(no subject)

The paper describes the Receiver-driven Layered Multicast (RLM)
protocol, which is a protocol for streaming data efficiently to a
number of receivers with varying bandwidth.

The protocol breaks a stream into a set of cumulative streams, where
the first stream contains the lowest quality signal, but is usable
independent of other streams, while the second stream can be combined
with the first for increased quality, the third can be combined with
the first and second, and so on. It is actually not necessary to
have cumulative streams, but the authors found that this gave the
most efficient utilization of bandwidth.

Receivers subscribe using IP multicast to as many streams as they can
without observing congestion. While the experiments necessary to
determine the ideal level at which to subscribe would interfere if
performed independently, the protocol involves a "shared learning"
strategy, whereby all subscribers to a given stream can learn from
the experiments performed by any other subscriber.

The protocol has some nice properties. It can be used without
modifications to the existing internet infrastructure, since it runs
on top of IP Multicast, and requires only modification at the end
hosts. It also allows for very efficient utilization of bandwidth.

My only real concern with RLM is that it may be incredibly easily
subverted by simple denial of service attacks. If an adversary
chooses to subscribe to the highest level of streams, regardless of
congestion, RLM will likely fail for all recipients sharing links
with this adversary. It is not entirely clear to me how this
scenario might be detected at the source and prevented.

With the exception of my concerns about susceptibility to attack, I
felt this was a strong paper: clear, concise, implement-able, and
showing significant benefit.
Received on Tue Oct 31 2006 - 07:37:42 EST