
I, of course, have no opinion :-)
On Apr 15, 2006, at 2:10 PM, Lisong Xu wrote:
> Hi everybody,
> Both congestion control-based and admission control-based
> approaches were proposed for multimedia streaming over the Internet
> a few years ago. Intuitively, admission control is more suitable
> for multimedia streaming than congestion control, since admission
> control can provide some guaranteed bandwidth.
> However, it seems that congestion control-based approaches are
> accepted by more people. I am wondering what the fundamental reason
> is. Is it because congestion control-based approaches are tcp-
> friendly, and then safer to implement in the Internet?
> Thanks in advance!
More information about the end2end-interest