A Comparison of Layering and Stream Replication Video Multicast Schemes Taehyun Kim and Mostafa H. Ammar Networking and Telecommunications Group Georgia Institute of Technology Atlanta, Georgia
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 DSG (Cheung, Ammar, and Li, 1996) SureStream of RealNetworks Intelligent streaming of Microsoft
Replicated Stream Multicast
Cumulative Layering 1 base layer + enhancement layers Base layer Independently decoded Enhancement layer Decoded with lower layers Improve the video quality Example RLM (McCanne, Jacobson, Vetterli, 1996) LVMR (Li, Paul, and Ammar, 1998) MPEG-2/4, H.263 scalability modes
Layered Video Multicast
Layering or Replication? Common wisdom states: “Layering is better than replication” But it depends on Layering bandwidth penalty Specifics of encoding Protocol complexity Topological placement of receivers
Bandwidth Penalty Information theoretic results R(P, 2 ) R(P, 1, 2 ) Packetization overhead Syntactically independent layering Picture header GOP information Macroblock information
Experimental Comparison
Comparison by DP J. Kimura, F. A. Tobagi, J. M. Pulido, P. J. Emstad, "Perceived quality and bandwidth characterization of layered MPEG-2 video encoding", Proc. of the SPIE, Boston, MA, Sept. 1999"
Providing a Fair Comparison Need to insure that each scheme is optimized Two dimensions Selection of stream/layer rates Assignments of streams/layers to receivers
Rate allocation Cumulative layering Optimal receiver partitioning algorithm (Yang, Kim, and Lam) Stream replication Cumulative rate allocation
Stream assignment Cumulative layering Assign as many layers as possible Stream replication Greedy algorithm
Comparison Methodology Model of network Topology Available bandwidth Placement of source and receivers Determine optimal stream rates and allocation Evaluate performance
Performance Metrics Average reception rate Total bandwidth usage Average effective reception rate Efficiency
Network Topology GT-ITM Number of server = 1 Number of receivers = 1,640 Number of transit domains = 10 Number of layers = 8 Amount of penalty = 25%
Data reception rate
Bandwidth usage
Effective reception rate
Efficiency
Effect of overhead
Effect of the number of layers
Clustered Distribution Topology consideration Layering favors clustered receivers Stream replication favors randomly distributed receivers Simulate when receivers are clustered within one transit domain
Effective reception rate
Protocol Complexity Layered video multicasting Multiple join for a receiver Large multicast group size Replicated stream video multicasting One group for a receiver Small multicast group size
Average group size
Conclusion Identified the factors affecting relative merits of layering versus replication Layering penalty Specifics of the encoding Topological placement Protocol complexity Developed stream assignment and rate allocation algorithm Investigated the conditions under which each scheme is superior
Optimal Quality Adaptation for MPEG-4 Fine-Grained Scalable Video Taehyun Kim and Mostafa H. Ammar Networking and Telecommunications Group Georgia Institute of Technology Atlanta, Georgia
Related Work (1/2) S. Nelakuditi, et al, “Providing smoother quality layered video stream,” NOSSDAV 2000 Goals Achieving smoother quality for layered CBR video using receiver buffer Minimizing quality variation (maximizing runs of continuous frames)
Algorithm Forward scan Switching between select and discard phase Entering select phase if buffer is full Entering discard phase if buffer is empty Backward scan Exploiting the residual buffer Extending each run
Bandwidth Model
Experimental Result
Related Work (2/2) D. Saparilla, et al, “Optimal streaming of layered video,” INFOCOM 2000 Goal Investigating the bandwidth allocation problem to minimize loss probability Modeling the source video and the available bandwidth by stochastic process
Main Result Static policy Allocating bandwidth in proportion to long run average data rate Optimal for infinite length, independent layering Threshold-based policy If the base layer buffer is below a threshold, allocate bandwidth to the base layer
Research Goal of MPEG4 FGS Quality Adaptation Maximization of the perceptual video quality by minimizing quality variation Accommodation of the mismatch between Rate variability of VBR video Available bandwidth variability
MPEG4 FGS Hybrid Scalability Base layer Enhancement layer FGS layer: improving video quality FGST layer: improving temporal resolution
Rate Variability
Quality Adaptation Framework C[k]: transmission resource constraint X[k]: cumulative data size S[k]: cumulative selected data size d: threshold
Optimal Quality Adaptation Threshold should be equal to the receiver buffer size to achieve Minimum quality variability Necessary condition of maximum bandwidth utilization
Online Adaptation Estimating the threshold point without assuming the available bandwidth information in advance The available bandwidth is estimated by an MA style linear estimator
Experiment Model
Bandwidth Variability TCP TFRC
Performance over TFRC Threshold-based streaming (Infocom’00) Online adaptation
Performance over TCP Threshold-based streaming Online adaptation
Conclusion Accommodated the mismatch between the rate variability and the bandwidth variability Developed an optimal quality adaptation scheme for MPEG4 FGS video to reduce quality variation Investigated the perceptual quality of different algorithms and options