Chun-Yuan Chang, Cheng-Fu Chou * and Ming-Hung Chen Presenter: Prof. Cheng-Fu Chou National Taiwan University
Introduction Model & Chunk Selection Strategies Practical P2P Streaming System & Dynamic Strategy-Switch Performance Evaluation Conclusion
Swarm-Based P2P Streaming Similar to “BitTorrent” Encourage users to contribute its outbound bandwidth and storage to speed up content distribution. PPLive, PPStream, CoolStreaming and GridMedia, etc
Two components Overlay construction Chunk swarming mechanism Buffer map exchange Chunk scheduling The chunk IDs the peer possesses
Content bottleneck problem No content to exchange even if outbound bandwidth is sufficient More diverse the content distribution is made, the less the content bottleneck is !!
Existing approaches Rarest-First E.g. CoolStreaming Infocom 2005 Random E.g. Chainsaw Infocom 2005 Hybrid ones (Deadline-First + Rarest-First) E.g. Bitos Infocom 2006 and Prime Infocom 2007 Network Coding E.g. R2 JASC 2007
System dynamics Peer churn Network core congestion Variable source streaming rate Content diversity Random chunk loss Content Importance Unequal content importance
With and without considering content importance
Introduction Model & Chunk Selection Strategies Practical P2P Streaming System & Dynamic Strategy-Switch Performance Evaluation Conclusion
Simple Model (ICNP 2007)
Recursive Formulation
Priority B(1)>B(2).. therefore
c h > c l Rarity is adopted to do a tie-break
Only c h can compete to each other
ComparisonScheduling efficiency Content Bottleneck IFGoodHigh RFPoorLow How can we support high scheduling Efficiency and maintain the scalability at the same time?
When population size is not large, we can enjoy throughput and scheduling efficiency simultaneously There exist a good balance between content diversity and content importance
Introduction Model & Chunk Selection Strategies Practical P2P Streaming System & Dynamic Strategy-Switch Performance Evaluation Conclusion
Receiver Side
Supplier Side
As a receiver: Detect if the number of retrieval chunks in the request window is zero. If it does, send a signal to itself. No scheduling process will be performed. If it does not, just subscribe to all desired chunks and assign each desired chunk to a peer who possesses the chunk in a random fashion As a sender: Check if the event of content bottleneck is captured. If it does, conduct RAND on each requested packet. Otherwise, conduct IF on each requested packet.
Introduction Model & Chunk Selection Strategies Practical P2P Streaming System & Dynamic Strategy-Switch Performance Evaluation Conclusion
Simulator GridMedia Project Settings:
Video Trace: Encoded by H.264 (JM16.0) Concatenated by different types of CIF video sequences, which include high motion and low motion video sequences Fixed the quantization parameters (QP) for I,P,B frame in encoding
Delivery Ratio: the ratio of the number of chunks that arrive before playback deadline to the number of chunks that should arrive before playback deadline. PSNR (dB): the rendered video quality compared with the raw video sequence. The ffmpeg is used as our decoder.
RAND: peers always serve the chunk in random fashion IF-IPB: peers always serve the chunk with highest priority with respect to –IPB. PR-IPB: the prioritized random scheduling in [10] UL-IPB: the utility-like approach in [15]
Scalability
Scheduling Efficiency
PSNR over time with 4,500 peers underloadoverload
Point out the trade-off between content diversity and content importance A simple but effective content bottleneck detector is proposed to strike the balance between content diversity and content importance