Receiverdriven Layered Multicast RLM

1
Receiver-driven Layered Multicast(RLM)

• Steven McCanne
• Van Jacobson
• Martin Vetterli

Presented by Krishna Bhimavarapu Thursday, 24
January 2002
2
Introduction

• Source-based, rate-adaptive, multimedia
  applications adjust
• their transmission rate to match available
  network capacity
• This does not work well for heterogeneous
  multicast environment, like the Internet
• Reasons
• - no single target rate
• - low-capacity regions of the network suffer
  congestion
• - high-capacity regions are underutilized
• Such problems are not just theoretical. They
  impact the daily use of Internet
  remote-conferencing
• This paper addresses some questions related to
  layered multicast transport through the design
  and simulation of an experimental network
  protocol called Receiver-driven Layered Multicast
  (RLM)

3

• Example U. C. Berkeleys broadcasting a seminar
  over the
• campus network and the Internet
• One approach is to combine a layered compression
  algorithm
• with a layered transmission scheme
• - selective forwarding only the number of
  layers that any link can manage
• - but mechanisms to determine, communicate and
  execute this are
• unavailable
• - a novel mechanism based on IP Multicast was
  suggested by Deering
• - the different layers of the hierarchical
  signal are stripped across multiple
  multicast groups

4
The Network Model

• RLM works within the existing IP model
• does not require any new
  machinery in the network
• Assumptions
• For single-source 1. Only best effort,
  multipoint packet delivery
• 2. The delivery efficiency of IP Multicast
• 3. Group-oriented communication
• For multiple, simultaneous sources
• Receivers can specify
  their group membership on a per-source basis
• Session is a set of end-systems communicating via
  a common set of layered multicast groups
• RLM is most easily conceptualized in a network
  where all groups follow
• the same route but this not a requirement

5
RLM PROTOCOL

• Source takes no active role
• Protocol is run at the receiver, where adaptation
  is carried out by joining and leaving groups
• - on congestion, drop a layer
• - on spare capacity, add a layer
• 1. Capacity Inference
• - the receiver must determine its
• current level of subscription too high or
  too low
• - one way is to monitor link utilization and
  explicitly notify end- systems but this
  renders deployment difficult
• - the approach in RLM is to to carry out active
  experiments by spontaneously adding layers at
  well chosen times
• - this spontaneous subscription to the next
  layer in the hierarchy is called
  join-experiment

6

• 2. RLM Adaptation
• - join-experiment cause transient congestion
  that impact the quality of
• the delivered signal
• - so, minimize the frequency and duration of
• experiment without affecting the algorithm
• - done through a learning algorithm
• - a separate join-timer for each level of
  subscription
• and exponential back off to problematic layer
• - detection-time time taken for a local layer
  change
• to be fully established in the network and
  for the
• resulting impact to be detected back at the
  receiver
• 3. Scaling RLM
• - the aggregate frequency of join-experiment
• increases as session membership grows
• - avoid problems by scaling down the individual
• experiment rate independent of session size
• (drawback is learning rate decreases)
• - solution is shared learning

7

• 4. RLM State Machine
• - Tj is randomized to avoid protocol sync
  effects
• - Td is set to a scaled value of the
  detection-time
• estimator
• - loss rate is measured with short term
  estimator and
• action is taken if estimator exceeds
  threshold (LgtT)
• - when the join-timer expires, a message is
  broadcasted
• to the group and a layer is added, the
  receiver notes
• the start of experiment
• - the experiment is in progress if the time is
  less than k1TD k2sD
• - when loss occurs in S state, resulting action
  depends on the
• presence of join-experiment
• - if a join-experiment has failed, then drop the
  offending layer, back
• off the join-timer and enter D state
• - if not certain then enter M state to look for
  longer term congestion
• - if we are not conducting the experiment, then
  transition to H state

8

• 5. Protocol State Maintenance
• - detection-time estimator and join-timers must
  be
• dynamically adapted to reflect changing
  network
• conditions
• - the join-timer is multiplicatively increased
  when the
• join-experiment fails
• TkJ lt- min (aTkJ , TmaxJ) a gt 1
• - the join-timer undergoes relaxation in
  steady-state
• - the longer a receiver is in steady-state at
  some level, the more
• likely it is for that level to be stable
• TkJ lt- max (ßTkJ , TminJ) ß lt1
• - the detection-time reflects the latency
  between time when the local
• action is carried out and the impact
  reflected back at the receiver
• - each time the join-experiment fails, the
  detection-time estimator is
• fed the new latency measurement
• - measurement, Di, is passed through low-pass
  filter with gains g1,g2
• sD lt- (1-g2) sD g2Di TD
• TD lt- (1-g1) TD g1Di
  

9
Simulations

• In a real network, performance is affected by
  cross-traffic and competing groups
• RLM protocol was implemented on the LBNL network
  simulator, ns
• Packets are generated at times
• T0 0
• Tk Tk-1 ? Nk, k gt 0
• Nk is zero-mean noise process ? - fixed
  interval chosen to meet the target bit-rate
• Variable parameters network topology, link
  bandwidths and latencies, number and rate
• of transmission layers and placement
  of senders receivers
• Fixed parameters the routing discipline,
  the router queue size, and packet size
• In all the simulations link bandwidth is 1.5
  Mb/s, traffic sources are modeled as six layer
  CBR stream at rates 32x2m kb/s,m05, the start
  time of each receiver is randomly chosen
  uniformly on the interval 30,120 seconds

10

• 1. Evaluation Metrics
• Two metrics that reflect the quality of
  real-time loss-tolerant stream
• - worst-case loss rate over varying
  time scales
• - throughput
• 2. Experiments
• - protocols delay scalability
• - scalability with respect to
• session size
• - Performance in the presence of bandwidth
  heterogeneity
• - superposition of a large number of independent
  sessions
• 3. Results
• Latency Scalability
• - duration of join-experiment is roughly twice
  the
• link latency plus the queue build-up time
• - varied link delay from 1ms to 20sec and
  computed
• worst-case loss rate to explore delay
  sensitivity
• - slide a measurement window (1,10,100 sec)
  over the packet process

11

• Session Scalability
• - plotted the maximum loss rate against the
  number of
• receivers for each time scale
• - compute the maximum loss rate by taking the
  maximum
• across the worst-case loss rates of each
  receiver
• - worst-case loss rates are independent of
  session size
• Bandwidth Heterogeneity
• - works well even in the presence of large set
  of
• receivers with different bandwidth
  constraints
• - worst-case loss rates are comparable
• - short-term congestion periods can last
  longer at larger
• session sizes, impact is reduced by
  limiting detection time estimator
• Superposition
• - bottleneck bandwidth was scaled in
  proportion to the
• number of pairs and the router queue
  limit scaled
• to twice the bandwidth-delay product
• - converged to an aggregate link utilization
  close to one

12
Network Implications

• Implications of RLM on other components in a
  comprehensive system
• 1. Receiver-consensus
• - an important requirement is all users must
  cooperate
• 2. Group Maintenance
• - performance of RLM depends on the join/leave
  latencies
• - when a receiver joins a new group, the host
  informs the next-hop router,
• which propagates a graft message in order to
  instantiate the new group
• - the leave-case is more complicated when a
  host drops a group, it broadcasts
• a leave group message when no members
  remain, the router sends a prune
• message to suppress the group
• 3. Fairness
• - an aggregate performance metric depends on how
  group fairness is
• defined
• - RLM alone does not provide fairness if
  machinery for fairness is added to
• the network, RLM should work efficiently

13
The Application

• A layered source coder is developed to complement
  the layered transmission system goal being to
  design, build and evaluate all components that
  contribute to a scalable video transmission
  system
• Two key features of the layered coder are
  resilience to packet loss and low complexity
• The RLM and layered codec are being implemented
  in the UCB/LBNL video conferencing tool, vic

Future Work

• Needs a better evaluation metric - level of
  quality perceived by the user
• Experiment with algorithms that dynamically
  adjust the bit-rate allocation of the different
  compression layers
• Interactions among multiple RLM sessions in the
  context of different scheduling algorithms
• Explore interactions with other
  bandwidth-adaptive protocols like TCP
• Improve modeling and analysis of the problem

14
SUMMARY

• The paper proposed a framework for the
  transmission of layered signals over
  heterogeneous networks using receiver-driven
  adaptation
• Evaluated the performance of RLM through
  simulation and showed that it exhibits reasonable
  loss and convergence rates under various scaling
  scenarios
• Focused on complete systems design
• Described the work on a low-complexity,
  error-resilient layered source coder. This when
  combined with RLM provides a comprehensive
  solution for scalable multicast video
  transmission in heterogeneous networks