From: Chandrika Jayant (cjayant@cs.washington.edu)
Date: Wed Nov 10 2004 - 09:51:38 PST
"Receiver- driven Layered Multicast"
Written by McCanne, Jacobson, and Vetterli
Reviewed by Chandrika Jayant
This paper proposes a new approach to layered multicast in
heterogeneous environments Each receiver subscribes by specifying its
level of subscription and joining a subset of IP-Multicast groups. The
authors present Receiver-driven Layered Multicast (RLM), a
rate-adaption protocol where the receivers adapt to bandwidth
heterogeneity and network capacity variations like congestion- keeping
bandwidth neither all used up or underutilized. Most work in this area
has been source-based rate adaptation, which doesn't deal well with
optimizing bandwidth over varied receiver bandwidth. Other work in
receiver based adaptation seems not to have been experimented with or
implemented much at the time of this paper (1996). The popularity of
multimedia applications that started before this paper makes this an
exciting topic of the time.
This algorithm, by allowing receivers to add and drop layers, makes
sure that the network is not over- or under-utilized. Learning about
congestion is easy (from packet loss data), but being aware of
underutilization is harder. Simply put, the algorithm uses adaptive
join-timers and detection-time estimators to figure out when are
optimal times to try adding more layers. Scaling in the simulations
works well in RLM because of conservative shared learning. This method
does rely on a packet drop policy. The authors say, and it is obvious,
that priority-drop policies would mess up RLM's scalability. This
seems like it would be a significant deployment problem. Scaling
according to latency and session size seems to be successful under the
paper's assumptions.
I really liked the solid structure of this paper where not only was an
idea presented, but the performance was simulated (even if only
briefly) on different network topologies, overall network implications
were discussed, and a layered source coder is built. This shows that
the authors are thinking of the big picture, deployment issues, and
how RLM could fit into the existing network, which is lacking in many
papers. Some future work could be working out the unfairness of
superposition topologies, seeing if RLM works with bursty sources, and
interacting with other bandwidth-adaptive protocols like TCP.
This archive was generated by hypermail 2.1.6 : Wed Nov 10 2004 - 09:51:44 PST