Receiver-driven Layered Multicast Review

"Receiver-driven Layered Multicast" describes a system that uses a layered
source coding algorithm combined with a layered transmission system to
create an efficient IP multicast system that is capable of adapting to
heterogeneous network conditions at the receivers. No changes are required
to the existing multicast networks to implement this system.

The system is centered around the idea of shifting the burden of adjusting
the transmission rate from the source, which can never set a target rate
agreeable to all receivers, to the receivers which can adjust the rate in
response to the capacity of the link between them and the source as well as
in response to dynamic network conditions. This is achieved by coding the
media stream into several layers, each of which adds an extra level of
quality. Receivers add or remove levels of quality by managing their
subscriptions to multicast groups.

The process of joining and leaving layers is controlled by a timer. If the
receiver has experienced a significant amount of loss, the receiver may
unsubscribe from the top layer when the timer expires. Similarly, new
groups are joined when the timer fires and there has been no significant
loss. The timer period is increased multiplicatively after congestion is
detected. Initially, the receiver will "ramp up" by joining groups until
congestion occurs.

The notion of "detection time" is essential to the effectiveness of this
scheme. It refers to the time necessary for the layer change to be fully
established in the network and for the resulting impact to be detected.
The value is adjusted through failed join experiments. The system
implements shared learning by having receivers inform others in their groups
of the outcome of their experiments. However, the individual receivers must
still perform their own experiments to make a decision about adding or
removing layers.

Simulation results with ns are presented and show that the framework has
reasonable loss and convergence rates under a variety of scenarios.
Evaluation of users' perception of quality is left to future work.

I think this paper combines two clever ideas, namely incremental coding and
multicast groups corresponding to each layer, into what could be a useful
and practical framework. I do have some doubts about the value of the
shared learning algorithm which informs the entire group of the outcome of a
join experiment. This information is not applicable to other receivers
unless they are in nodes that are "close" to each other. For example, for a
sender/receiver arrangement such as R1-S-R2, if R1 and R2 are receivers in
the same group, the outcome of a failed join experiment for R1 has little
predictive value for a join experiment for R2. Shared learning is useful
for reducing interference between concurrent experiments.
Received on Tue Oct 31 2006 - 00:41:46 EST