Professional Documents
Culture Documents
August 2010
August 2010
Outline
Outline
August 2010
1 / 29
Outline
Outline
1
Introduction Motivations Denitions Overlay Topology Design The Strategy Chunk Signaling and Scheduling Performance Evaluation Model
Network Scenario Video Parameters
Results
3
4 5
Introduction
Introduction
August 2010
3 / 29
Introduction
Motivations
August 2010
4 / 29
Introduction
Motivations
August 2010
4 / 29
Introduction
Denitions
Basic concepts
A source peer splits the video streams in small chunks that are injected inside the overlay. Peers start exchanging chunks according to some scheduling scheme. Chunks are exchanged among peers that are neighbors of each others. P2P streaming systems, dierently from le sharing, have to face with the real-time constraint!
Every chunk must be received within a deadline Dmax also called playout delay.
August 2010
5 / 29
Introduction
Denitions
Assumptions
Peers are Internet nodes and tipically their upload bandwidth is much lower than the downlink one. Peers upload bandwidth and latencies between peers are supposed to be known somehow. We neglect the eect of churning since scheduling dynamics are much faster. Peers exchange signaling information to trade chunks.
August 2010
6 / 29
Introduction
Denitions
Assumptions
Peers are Internet nodes and tipically their upload bandwidth is much lower than the downlink one. Peers upload bandwidth and latencies between peers are supposed to be known somehow. We neglect the eect of churning since scheduling dynamics are much faster. Peers exchange signaling information to trade chunks.
August 2010
6 / 29
Introduction
Denitions
Assumptions
Peers are Internet nodes and tipically their upload bandwidth is much lower than the downlink one. Peers upload bandwidth and latencies between peers are supposed to be known somehow. We neglect the eect of churning since scheduling dynamics are much faster. Peers exchange signaling information to trade chunks.
August 2010
6 / 29
Introduction
Denitions
Assumptions
Peers are Internet nodes and tipically their upload bandwidth is much lower than the downlink one. Peers upload bandwidth and latencies between peers are supposed to be known somehow. We neglect the eect of churning since scheduling dynamics are much faster. Peers exchange signaling information to trade chunks.
August 2010
6 / 29
August 2010
7 / 29
The Strategy
Intuition
To avoid long trading phases, peers with short end-to-end latencies should be connected to each other. To speed up chunk replication, high bandwidth peers should be well connected to each other (to the source). The number of neighbors of a peer have to choose, should be related to its upload bandwidth.
August 2010
8 / 29
The Strategy
Intuition
To avoid long trading phases, peers with short end-to-end latencies should be connected to each other. To speed up chunk replication, high bandwidth peers should be well connected to each other (to the source). The number of neighbors of a peer have to choose, should be related to its upload bandwidth.
August 2010
8 / 29
The Strategy
Intuition
To avoid long trading phases, peers with short end-to-end latencies should be connected to each other. To speed up chunk replication, high bandwidth peers should be well connected to each other (to the source). The number of neighbors of a peer have to choose, should be related to its upload bandwidth.
August 2010
8 / 29
The Strategy
The Strategy
How many neighbors?
Being Kp the number of neighbors of a peer p chooses, we set Kp = max (3, 10 Bp /rs ) where Bp is the upload bandwidth of peer and rs is the average video rate of the stream.
A pull mechanism
Every peer periodically generates oer messages to publish the list of its useful chunks. Neighbors reply to each oer with a select message in which they specify the chunk they need. Once the select message is received, the chunk is then transmitted. When the chunk is received, an acknoledgement message is sent back to the trasmitter.
P5 P2
OFFERS SELECTS
P1 P2 Np
P7
OFFERS
P4
SELECTS
Np
Chunk #1 to Peer 2
time
Chunk #1 to Peer 5 Chunk #2 to Peer 1 Chunk #2 to Peer 2 Chunk #2 to Peer 7 Chunk #3 to Peer 4
August 2010
10 / 29
August 2010
11 / 29
Nework Scenario
Our simulation involved 2000 peers partitioned in four classes according to their upload capacity: Class 1: 5.0Mb/s 10%; Class 2: 1.6Mb/s 10%; Class 3: 0.64Mb/s 10%; Class 4: 0Mb/s. We consider 4 scenarios, with increasing heterogeneity. The average upload bandwidth is E [Bp ] = 1.3Mb/s in all cases. Class H = 0.01 H = 0.05 H = 0.10 H = 0.15
S. Traverso (Politecnico di Torino)
1 1 5 10 15
4 20 20 20 20
August 2010 12 / 29
0.05
0.1
0.25
0.3
0.35
The videos consists of 1000 frames 40s of visualization. H.264/AVC codec has been adopted for encoding sequences. Hierarchical structure of GOP: frames can be IDR, P, B or b. Intra frames (IDR) carry valuable information (bigger), inter frames (P,B or b) carry dierential information (and are much smaller).
August 2010
14 / 29
Results
The Impact of
45 40 PSNR [dB] 35 30 25 EVQ Dmax=6s Dmax=5s
Dmax=4s Dmax=3s 1
Results
The Impact of II
80 60 Latitude [Deg] Latitude [Deg] 40 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150 80 60 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150 Latitude [Deg] 40 80 60 40 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150
Figure: = 0.0
Figure: = 0.1
Figure: = 1.0
August 2010
16 / 29
Results
The Impact of II
80 60 Latitude [Deg] Latitude [Deg] 40 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150 80 60 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150 Latitude [Deg] 40 80 60 40 20 0 -20 -40 -60 -80 -150 -100 -50 0 50 Longitude [Deg] 100 150
Figure: = 0.0
46 44 PSNR [dB] 42 40 38 36
Figure: = 0.1
Figure: = 1.0
Results
Dmax=4s Dmax=3s 1
Fixed Kp = 20 (left) and variable Kp = max (3, 10 Bp /rs ) (right) for = 0.9.
August 2010
17 / 29
Results
Network stress
0.1 0.09 Network stress [ms] 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0 0.2 0.4 Dmax=3s Dmax=4s Dmax=5s Dmax=6s 0.6 0.8 1
Figure: The network stress index, i.e. the average distance covered by chunks expressed it terms of the corresponding latency ( = 0.9).
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 18 / 29
Results
Robustness
46 45.5 45 PSNR [dB] 44.5 44 43.5 43 42.5 42 41.5 41 0.025
PSNR [dB]
45 40 35 30 25 Paris Foreman Pink Paris EVQ Foreman EVQ Pink EVQ 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.22
0.415
0.61
0.805
Figure: Average PSNR versus for Figure: Average PSNR versus for dierent degrees of heterogeneity H with dierent video sequences with = 0.9, H = 0.10 and variable Kp . Dmax = 5s and = 0.9.
August 2010
19 / 29
Video-Aware Schedulers
Video-Aware Schedulers
August 2010
20 / 29
Video-Aware Schedulers
10
15
20 25 Time [s]
30
35
40
55 50 45 40 35 30 25 20 15
PSNR [dB]
10
15
20 25 Time [s]
30
35
40
Figure: PSNR variation of a random peer versus time. = 0.6 (rs = 780kb/s, top) and = 1.0 (rs = 1290kb/s, bottom).
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 21 / 29
Video-Aware Schedulers
The Strategy
We can assign priority to chunks transporting precious frames. To avoid chopping frames into several chunks, we set one frame per chunk. Priority is assigned to frames based on the amount of degradation they might induce if lost: q is the number of subsequent frames that would be aected by the lost of that frame. A weight q is given to every chunk encapsulating a frame. With > 0 we assign a larger weight to more important frames.
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 22 / 29
Video-Aware Schedulers
The Strategy
We can assign priority to chunks transporting precious frames. To avoid chopping frames into several chunks, we set one frame per chunk. Priority is assigned to frames based on the amount of degradation they might induce if lost: q is the number of subsequent frames that would be aected by the lost of that frame. A weight q is given to every chunk encapsulating a frame. With > 0 we assign a larger weight to more important frames.
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 22 / 29
Video-Aware Schedulers
Results
Slight Improvements
42 Pink 40 PSNR [dB] 38 36 34 32 5 10 15 20 25 PeerID =0.0 =0.5 =1.0 =2.0 30 35 40 45
Figure: PSNR for dierent peers and values of with = 1.1, Dmax = 5s and H = 0.10. Pink by Aerosmith video sequence.
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 23 / 29
Video-Aware Schedulers
Results
Slight Improvements II
40 38 36 PSNR [dB] 34 32 30 28 26 5 10 15 20 25 PeerID =0.0 =0.5 =1.0 =2.0 30 35 40 45 Paris
Figure: PSNR for dierent peers and values of with = 1.1, Dmax = 5s and H = 0.10. Paris video sequence.
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 24 / 29
Video-Aware Schedulers
Results
Figure: PSNR for dierent peers and values of with = 1.1, Dmax = 5s and H = 0.10. Foreman video sequence.
S. Traverso (Politecnico di Torino) Overlay Topology Design Tradeos August 2010 25 / 29
Conclusions
Conclusions
August 2010
26 / 29
Conclusions
Conclusions
We provided guidelines for the design of the overlay topology and the chunk scheduling algorithm. By carefully designing the overlay topology we can partially localize the trac and improve the user QoE. By prioritizing chunks that encapsulate valuable pieces of information at the scheduler level, system performance can be slightly improved in overloaded conditions.
August 2010
27 / 29
Conclusions
Conclusions
We provided guidelines for the design of the overlay topology and the chunk scheduling algorithm. By carefully designing the overlay topology we can partially localize the trac and improve the user QoE. By prioritizing chunks that encapsulate valuable pieces of information at the scheduler level, system performance can be slightly improved in overloaded conditions.
August 2010
27 / 29
Conclusions
Conclusions
We provided guidelines for the design of the overlay topology and the chunk scheduling algorithm. By carefully designing the overlay topology we can partially localize the trac and improve the user QoE. By prioritizing chunks that encapsulate valuable pieces of information at the scheduler level, system performance can be slightly improved in overloaded conditions.
August 2010
27 / 29
Q&A
Q&A
August 2010
28 / 29
Q&A
August 2010
29 / 29