Download presentation
Presentation is loading. Please wait.
Published byValentine Sullivan Modified over 9 years ago
1
Scheduling P2P Multimedia Streams: Can We Achieve Performance and Robustness? Luca Abeni, Csaba Kiraly, Renato Lo Cigno DISI – University of Trento, Italy kiraly@disi.unitn.it
2
IMSAA 2009, Bangalore, 9-11 December 2009 2 P2P Multimedia Streaming P2P is cool, but why streaming? Think of out-of-country TV broadcasting easier to get Internet connection than a satellite dish Think of the cost of starting a new TV channel traditional TV broadcasting vs. client-server vs. P2P P2P-TV could become one of the dominant multimedia applications on the Internet Some systems already deployed: PPLive, TVAnts, CoolStreaming, … with hundreds of channels already available
3
IMSAA 2009, Bangalore, 9-11 December 2009 3 P2P Multimedia Streaming contd. P2P-TV is resource-hungry previously unseen traffic volumes to/from the users 1+ mbit/s sustained download Even higher upload (if available) P2P-TV is challenging to design large peer count with heterogeneous networking resources This is not VoD, potentially millions of users watching the same live channel tight delay constraints This is not file sharing, delay is the design objective
4
IMSAA 2009, Bangalore, 9-11 December 2009 4 Achieve Performance & Robustness Several design challenges organizing and maintaining the P2P overlay scheduling information transmission between peers etc. In this work, we concentrate on scheduling for chunk-based P2P streaming study different combinations of peer and chunk selection strategies propose a new peer selection strategy that achieves both performance and robustness
5
IMSAA 2009, Bangalore, 9-11 December 2009 5 Outline of Talk P2P streaming systems, definitions The scheduling problem Chunks selection strategies (RUc, LUc, DLc) Peers selection strategies (RUp, MDp, ELp, BA W p) The optimal ones … are these robust? Bandwidth-Aware ELp Algorithm (BAELp) Algorithms Comparison
6
IMSAA 2009, Bangalore, 9-11 December 2009 6 P2P Streaming Systems 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 The system is unstructured No fixed distribution tree Each peer is connected to a small subset of the other peers (neighbourhood) Chunks are exchanged among neighbour peers
7
IMSAA 2009, Bangalore, 9-11 December 2009 7 The Scheduling Problem Each peer Receives chunks from the other peers Redistributes chunks to neighbour peers Scheduling decision at the sender peer Which chunk to send? (chunk selection) To which neighbour send a chunk? (peer selection) 2 variants Chunk first selection (XXc/XXp) Peer first selection (XXp/XXc) We concentrate on chunk first selection!
8
IMSAA 2009, Bangalore, 9-11 December 2009 8 Chunk Selection Random Useful (RUc): select among the chunks useful to at least one neighbour with uniform random choice Rationale: If there is enough bandwidth, sooner or later useful chunks get there easy to implement, widely used as baseline performance Latest Useful (LUc): Rationale: spread new chunks as fast as possible Shown to be fragile: older chunks can be "overtaken“ by newer ones, stopping their diffusion This fragility increases as neighbourhood size is reduced
9
IMSAA 2009, Bangalore, 9-11 December 2009 9 Chunk Selection contd. Deadline-based scheduler (DLc): Rationale: embed meta-information in the chunk instance Each copy of each chunk is associated a scheduling deadline, initialized to the chunk generation time Deadline of the chunk instance in the sender peer is postponed each time chunk is sent The useful chunk with the earliest deadline is selected shown to overcome problems of LUc No “overtaking” effect good performance with small neighbourhood size We will use DLc in this paper!
10
IMSAA 2009, Bangalore, 9-11 December 2009 10 Peer Selection Random Useful Peer (RUp): Uniform random choice among the peers that need the given chunk Bandwidth Aware Peer scheduler (BA W p): Rationale: peers with high upload bandwidth has high redistribution potential randomly selects a target (as in RUp); the probability of selecting P j is proportional to its output bitrate.
11
IMSAA 2009, Bangalore, 9-11 December 2009 11 Peer Selection contd. Earliest-Latest Peer (ELp): Rationale: key to fast diffusion is to choose a peer that can re-distribute the chunk Check the latest chunk owned by each peer And select as a target the peer with the earliest latest chunk
12
IMSAA 2009, Bangalore, 9-11 December 2009 12 The Optimal Ones ELp shown to be optimal in idealized conditions Homogeneous peers: for each peers upload bandwidth = stream bandwidth What happens in heterogeneous networks? BA w p Shown to achieve good performance in largely heterogeneous networks But it falls back to RUp for homogeneous networks! Are any of these robust to various network scenarios?
13
IMSAA 2009, Bangalore, 9-11 December 2009 13
14
IMSAA 2009, Bangalore, 9-11 December 2009 14 Bandwidth-Aware ELp Algorithm Goal: blend the best properties of bandwidth aware heuristics with ELp optimality 1 st approach: hierarchical scheduling EL BA p: use EL first. If there is a tie, apply BA among winners BA EL p: BA first, EL after
15
IMSAA 2009, Bangalore, 9-11 December 2009 15 Bandwidth-Aware ELp Algorithm 2 nd approach: weighted combination Instead of minimizing L(P j, t) the ID of the latest chunk of neighbour node P j Consider also Expected arrival of the chunk to P j, though the bandwidth of the sender s(P i ) Redistribution potential of P j through the bandwidth of the target peer s(P j ). Maximize: t − L(P j, t) + B w (s(P j )/s(P i )) Where BW is a weight assigned to the upload bandwidth
16
IMSAA 2009, Bangalore, 9-11 December 2009 16 Algorithms Comparison We use the P2PTVSim simulator Open source, event-driven, chunk level simulation available at http://www.napa-wine.euhttp://www.napa-wine.eu Critical resource is the overall upload bandwidth in the system We model the network as upload bandwidth limits at the peer’s access link Download bandwidth assumed to be unlimited We study three bandwidth distribution scenarios Each scenarion has a [0..1] heterogeneity parameter
17
IMSAA 2009, Bangalore, 9-11 December 2009 17 Bandwidth Distribution Scenarios We fix the average upload bandwidth at 1 (the source rate) The 3-class scenario ADSL like bandwidth distribution High-, mid- and low-bandwidth classes h: heterogeneity factor [0..1]
18
IMSAA 2009, Bangalore, 9-11 December 2009 18 Bandwidth Distribution Scenarios contd. Uniformly distributed scenario Peer bandwidth taken from a uniform distribution [1-Δ B,1+Δ B ] To avoid artifacts due to class-based distributions Free-rider scenario With peers that only leach, do not contribute
19
IMSAA 2009, Bangalore, 9-11 December 2009 19 3-class scenario 90 th percentile as a function of heterogeneity neighbourhood size 20 playout delay 50 600 peers 2000 chunks. Uniform scenario
20
IMSAA 2009, Bangalore, 9-11 December 2009 20 Excess resources What if excess upload bandwidth is available? Performance improves and differences diminish BAELp uses bandwidth more efficiently neighbourhood size 20; playout delay 50; Uniform with B = 0.8; N = 1000 peers, Mc = 2000 chunks.
21
IMSAA 2009, Bangalore, 9-11 December 2009 21 Free-riders What if some users don’t (or can’t) contribute? Non BA algorithms (even EL BA p) fail at 15-20% of free-riders BAELp remains top performer neighbourhood size 100; playout delay 50: F90 versus the fraction of the free riders. B = 1, N = 1000 peers, Mc = 2000 chunks.
22
IMSAA 2009, Bangalore, 9-11 December 2009 22 Summary and Future Work Summary We have compared several scheduling algorithms from previous literature, showing their weaknesses Designed the BAELp algorithm, which outperforms other algorithms in a large number of scenarios Our future work Formal analysis of BAELp, and its weight parameter Improve simulations with video trace driven chunk generation and evaluation of the received video quality
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.