Receiver-driven Layered Multicast

Paper Review: Receiver-driven Layered Multicast

The paper presents a new multicasting scheme optimized to work in the
Internet, a heterogeneous network. Efficiently broadcasting
multimedia content to a large number of nodes with varying download
bandwidths, while providing each with its desired bit rate, is not a
trivial problem. This paper attempts to answer some of the open
questions in layered multicast through their scheme called
Receiver-driven Layered Multicast (RLM).

The paper begins by presenting the network model assumed by their
scheme. Their assumptions are satisfied by the IP protocol, including
IP Multicast, and thus can be considered realistic. The paper
differentiates between their scheme, which is simulcast based, rather
than the method of distributing multiple copies of the same content at
different rates which is less efficient and less scalable.

In their model, the source simply transmits each layer of its signal
on a separate multicast group. The key aspects of the protocol are
carried out at the receiver by joining and leaving multicast groups
based on whether there is congestion or spare capacity. Congestion is
easy to detect through dropped packets. However, spare capacity is
more complicated to infer. They solve this problem by using
?join-experiments? where the receiver joins a new layer and if there
is congestion, this new layer is immediately dropped. The
join-experiments have the potential to cause congestion and thus a
join-timer is used in conjunction with a learning algorithm to prevent
this from occurring frequently. To further increase scalability,
join-experiment frequency decreases proportionally with the group size
as well as ?shared-learning? is used to increase information among the
receivers.

The timers must be dynamically modified to keep up with dynamic
network conditions. These timers are also proportional to the number
of receivers to allow for varying session sizes. This also means that
the congestion due to join-experiments is not affected by the session
size.

The authors present a set of simulations to show that their intuition
about scalability is true, but do not go as far as stating that the
system is in fact scalable. They propose two metrics in an attempt to
capture the true performance of the system. The first is worst-case
loss rate over varying time scales. The second metric is the time it
takes to converge to the optimal operating point, considered the
?throughput? of the system. The simulations were extensive and
examined a broad range of situations and performance issues.

One complaint about this paper is that the introduction section does
not provide enough general information about the ?current state? and
thus it is unclear exactly what the ?problem? with layered multicast
is that needs to be solved. An interesting, although unintended,
effect of the work is that the use of timers, exponential back-offs
and multiplicative increases renders their scheme to closely resemble
TCP. This supports the point that TCP is well-designed and its
methods of dealing with dynamic network conditions are widely
replicated. One of the key aspects of their design is that they
decided to make use of the IP multicast which to this day is not
widely utilized. IP multicast prevents the need for multiple unicast,
which places a great drain on the source. Although issues such as
fairness and feasibility seem open, the paper presents a novel
solution to a problem which is definitely in need of solving.
Broadcasting multimedia content is an integral part of the future of
the Internet. The layered approach is also very useful as the
Internet continues to grow, so does the degree of heterogeneity of its
participants.

-- Nadeem Abji
Received on Mon Oct 30 2006 - 19:32:39 EST