From: Lillie Kittredge (kittredl@u.washington.edu)
Date: Tue Nov 09 2004 - 23:46:10 PST
RLM
This paper discusses RLM, a protocol for distributing multimedia
applications at varying quality levels in a heterogeneous multicast
network.
The key deadly difference between multicast and unicast is that the sender
cannot accommodate all the clients with a single transmission rate. This
system multicasts the stream at multiple different levels, to which
clients can subscribe at increasing levels, only taking as much as they
need without causing congestion.
The most innovative thing in this paper as I saw it was the "shared
learning" technique, where, rather than each node testing the waters of
congestion for itself, which would take a long time and cause confusion,
nodes let the rest of the network know they're attempting to subscribe to
a new layer, so that the whole network can see the result. I find this an
interesting balance between emergent and explicit behavior.
I was also very glad to see this take into account the deployability of
the system. They base their protocol on existing behavior, avoiding
insisting on, for instance, an explicit notification of available
bandwidth. This attention to the state of things is reflected in their
focus on testing of RLM with drop-tail queueing. Though drop-tail
queueing is generally crappy, and specifically crappy in the context of
RLM and especially compared to RED, they know that drop-tail is much more
prevalent.
The other cool thing about this paper was that they're interested in
making a whole system. They've got the protocol to deal with getting the
highest quality media to the clients capable of getting it; now they're
working on the system for encoding those layers. That shows rather
considerable foresight, which I like.
This archive was generated by hypermail 2.1.6 : Tue Nov 09 2004 - 23:46:14 PST