Performance of Layered Multicast Video over IP Networks

1
Performance of Layered Multicast Video over IP
Networks

• Ashraf Matrawy
• Ioannis Lambadaris
• Broadband Networks Laboratory
• Department of Systems and
• Computer Engineering
• Carleton University

2
Outline

• Background
• Experiences with NS
• RLM with cross-traffic
• Results and discussion
• Experiments in progress
• Future plan

3
Background

• IP Multicast (1)
• Addressing
• IP class D address
• Allocated dynamically
• Dynamic registration
• Receivers join and leave
• Process is done using IGMP (Internet Group
  Management Protocol)

4
Background

• IP Multicast (2)
• Multicast routing
• Protocols to build distribution trees (DVMRP,
  MOSPF,..)
• QoS New routing protocols (QoS routing)
• Multicast backbone (Mbone)
• Multicast capable routers (Mrouter)
• Network of Mrouters (via tunnels)

5
Background

• QoS requirements for video multicast
• Dealing with heterogeneity
• Networks with different service capacities
• Scalability
• Number of receivers may grow significantly
• Control messages may overwhelm the network
• Upper bound on end-to-end delay
• Bandwidth guarantees
• Upper bound on packet loss rate

6
Background

• Applications and proposals
• Application performance is adapted to the network
  behavior
• More than one approach to follow
• Sender driven Scalable feedback control
• Receiver-driven Layered Multicast (RLM)
• Applications rely on IP multicast on Mbone

7
Background

• Scalable Feedback Control (SFC)
• Sender uses a single rate that is used to
  optimize the group reception quality
• Periodically, the sender probes the network and
  based on the information from receivers, it makes
  a decision about the sending rate
• Increase, decrease, or no change (within MAX_RATE
  and MAX_RATE)
• INRIA Videoconferencing System (IVS)

8
Background

• Problems with SFC
• Unfairness of using one rate due to heterogeneity
• How to pick up "good" values to increment or
  decrement the rate
• What fraction of the receivers should be
  congested in order to change the rate
• MIN_RATE and MAX_RATE chosen by the user may not
  be advantageous

9
Background

• Receiver-driven Layered Multicast (RLM)
• Signal is encoded into a number of layers
• The lowest layer contains basic information
• Incrementally combining the layers in sequence
  provides progressive enhancement
• End-to-end control compensates for the missing
  network layer support
• Layered coding layered transmission
  End-to-end control

10
Background

• RLM (2)
• Different layers are being multicast to different
  (multicast) addresses
• A receiver subscribes to as many layers based on
  bandwidth availability (by joining different
  multicast groups)
• When the network is congested, a receiver drops
  the last layer it subscribed to
• If the network has spare capacity, the receiver
  adds the next enhancement layer

11
Background

• RLM (3)

12
Background

• Join/Leave sequence in RLM
• Comment on this will follow in the results

4
3
Layer
2
1
13
Background

• Problems with current proposals(1)
• Scalability
• Increase of the number of receivers will increase
  the number of control messages exchanged
• Using multicast to distribute information about
  experiments overloads the network
• More processing load on the routers for pruning
• Delay in congestion detecting and recovery
• Routers will forward congestion warnings to
  receivers

14
Background

• Problems with current proposals (2)
• Fairness issues
• BW usage during join/leave operation and after
  detection of congestion
• Uniform dropping of packets
• Packets are dropped uniformly in time of
  congestion irrespective of their layer
• Maybe a start to use a priority dropping scheme
  in the network (new QoS architecture for the
  Internet)

15
Experiences with NS

• NS ver. 2.1b4a is used for our experiments and
  its assisting tools, nam, xgraph
• Discrete event simulator
• NS is implemented in C (for data modeling, e.g.
  packets) and Otcl (for control)
• Data networks components are implemented as
  objects
• We simulated RLM using existing NS classes

16
Experiences with NS

• Useful NS resources
• NS source code
• http//www-mash.cs.berkeley.edu/ns/
• Mailing list
• ns-users_at_mash.cs.berkeley.edu/
• Tutorial
• http//titan.cs.uni-bonn.de/greis/ns/ns.html/
• Help on code and class hierarchy
• http//www-sop.inria.fr/rodeo/personnel/Antoine.Cl
  erget/ns

17
RLM with cross-traffic

• Objectives
• RLM interaction with cross-traffic was not
  studied before.
• We monitored
• Subscription level of receivers
• Frequency of join/leave experiments
• Packet loss rate
• Using different settings of
• Number of layers
• Link capacities of the receivers
• Bottleneck link bandwidth

18
RLM with cross-traffic

• Setup
• one source for video (denoted by node 0)
• Zero, one or more ftp sources (over TCP)
• Three receivers with different link capacities
  (50kbps, 100kbps, and 250kbps)
• There is a bottleneck between senders and
  receivers with capacity B Mbps
• Video source is a set of CBR sources representing
  the layers
• Layers are sent at these rates (32k, 64k, 128k,
  256k, 512k, 1024k)

19
RLM with cross-traffic

• Preliminary simulation setup

7
3
0.5Mbps
100 kbps
B Mbps
1Mbps
250 kbps
0
1
2
4
50 kbps
0.5M bps
5
6
20
RLM with cross-traffic

• Experiments overview

21
RLM with cross-traffic

• Discussion of results
• With the introduction of cross-traffic
• Subscription level of the receivers is reduced
• Frequency of join experiments is higher
• Loss rate is NOT much higher due to the change of
  the subscription level of receivers
• With higher bottleneck bandwidth, the frequency
  of join experiments is higher
• With higher number of layers, the frequency of
  join experiments is higher

22
Experiments in progress

• Uniform vs. priority dropping with RLM
• New Internet architectures for QoS support
  priorities (e.g., diffserv)
• Introducing complexity in the network should not
  be considered before proving the benefits
• Use a 2-layer video source
• Give the basic layer a higher priority
• Routers should support priority forwarding
• Study the performance with uniform dropping and
  with priority dropping

23
Experiments in progress

• Uniform vs. priority dropping with RLM (2)
• Video source marks packets in two different ways
• Routers Rt1 and Rt1 forward packets according to
  their priority (marking)
• B is the bandwidth of the bottleneck

B