Paper review: Receiver-driven Layered Multicast

Paper: Receiver-driven Layered Multicast

Name: Robert Danek
Course: CS2209, Fall '06

        This paper discusses the problem of multicasting data streams, such as
real-time multimedia, on the internet. The major issue with doing this
is that the internet is heterogeneous; a rate of transmission that may
seem appropriate for one part of the internet may overwhelm or
underwhelm another part of it. Also, congestion may occur, and this
necessitates varying the rate of transmission as the amount of
congestion varies. Unfortunately, the sender cannot intelligently vary
the rate of transmission when multicasting data, since congestion may
not be occurring consistently on the paths to all receivers.

        To solve the above problems, the authors of this paper propose a
scheme called Receiver-driver Layered Multicast (RLM). In this scheme
the signal being transmitted is encoded into a number of layers such
that combining them provides a more refined signal. The layers are
"striped" across different multicast groups, and receivers can then
adjust to the available bandwidth by adding and removing layers, which
is done by joining and leaving the different multicast groups.

        The way receivers adapt to congestion is by performing what is called
"join-experiments". When no congestion is detected, receivers will join
a new group (thus adding a new layer to the signal being received), and
see if it causes congestion. If not, then the receiver has now better
utilization of the available bandwidth.

        In order to reduce the amount of join experiments that occur during
the same time, RLM uses the idea of "shared learning". With this, when
a node is about to perform a join-experiment on some layer, it notifies
the other members in the group of this fact through a multicast
message. Thus, if a node in the group detects packet loss, it knows it
was due to a failed join-experiment taking place; it does not need to
conduct its own join-experiment to determine that adding another layer
would cause congestion.

        This was an alright paper. One of the drawbacks was that in the
simulations the authors only explored simple network topologies. They
concede that real networks are more complex and that RLM may have
issues with oscillatory behaviour and scalability in that case. The
authors also go on to discuss some of the other implications of RLM. In
particular, RLM requires that all nodes downstream of any link
cooperate; if a node decides to add a layer that causes congestion on a
link, then all the nodes downstream of the link will suffer the effects
of that congestion. It is also important that latencies for joining and
leaving a multicast group be fairly low for RLM to perform well.
Received on Mon Oct 30 2006 - 23:40:32 EST