A Comparison of Heterogeneous Video Multicast Schemes: Layered Encoding or Stream ReplicationBy

Taehyun Kim and Mostafa Ammar

Presented by

Neyaz Shafi and Kaizad Avari

Taehyun Kim B.S. degree in control and instrumentation

engineering from Seoul National University, Korea (1992)

M.S degree in electrical engineering from Korea Advanced Institute of Science and Technology, Taejon (1994)

Ph.D. in electrical and computer engineering from Georgia Institute of Technology, Atlanta (2005)

Research Objective: Development of efficient algorithms and protocols for scalable multimedia delivery.

http://www.cc.gatech.edu/computing/Telecomm/people/Phd/tkim/

Freescale Semiconductors

Mostafa H. Ammar

Advisor to Mr. Kim during his Ph.D.

B.S. and M.S degrees from MIT

Ph.D degree in electrical engineering from University of Waterloo, Ontario, Canada.

Regents' Professor ( generally granted to a small percentage of the top tenured faculty who are regarded as particularly important in their respective fields of study) with the School of Computer Science at the Georgia Institute of Technology

http://www.cc.gatech.edu/fac/Mostafa.Ammar/

Professor at Georgia Tech since 1985

What is multicasting? One-to-many communication in a network

Information is addressed to a group of destination computers simultaneously

Two ways to implement : may be implemented at the Internet Later using IP multicast or at the Data Link Layer using one-to-many addressing and switching such as Ethernet Multicast Addressing, Asynchronous Transfer mode (ATM) point-to-multipoint virtual circuits (P2MP) or Infiniband multicast

What is the paper about?

1. Introduction Heterogeneous transmission resources and end system capabilities

Difficult to agree on acceptable traffic characteristics

Three ways around this problem:

1. Multicast replicated streams at different rates

2. Multicast video encoded in cumulative layers

3. Multicast video encoded in noncumulative layers

General belief that layering approach is better is not accurate!

OVERHEAD due to layer encoding has to be considered (key observation that serves as the basis of this paper)

Replicated Stream Approach

Several streams

Each has IDENTICAL CONTENT at different data rates.

Each stream reaches its multicast GROUP

Each receiver SUBSCRIBES to a stream that seems adequate

Receivers can switch streams

Cumulative Layering Approach

Encoded in BASE layer + ENHANCEMENT layers

Can decode base layer independently

Needs (k-1)th layer to decode kth layer

PROGRESSIVE refinement

Subscribe to atleast layer 1 multicast group

Additional layers add quality to video stream

Noncumulative Layering Approach

Encoded in two or more INDEPENDENT layers

Each layer can be decoded independently

Multiple Description Coding (MDC) can be used

Varies the video stream based on different properties.

2. Overhead in Layered Video

Rnl – Data rate of a video encoded in a single stream (non layered) including all protocol and packetization overhead

M cumulative layers in layered encoded video.

Ri – Data rate of the ith layer including all protocol and packetization overhead.

Results in literature show a different picture.

Equality is rarely achieved and the rate required by layered video can be as much as 20-30% higher.

Substantiating the claim

A. Information Theoretic Results

B. Packetization Overhead

C. Experimental Evidence

D. Protocol Overhead

A. Information Theoretic Results

These results are derived in terms of the rate distortionfunction which describes the required rate toencode a memoryless source at a maximum distortion of delta.

The distortion is a measure of the quality degradationrepresented by the encoding of the source.

The general result is that, for the same source and thesame distortion, a successively refined (i.e., layered) encoding requires at least as much data rate as a nonlayered encoding [15].

While equality is possible, it requiresa strict Markovian condition to apply to the source andis generally not achievable. Moreover, the result in [14]shows that the performance of the layered encoding is notbetter than that of nonlayered encoding for a finite-lengthblock code, even if the Markovian condition holds.

B. Packetization Overhead

For certain scalability modes, enhancement layers aredesigned to be syntactically independent of one another.

Along with the residual information, a data stream needsto also carry syntactic data, such as picture header, startcodes, group of pictures (GOP) information, and macroblock header.

This means incurring a large amount of overheadespecially at low data rates [17].

C. Experimental Evidence Video Quality versus Data Rate

MPEG-2 SNR scalability and nonscalability mode

Thevideo quality is measured in peak signal-to-noise ratio(PSNR) by varying quantization step size

A layeredstream has two layers consisting of a base layer and an enhancement layer.

Key observations from experiment

Layered stream requires more data rate than a nonlayered stream to provide the same quality

The difference ranges from 0.4% at 27.7 dBto 117% at 23.2 dB

Difference is expected to grow as the number of layers increases, since the accumulation of the redundancy leads to the increase of theoverall distortion in layered video encoding

More extensive experiment

In reference [16] - “Perceived quality and bandwidth characterization of layered MPEG-2 video encoding”

The authors investigated the impact of the number of layers, bit rates, and packet loss on the perceptual video quality as determined by subjects scoringthe quality of the video, when MPEG-2 data partitioning and SNR scalability are employed

difference ranges from nearly 0 for the highest quality video (scoring close to 4.5) to 57% for fair quality video (scoring close to 3)

For a score of 4 (good quality video), the overhead varies from 2%(the flower sequence) to 49% (the basket sequence)

D. Protocol Overhead

The nature of the subscription to multiple layers in layered video multicasting may cause additional overhead, as the receiver needs to manage these multiple subscription

The amount of bandwidth overhead is increased, as the group size of a multicast group grows.

The subscription of multiple layers requires more buffersize and better synchronization capability than replicatedstream video multicasting.

3. Optimizing Stream Rates

For a fair comparison, we need to ensure that each scheme is optimal

Need to determine no. of streams and their rates for both schemes.

1. RATE ALLOCATION ALGORITHM can determine the data rate

for each stream

2. STREAM ASSIGNMENT ALGORITHM can determine the

reception rate for each receiver

The goal is to maximize the bandwidth utilization for

1. A given network,

2. A particular set of receivers and

3. Given available bandwidth on links

Terminology

A network graph G = (V, E), where

V is a set of vertices representing hosts and routers

E is a set of edges defined over the V x V network

A SET OF RECEIVERS is

An ISOLATED RATE for each receiver is the reception rate for that receiver when there are no constraints

Bandwidth function is a measure of the residual bandwidth available

on link ej.

Cumulatively layered multicasting

A session is defined by

where αi is data rate of layer ‘i’ and ‘m’ is the no. of layers.

The authors used an optimal receiver partitioning algorithm to determine optimal stream rates using dynamic programming

This maximizes overall EFFECTIVE reception rate

1. Rate Allocation

2. Stream Assignment The reception rate is the sum of stream rates that does not exceed

the isolated rate.

Replicated Stream Multicasting

A replicated stream multicast session is denoted by

where βi is the data rate of a replicated stream and ‘m’ is the number of replicated streams

β1 corresponds to the base layer of the cumulative layering

If a receiver can join ‘k’ layers, then in a replicated stream,

1. Rate Allocation

Replicated Stream Multicasting

Receiver subscribes to either BASE or higher level stream.

Define δi , the data rate of the stream, such that

Define Ωi as a SET OF RECEIVERS such that

where Փδ is the rate allocation function

Set up two objectives for stream assignment:

1. Min reception rate for all receiver is greater than zero

2. Maximize

2. Stream Assignment

Replicated Stream Multicasting

Non-cumulatively Layered Multicasting

A session is defined bywhich signifies the rate of the stream,

‘m’ being the number of streams.

Set of receivers assigned is defined bywhere Փγ is the stream rate

function.

A receiver can subscribe to ANY SUBSET of receivers.

A stream γ = 1,2,4 in this mode would equal a cumulative stream consisting of 7 layers, 1,2,3,4,5,6,7 since the previous is required to decode any level

Two objectives to assign streams:

1. The min reception rate is greater than zero for all streams

2. Maximize

1. Rate Allocation

2. Stream Assignment

Non-cumulatively Layered Multicasting

4. Models in Experiments

The main goal in the experiment is to evaluate the impact of the parameters, such as the amount of layering overhead and the topological placement of receivers, on the video reception quality

All schemes use the rates and stream assignment as determined by the algorithms in previous section.

Network Model

GT-ITM [27] used to generate 100 different transit-stub graphs representing hierarchical Internet topologies

The graphs consist of 1,640 nodes including ten transit domains, four nodesper transit domain, four stubs per transit node, and ten nodes in a stub domain

2.4 Gbps to transit-to-transit edges; 10 Mbps and 1.5 Mbps to stub-to-stub edges; and 155 Mbps, 45Mbps, and 1.5 Mbps to transit-to-stub edges

The available link bandwidth is chosen uniformly randomly in the range 1%–80%of the full capacity of the edge.

Layering Overhead Model

The number of cumulatively layered video streams and the number of replicated video streams are 8, and the number of noncumulatively layered video streams is 4

In this paper, the authors consider a dynamic overhead model. The dynamic overhead model captures the notion of the dynamically varying nature of the layering overhead. The model is based on the experimental results in [20]

Based on experimental results, authors modelled the layering overhead by linear interpolation

Layering Overhead = 520 – 1.6*R(data rate of base layer) when R <= 325 kbps

= 0 when R(data rate of base layer) >325kbps

Performance Measures

the average reception rate which is the average data rate received by a receiver

average effective reception rate where the effective reception rate at a receiver is defined by the amount of data received less the layering overhead

total bandwidth usage calculated by adding the total traffic carried by all links in the network for the multicast session—including all layers and all replicated streams

Bandwidth usage efficiency defined by The efficiency is a measure of delivered data rate contributing tothe video quality for each unit of bandwidth used in the network

5. Results – Random Distribution

Methodology - Randomly select a server and receivers from a set of nodes in the graph. Receivers are selected from all domains which results in random distribution of receivers. Authors then investigate the performance of the video multicast schemes by varying the number of receivers . Dynamic overhead model is used.

• average videoquality of replicated stream video multicasting is the best. •The efficiency of replicated stream video multicasting is also the bestin

Results – Random Distribution

Experiment results in the dynamic model under the random receiver distribution: (a) Reception rate, (b) effective reception rate, (c) total bandwidth usage, and (d) efficiency. Replicated stream multicasting shows the largest effective reception rate in (b) and the best bandwidth usage efficiency in (d).

Results – Clustered Distribution

Methodology - receivers are chosen within only one transit domain and a sender is selected from another domain. Multiple streams share the bottleneck link

•layered video multicasting is more efficientthan replicated stream video multicasting

•the performance of layered video multicasting is improved but that of replicated stream video multicasting is degraded

•performance characteristics are changed in favour of layered video multicasting, whenreceivers are clustered in a small number of domains.

Results – Clustered Distribution

Experiment results in the dynamic model under the clustered receiver distribution: (a) Reception rate, (b) effective reception rate, (c) total bandwidth usage, and (d) efficiency. Both cumulatively and noncumulatively layered video multicasting achieves greater data reception rate and greater effective reception rate than that of replicated stream multicasting.

6. Protocol Complexity

Receiver Driven Layered Multicast (RLM)

Receiver decides

Receivers perform Join Experiments to try establishing connections

These incur bandwidth overhead

Shared learning mechanism requires significant amount of state information.

For Layered multicasting, the group size is given by

Needs 2 unicast messages and 1 multicast message for join experiment

For replicated multicasting, the group size is given by

Needs 4 unicast messages and 1 multicast message for join experiment

COST of multicast message is dominant!

Protocol Complexity

Bandwidth Buffer Size

Conclusion

Conclusion is in line with expectations

Replication is better suited to a RANDOM DISTRIBUTION of hosts

Layering is better suited to a CLUSTERED DISTRIBUTION of hosts

EFFECTIVE reception rate of a layered approach is significantly lower than the channel reception rate.

Questions

Thank you!