130 likes | 270 Views
Realistic Media Streaming over BitTorrent. George Xylomenos Mobile Multimedia Laboratory Greece. Outline. Context Motivation Advantages Streaming over BitTorrent Experimental setup Number of stall periods Average stall duration Download and seeding time Conclusion. Context.
E N D
Realistic Media Streaming over BitTorrent George Xylomenos Mobile Multimedia Laboratory Greece
Outline • Context • Motivation • Advantages • Streaming over BitTorrent • Experimental setup • Number of stall periods • Average stall duration • Download and seeding time • Conclusion
Context • The ICT PURSUIT Project • The Internet mostly disseminates data • Publish-Subscribe Internet Technology • Clean slate approach to Future Internet • What does that have to do with BitTorrent? • PURSUIT was motivated by content distribution • BitTorrent is the perfect benchmark for this! • Now if we only had a good BitTorrent simulator…
Motivation • Why a BitTorrent simulator? • BitTorrent swarms exhibit very complex behavior • Many mechanisms and strategies are at play • Traces are hard to gather and understand • Fully distributed systems are hard to monitor • Performance predictions are simply guesses • What happens when we modify a strategy? • Existing simulators were not detailed enough • Some omit large parts of the protocol • Others only work over abstract networks • Most are custom-built and hard to extend
Advantages • Why bother with our simulator? • It operates over the OMNeT++ platform • You can use it with everything available for OMNeT++ • Example: OMNeT++ supports two types of network • InetUnderlay: hosts with full TCP/IP stacks • SimpleUnderlay: simple and fast abstract hosts • It incorporates nearly all BitTorrent details • All policies and options are present and tunable • If something is missing, feel free to add it! • Extra features that simplify simulations • Asymmetric links, churn model, GT-ITM topologies • To probe further • K. Katsaros, V. Kemerlis, C. Stais, and G. Xylomenos, A BitTorrent module for the OMNeT++ simulator, IEEE MASCOTS, 2009
Streaming over BitTorrent • A case in point: Streaming over BitTorrent • A download window favors “sequential” downloads • Window moves on playback or piece download • If a piece is not available for playback, the player stalls • Three different proposals exist, but… • Undocumented simulation setups • Different evaluation metrics • Retransmissions and congestion are ignored • This paper avoids all these problems • Realistic simulations with the same assumptions • To probe further • C. Stais, G. Xylomenos, and A. Archontovasilis, A comparison of streaming extensions to BitTorrent, IEEE ISCC, 2011
Streaming over BitTorrent • Fixed-Size Window (FSW) • Fixed sliding window from first non-available piece • Rarest-first only within the window • High-Priority Set (HPS) • Fixed size window of non-downloaded pieces • Download outside the window with probability 1-p • Stretching Window (SW) • Also uses an HPS but only downloads inside it • Bounded distance between first and last piece
Streaming over BitTorrent Player Begin End OK OK D OK OK 1 2 3 4 5 6 7 8 FSW Player Begin HPS 3 4 7 8 End OK OK D OK OK 1 2 3 4 5 6 7 8 HPS Player Begin HPS 3 4 7 End OK OK D OK OK 1 2 3 4 5 6 7 8 SW
Experimental setup • Main parameters • Video: 200 MB @ 256 kbps (106 min) • 112 KB pieces (1 GOP = 3.5 seconds) • 4 core / 192 access routers • ADSL access links: Uplink 1-2 Mbps, Download 4-24 Mbps • 1 seeder and 120 peers, incremental joins • 50% of peers stay in the swarm until playback completes • Prefetch 1 or 5 pieces before playback • Window size 2% or 8% of file (36 or 147 pieces) • HPS probability 80% • SW limit 50 or 200 pieces • Two player modes • B1: the player stalls until the next piece completes • B2: the player stalls until the next three pieces complete
Number of stall periods • How many times will a user experience a stall? • 30% more when we stall only for a single piece (left) • In all cases, very few stalls for 100+ minutes • HPS is the worst performer, despite its complexity
Average stall duration • How long does each stall last? • 30% more when we stall for three pieces (right) • Stall durations vary from noticeable to quite long • HPS is the worst by far with a larger window
Download and seeding time • Download time: same for B1 and B2 • Better with HPS, but always lower than the video duration • Seeding time: same for B1 and B2 • Also better with HPS, as the download finishes earlier
Conclusion • Which scheme is best? • HPS leads to visibly worse user experience • More stalls with longer durations • FSW and SW are acceptable and very close • SW is not worth the extra complexity over FSW • FSW leads to 1-2.5 stalls of 40-60 seconds for 100+ minutes • The two buffering modes offer the expected tradeoff • Either fewer or longer stall periods • Future work • Explore additional parameters and tricks • Study individual peer performance