Transcript A Comparison of Layering and Stream Replication Video

A Comparison of Layering and Stream
Replication Video Multicast Schemes
Taehyun Kim and
Mostafa H. Ammar
Content





Research Goal
Replication VS Layering
Experimental Comparison
Results
Conclusion
Research Goal

A systematic comparison of video
multicasting schemes designed to deal with
heterogeneous receivers
–
–
–
Replicated streams
Cumulative layering
Non-cumulative layering
Stream Replication




Multiple video streams
Same content with different data rates
Receiver subscribes to only one stream
Example
–
–
SureStream of RealNetworks
Intelligent streaming of Microsoft
Replicated Stream Multicast





R1, R2 and R3 are from different
domain
Receivers subscribe to only one
stream
R1 joins the high quality stream
(8.5Mbps)
R2 receives the medium quality
stream (1.37Mbps)
R3 joins the low quality stream
(128kbps)
Cumulative Layering


1 base layer + enhancement layers
Base layer
–

Enhancement layer
–
–

Independently decoded
Decoded with lower layers
Improve the video quality
Example
–
MPEG-2 scalability modes
Non-Cumulative Layering



Video is encoded in two or more independent
layers
Receiver can join any subset of the video
layer without joining the layer 1 multicast
group
Example
–
Multiple description coding (MDC)
Layered Video Multicast



R1 subscribes to all video layers (10
Mbps)
R2 joins enhancement layers 1 and
the base layer (1.5 Mbps)
R3 just receives the base layer
(128kbps)
Layering or Replication?

Common wisdom states:
–

“Layering is better than replication”
However, it depends on
–
–
–
–
Layering bandwidth penalty
Specifics of encoding
Protocol complexity
Topological placement of receivers
Layered Video Multicast


Considering 20% overhead, the
data rates contributing to the video
quality are 8Mbps, 1.2Mbps and
102.4Kbps
Stream Replication: video quality
are 8.5Mbps, 1.37Mbps and
128kbps
Bandwidth Penalty

Information theoretic results
–
–

Recent results showed that the performance of layered
coding is not better than that of non-layered coding
Increase the number of layers => significant quality
degradation
Packetization overhead
–
Enhancement layers carry:



Picture header
GoP information
Macroblock information
Experimental Comparison



Non-layered streams has better
video quality
Difference in data rates ranges from
0.4% at 27.7dB PSNR to 117% at
23.2dB PSNR
For a good quality video, the
overhead is around 20%
Providing a Fair Comparison


Need to insure that each scheme is optimal
Two dimensions
–
Stream assignment algorithm

–
Rate allocation algorithm


Determine the reception rate of each receiver by aggregating
the data rates of the assigned streams
Determine the data rate of each stream
Goal
–
Maximize the bandwidth utilization by each scheme for



a given network
a particular set of receivers and
given available bandwidth on the network links
System Model

Model the network by a graph G = (V, E)
–
–
V is a set of routers and hosts
E is a set of edges representing connection links


C  ci | ci V , i  1,..., n
n is the number of receivers

Isolated rate
–
The reception rate of the receiver if there is no
constraint from other receivers in the same
session
Stream Assignment

Cumulative layering
–
Define   i | i  R  , i  1,..., m

–
i is the data rate of a stream and m is the number of
layers
Assign as many layers as possible


Compute the isolated rates
Assign  i that does not exceed the isolated rate
Stream Assignment

Stream replication
–
Define    i |  i  R  , i  1,..., m


–
i is the data rate of a replicated stream and m is the number of
replicated streams
Set of receivers assigned to stream i,   c j |   (c j )   i


Two objectives


Minimum reception rate for all receivers is greater than zero
m
Maximum Z   i 1 i  i subject to i  i  b j
ej

Greedy algorithm
–
–
Allocate 1 to all receivers to satisfy the minimum reception rate constraint
Receiver is assigned a stream that has not been assigned and has the
maximum value of group size and stream rate product
Stream Assignment

Non-cumulative layering
– Define    |   R , i  1,..., m

i


–
i
i is the data rate of a non-cumulatively layered stream and m is the
number of streams
Set of receivers assigned to stream i, '  c j |  i   (c j )


Two objectives


Minimum reception rate for all receivers is greater than zero
m
Maximum Z  i 1 i'  i subject to i  i  b j
ej
Rate Allocation

Cumulative layering
–
Optimal receiver partitioning algorithm (Yang, Kim and Lam 2000)
determines the optimal rates of layer i, i





Receivers are partitioned into K groups (G1, G2,…, GK)
Objective is to maximize the sum of receiver utilities
Dynamic programming algorithm is used to find an optimal partition
For a given partition, an optimal group transmission rate can be
determined
Stream replication
–
Stream rates, i, are allocated based on the optimal cumulative
layering rate
1
i   i
 j 2  j
i 1
2im
Rate Allocation

Non-cumulative layering
–
–
–
–
Receiver can subscribe to any subset of layers
without joining the base layer
={1,2,4} => isolated rates of {1,2,3,4,5,6,7}
2m-1 different link capacities with m noncumulative layers
 1  1
i are allocated based on i =>
 2  1   2
 1   2  1   2   3
 3  1   2   3   4

Performance Metrics

Average reception rate
–

Average effective reception rate
–

Amount of data received less the layering overhead
Total bandwidth usage
–

Average rate received by a receiver
Adding the total traffic carried by all links in the network for
the multicast session
Efficiency
–
total effective reception rate / total bandwidth usage
Network Topology

Georgia Tech Internetwork Topology Models (GT-ITM)
–
–
–
–
–
–
–
–
1 server
1640 nodes with 10 transit domains
4 nodes per transit domains, 4 stubs per transit node, 10
nodes in a stub domain
transit-to-transit edges = 2.4Gbps
stub-to-stub edges = 10Mbps and 1.5Mbps
transit-to-stub edges = 155Mbps, 45Mbps and 1.5Mbps
number of layers = 8
amount of penalty = 20%
Date Reception Rate


Cumulative layering
can receive more data
Number of layers in
cumulative layering is
twice as many as that
of non-cumulative
layering
Cumulative
Non-cumulative
Replication
Bandwidth Usage

Bandwidth consumption
of cumulatively layered
multicasting is the largest
Cumulative
Non-cumulative
Replication
Effective Reception Rate

Only 80% of data
contributes to improving
the video quality
Cumulative
Non-cumulative
Replication
Efficiency

Replicated stream video
multicasting is more efficient
Cumulative
Non-cumulative
Replication
Effect of Overhead

Layering overhead of more
than 7% tends to favor the
replicated stream
approach
Effect of the number of layers


Efficiency of stream
replication is always
greater than that of
cumulative layering
The effect is not so
significant
Narrow Distribution
Wide distribution


Narrow distribution
The layering approach achieves better bandwidth efficiency when multiple streams share
the bottleneck link
In narrow distribution, the reception rates in Figure (a) is larger than that of Figure (b) by
1.63Mbps
Efficiency

Compared to the wide
distribution results, the
performance of replicated
stream video multicast is
degraded
Cumulative
Non-cumulative
Replication
Protocol Complexity

Receiver-driven Layered Multicast (RLM)




Layered video multicasting
–
–

Receivers decide whether to drop additional layer or not
Join experiment incur a bandwidth overhead
Receivers send a join message and multicast a message
identifying the experimental layer to the group
Receiver can join multiple groups
Large multicast group size
Replicated stream video multicasting
–
–
Receiver only join one group
Small multicast group size
Average Group Size


Group size in
cumulatively layered
video multicasting is
twice as large as that
in stream replication
More bandwidth to
multicast a message
reporting the “join”
experiment
Conclusion

Identified the factors affecting relative merits of
layering versus replication
–
–
–
–


Layering penalty
Specifics of the encoding
Protocol complexity
Topological placement
Developed stream assignment and rate allocation
algorithms
Investigated the conditions under which each
scheme is superior