Effects of P2P Streaming on Video Quality Csaba Kiraly, Luca Abeni, Renato Lo Cigno DISI – University of Trento, Italy
ICC 2010, Cape Town, May Problem domain P2P Streaming, also known as P2P-TV Examples you might know PPLive, SoapCast, TVAnts, etc. How they work? A source generates encoded audio/video This media stream is divided into chunks Various peers receive the encoded media and contribute to the diffusion, by forwarding received chunks to other peers Live stream, so delay does matter!
ICC 2010, Cape Town, May Studying chunk diffusion Numerous simulators are available to study these systems P2PTVSim, PeerSim, SSSim, etc. These provide answers like 0.8 Mb/s with 4% chunk loss ↔ 0.7 Mb/s with 2% loss
ICC 2010, Cape Town, May Our Contribution Methodology and tools for the comparison of P2P-TV systems through the evaluation of received video quality As seen by the user Simulation driven by real video traces Instead of simplifying assumptions, like “… lets assume the video is 1Mb/s CBR …” Initial evaluation using the new tool Choice of encoding rate Confronting chunkization schemes Various codecs
ICC 2010, Cape Town, May Methodology and Tools P2P Simulator overlay topology chunk rate and size chunk and peer schedulers chunk loss Encoder Chunkizer Remove lost chunks Fill missing frames Compare Raw video stream chunk time and size trace corrupted stream refilled stream chunks PSNR, SSIM, etc. Codec Encoding rate GOP size etc. Chunk forming: fixed size 1 chunk = 1 GOP 1 chunk = N frames Use codec’s error concealment replicate last decoded frame
ICC 2010, Cape Town, May Simulation parameters 1000 peers Push based operation based on buffer map of neighbours Overlay unstructured random regular graph overlay with degree 20 Network Access link constrained Homogeneous upload bandwidth of 1 Mb/s Download bandwidth is not a bottleneck Raw stream: “foreman” sequence looped 4 times Encoding: H.264 using ffmpeg and x264
ICC 2010, Cape Town, May Streaming rate vs. chunk loss Curves became flat: Quality gained by increased encoding rate is lost during transmission
ICC 2010, Cape Town, May Blind vs. media-aware chunkization Comparing chunk creation policies Blind: each chunk has same fix size, independent of stream structure Media-aware: respect frame boundaries, e.g. 1 chunk = 1 GOP (Group Of Pictures)
ICC 2010, Cape Town, May The effect of schedulers Comparing different chunk and peer selection policies a good scheduler ensures lossless delivery with low delay Guarantees unaffected PSNR to the users.
ICC 2010, Cape Town, May Choosing the right codec Evaluated received video quality as a function of video rate with 4 codecs No real surprises H.263 < MPEG2 < MPEG4 < H.264 Slight differences in the optimal working point Because of different error concealment implementations
ICC 2010, Cape Town, May Conclusions Simulating the P2P system and evaluating the quality of video “as seen by users” is feasible The proposed methodology allows joint evaluation of media encoding, chunkization strategies, and “traditional” peer parameters, such as scheduling and overlay algorithms Tool-chain available as open source GPL code: Questions?