Receiver-Driven Layered Multicast

One-line summary: Each IP mcast group adds an incremental "layer" of a multimedia session transmission. Receiver's decision to subscribe/unsubscribe groups is based on periodically running adaptive-timer-controlled "experiments" to see if adding a layer will cause congestion. Other receivers near you observe your experiments, learn from them, and conditionally suppress their own experiments when others in progress, so number of experiments doesn't grow linearly with number of participants.

Overview/Main Points

• Idea: deliver multiple layers of multimedia/realtime signals on different "channels" (IP mcast grps).
• RLM focuses on "cumulative" (adding layers increases quality incrementally) but can also work with "simulcast" (each channel carries a complete copy of the signal, but at different qualities/rates).
• Why not use priority-dropping of packets to achieve? Priority-drop curve has no unique maximum (see fig 2 in paper), therefore no convergence to single stable operating point, and a misbehaving user has no incentive to reduce the requested rate so would drive network into constant congestion.
• RLM doesn't ensure fairness itself, but should coexist well with router-based schemes that do.
• Join experiment: test to see if adding another layer will improve your quality or cause congestion (lower your quality).
  • Separate join-timer w/exponential backoff used for each layer of subscription.
  • If join-experiment outlasts detection-time without congestion, experiment is successful and the layer is added. If expt. fails, the layer is quickly removed. Detection time and its variance are tracked and adapted using failed join-expts.
  • Scaling: "shared learning". Local receivers listen and learn from (adjust their join-timers from) local peers' failed join-expts.
  • Local receivers suppress starting an expt if they detect one in progress (reduce probability of expts interfering w/each other).
  • Exception: allow overlap of an expt. at same or lower level with one already in progress, to allow newer receivers with lower subscription levels to converge to optimum point more quickly.
• Protocol operation:
  • Formally described by a 4-state machine: steady state, hysteresis, measurement, drop-layer. See state diag. in fig 6.
  • Join-timer backoff: multiplicative increase of join timer when an expt fails, clamped to a maximum which is based on current number of receivers. This qty is estimated from RTCP control messages, so that aggregate join-expt rate is fixed indept of session size.
  • Join-timer relaxation: (slower) geometric decrease by Beta at every detection-timer interval, clamped to a minimum.
  • Detection-time estimate: since detection is the time between initiation of an experiment and feedback as to what it caused, detection time is computed by correlating failed-experiment start times with onset of congestion. Congestion is detected by latency measurements from failed experiments, passed through a low-pass filter.
• Simulations:
  • Claim: don't necessarily show that RLM is absolutely scalable, but do show that scaling behavior matches authors' intuitions in coming up with the mechanisms.
  • CBR streams have interarrival times that are constant perturbed by zero-mean noise process. Fails to capture burstiness in video streams, but bursty sources can be smoothed by applying rate control. (Therefore not clear how applicable RLM is to bursty sources.)
  • Simulated 4 different topologies (fig 7):
    1. point to point w/bottleneck link (PP), one source, one receiver
    2. PP with one "endpoint" being a subtree (LAN), (one source, many receivers)
    3. PP with subtrees hanging off a couple of intermediate links (one source, many receivers)
    4. PP with both endpoints being subtrees: many sources, many receivers
  • Latency scalability: as expected, packet loss rate increases rapidly (exponentially!) with join/leave latency, since it takes longer for failed experiments to be stopped and timers to be backed off, prolonging congestion periods.
  • Session scalability (topology 2): packet loss rate is about constant independent of number of receivers. Therefore shared-learning mechanisms seem to be aiding scalability.
  • Bandiwdth heterogeneity: packet loss rate is about constant across receivers with many different bandwidth constraints, though short-term congestion periods slightly larger at larger session sizes.
  • Superposition of sessions (topology 4): when independent single-source/single-receiver sessions share a common link, the bottleneck link utilization steadied out at close to 1, but often was unfair (though there wasn't starvation).
• Network implications/assumptions:
  • All users assumed to cooperate
  • Performance depends critically on mcast join/leave latency
  • When multiple sources, "good" performance not well defined (partitioning across sessions or users?). Similarly: what is "good" interaction with TCP?
• Interesting future work: since video compression used produces a prefix code, can partition bit rate arbitrarily among layers, and vary allocation dynamically.

Relevance

Flaws

• Bursty sources - ugh
• Interaction of multiple RLM sessions that share some but not all of the multicast tree? What effects on fairness?