Summary: RLM Protocol

Summary: RLM Protocol

The paper presents design and simulation of a Receiver-driven Layered
Multicast (RLM) protocol for scalable multicast transmission in
heterogeneous networks. RLM assumes that receivers can specify group
membership on a per-source basis. It also assumes on the delivery efficiency
of IP multicast and does not guarantee packet ordering and minimum
bandwidth.

One of the key features of is that is that RLM receivers adjust transmission
rate to match the available capacity in the network. RLM framework transmits
layered signals over heterogeneous networks using receiver driven
adaptation. In RLM, receivers implicitly define the multicast trees by
expressing their interest in receiving flows. Each receiver runs the
following simple algorithm: 1) on congestion, drop a layer. 2) On spare
capacity, add a layer.

RLM carries out join-experiments by spontaneously adding layers at "well
chosen" times. It minimizes frequency and duration of join-experiments
without impacting algorithm's ability to track changing network conditions
(convergence rate). It does this by initializing the detection time
estimator with a large value (like mean etc.) and further adapting it using
failed join-experiments. This learning strategy is implemented by managing a
separate join-timer for each level of subscription and applying exponential
back-off to problematic layers. Scaling in RLM is achieved by shared
learning algorithm.

The key advantage of shared learning algorithm is that every receiver need
not run individual experiments to discover congestion on their own. Shared
learning process determines what does not work as opposed to what works.
However, success/failure decision is based on local observations, not on
global outcome i.e. on the path from source to receiver.

RLM maintains a state machine to shift between different states – steady
state, hysteresis, measurement, and drop state. In addition to current state
identifier, the receiver algorithm maintains current subscription level,
detection time estimator and join-timers.

Simulations were done on various topologies to characterize throughput and
worst case loss rate. Simulation results:
1) Duration of join experiment is roughly twice the link latency plus queue
buildup time.
2) Worst case loss rates are independent of session size.
3) RLM works well even in the presence of large sets of receivers with
different bandwidth constraints.
4) Long term loss rates are under 1% while medium loss rates are a few
percent.
5) Performance of RLM depends critically on join/leave latencies.

Though the protocol performs reasonably well, the performance on real
networks will be affected by cross-traffic and competing groups because they
could add noise to the measurement process and could potentially lead to
oscillations. Also, protocol does not provide fairness e.g. if a misbehaving
receiver keeps sending packets without caring for the congestion in the
network, throughput of near by receivers could suffer significantly.
Finally, as acknowledged by the paper, protocol evaluation doesn't study
ultimate quality received by the user.
Received on Tue Oct 31 2006 - 06:26:15 EST