congestion-aware video streaming over peer-to-peer networks eric setton information systems...

Congestion-Aware Video Streaming over

Peer-to-Peer Networks

Congestion-Aware Video Streaming over

Peer-to-Peer Networks

Eric SettonEric Setton

Information Systems LaboratoryInformation Systems LaboratoryStanford UniversityStanford University

22Eric Setton: Congestion-aware video streaming over peer-to-peer networks January 31st 2006

Multimedia Distribution Multimedia Distribution over the Internetover the Internet

• BitTorrent, Kazaa, Napster

• Self-scaling, peers relay data

• Typical latency: a few hours ~ a few days

P2P file transfer networkP2P file transfer network

Seed peer

Content delivery networkContent delivery network

Media server(s)

. . . . . . . . . . . .

• Content providers, Akamai

• Over-provisioned infrastructure e.g. 1500 servers deployed by AOL for Live 8 concert in July 2005

• Typical latency: a few seconds


Live Multicast over Peer-to-Peer Live Multicast over Peer-to-Peer

Challenges• Limited bandwidth • Delay due to multi-hop transmission• Dynamic behavior of peers

Related work• Coopnet

[Padmanabhan, Wang and Chou, 2003]• Coolstreaming

[Zhang, Liu, Li and Yum, 2005]

Approach• Determine encoding rate as a function of network bandwidth• Adapt media scheduling to network conditions and to content• Build and maintain multicast trees, request retransmissions to mitigate losses

… …Video stream


• Distortion model for throughput-limited streaming– Server-client scenario– Extension to peer-to-peer

• Packet scheduling for low-latency video streaming – Congestion-distortion optimized (CoDiO) scheduling– Low complexity (CoDiO light) scheduler

• Peer-to-peer video streaming– Distributed peer-to-peer protocol– Receiver-driven CoDiO light retransmission requests– Sender-driven CoDiO light prioritization

Presentation OutlinePresentation Outline


Video Distortion with Self CongestionVideo Distortion with Self Congestion

Goodpicturequality

Badpicturequality

Bit-rate (kb/s)

Self congestioncauses late loss


ModelModeling Streaming Performanceing Streaming PerformanceP

SN

R (

dB

)

Plo

ss

Bit-rate (kb/s) Bit-rate (kb/s)

[Stuhlmüller, Färber, Link and Girod, 2000]


Influence of T and Influence of T and KK

Foreman Salesman

Simulations over ns-2

Bandwidth 400 kb/s


Bandwidth 400 kb/s

165 % provisioning 110 % provisioning


Extension to Peer-to-PeerExtension to Peer-to-Peer

“Sender” peer

“Sender” peer

“Sender” peer

“Receiver” peer


Multi-Sender ScenarioMulti-Sender Scenario


Bandwidth reserved at each sender 200 kb/s

Latency constraint T=0.4 s


Bandwidth reserved at each sender 200 kb/s

Latency constraint T=0.4 s

Foreman Salesman


1 sender

380 kb/s, 36 dBHighest sustainable video quality

420 kb/s, 33.7 dB

Foreman Sequence


Streaming Streaming over aover a Bottleneck Bottleneck Link Link

Random cross traffic

Low bandwidthuplink

Video traffic

Acknowledgments

High bandwidth links

C


Related WorkRelated Work

• TCP-friendly rate-control– Indicates average rate as a function of collected statistics– Does not indicate any particular schedule

• Rate-Distortion optimized scheduling (RaDiO)– Formalization of the multimedia scheduling problem

– Adapted the framework to video streaming– Large gains compared to sequential scheduler

[Chou and Miao, 2001]

[Chakareski and Girod, 2002-5]

[Kalman and Girod 2003-5]

[Floyd et al., 1997]


Congestion-Distortion Congestion-Distortion Optimized Scheduling (CoDiO)Optimized Scheduling (CoDiO)

• Congestion as a new metric–Adaptive to network conditions–Reflects the impact of a sender–Unbounded as it reaches capacity

Decide which packets to send (and when) to maximize picture quality while minimizing network congestion

CRate

Congestion Δ(seconds)

• Principle of CoDiO


Estimating Estimating SSelfelf C Congestionongestion

defined as average end-to-end delay

t0 t1 t2 t3 t4 t5 t6

Siz

e o

fb

ottle

ne

ck q

ue

ue

R(t3)decrease rate C, the capacityR(t0)

R(t1)

R(t4)

average queue size


• Decoding buffer state sets the display outcome

• Expected distortion over a horizon of M frames is

Predicting Video DistortionPredicting Video Distortion

P PI P P P

Original picture Picture shown Error picture


Parameterized Delay DistributionParameterized Delay DistributionP

rob

abili

ty

dis

trib

uti

on

delay

• Self congestion affects bottleneck queue length• Model distribution by exponential with varying shift


Finding the Finding the BBest est TTransmission ransmission OrderOrder

• Optimal schedule is

• Why is this difficult ?– Large search space– Tight coupling between schedules

• How to solve this problem ?


Find a Good Schedule… at Random!Find a Good Schedule… at Random!

I B B B P

B

IB

B

P

Pictures to sendSchedule

I P B BBP I BBB I

B


CoDiO Light SchedulerCoDiO Light Scheduler

• Select iteratively most important video packet to transmit or retransmit

• Space transmissions to limit congestion over bottleneck

• Sufficiently simple to be run at each peer

PBP PP B B P PI B B B P …… PBP PP B B P PI B B B P …… PP PP B P PI B B B P …… B B P


CoDiO Scheduling PerformanceCoDiO Scheduling Performance


Packet loss rate 2%

Bandwidth 400 kb/s

Propagation delay: 50ms


Packet loss rate 2%

Bandwidth 400 kb/s


30 %

25 %

Mother & Daughter News


H.264/AVC @250 kb/sBandwidth 400 kb/s, propagation delay 50 ms

2 % packet loss0.6 second latency

CoDiO ARQ


CoDiO vs. RaDiOCoDiO vs. RaDiO

Sequence: Mother & Daughter

Packet loss rate 2%

Bandwidth 400 kb/s


Sequence: Mother & Daughter

Packet loss rate 2%

Bandwidth 400 kb/s


40 %


Sequence: Foreman

Packet loss rate 2%

Bandwidth 400 kb/s


Sequence: Foreman

Packet loss rate 2%

Bandwidth 400 kb/s


60 %

CoDiO vs. RaDiOCoDiO vs. RaDiO


Join ProcedureJoin ProcedureInitial join• Send request to source peer• Receive information on session

– List of connected peers– Video source rate– Bandwidth provisioning factor– Number of multicast trees

Probe peers• Find peers with available bandwidth• Select parents from candidates to

– Maximize diversity– Minimize tree height

Connect to multicast trees• Request attachment• Bandwidth reserved at each parent


DisconnectionsDisconnections

• Detection

• Reconnection

• Loss mitigation


Experimental SetupExperimental Setup• Network/protocol simulation in ns-2

– 300 active peers – Random peer arrival/departure

average: ON (5 min)/OFF (30 s) – Over-provisioned backbone– Typical access bandwidth distribution– Delay: 5 ms/link + congestion

• Video streaming– H.264/AVC encoder @ 250 kb/s– 15 minute live multicast

[Sripanidkulchai et al., 2004]

Downlink Uplink Percentage

512 kb/s 256 kb/s 56% 3 Mb/s 384 kb/s 21%1.5 Mb/s 896 kb/s 9% 20 Mb/s 2 Mb/s 3% 20 Mb/s 5 Mb/s 11%

Downlink Uplink Percentage

512 kb/s 256 kb/s 56% 3 Mb/s 384 kb/s 21%1.5 Mb/s 896 kb/s 9% 20 Mb/s 2 Mb/s 3% 20 Mb/s 5 Mb/s 11%


Join and Rejoin LatenciesJoin and Rejoin Latencies

Simulations over ns-2, 300 peers

Number of trees: 4

Retransmissions enabled


Number of trees: 4



Protocol ScalabilityProtocol Scalability


Video traffic

Control traffic

Protocol overhead


Video traffic

Control traffic

Protocol overhead


P PI B B B P ……

Receiver-Driven CoDiO Light Receiver-Driven CoDiO Light Retransmission RequestsRetransmission Requests

• Determine missing packets• Iteratively request most important packet• Limit number of unacknowledged retransmissions

P PI B B B P ……


% peersconnected to all 4 trees


With receiver-driven CoDiO light

Without retransmissions

Performance of Receiver-Driven Performance of Receiver-Driven CoDiO LightCoDiO Light




With content-oblivious retransmissions

Without retransmissions

Performance of Content-Oblivious Performance of Content-Oblivious Retransmission SchemeRetransmission Scheme


Receiver-driven CoDiO light No retransmissions

P2P Video Multicast: 64 out of 300 Peers

H.264 @ 250 kb/s2 second latency for all streams


P2P Video Multicast: 64 out of 300 Peers

H.264 @ 250 kb/s2 second latency for all streams

Receiver-driven CoDiO light No retransmissions


• Scheduler iteratively selects

• Intervals between transmission sufficient to– Mitigate any congestion of the uplink– Reserve rate for control traffic

Sender-Driven CoDiO Light PrioritizationSender-Driven CoDiO Light Prioritization

Sender

PI B P B P B

7 1 6 1 4 1 2


Performance ComparisonPerformance Comparison


Number of trees: 4



Number of trees: 4


30 %40 %

Foreman Mother & Daughter



Number of trees: 4



Number of trees: 4


35 %50 %

Foreman Mother & Daughter

Performance ComparisonPerformance Comparison


Sender-driven CoDiO light33.71 dB

Without prioritization30.17 dB

H.264 @ 250 kb/s0.8 second latency for all streams

Average Video Sequence for 75 Peers


Summary of ContributionsSummary of Contributions• Distortion model for throughput-limited streaming, captures influence of:

– Capacity– Latency constraint– Sequence dependence

• Packet scheduling for low-latency video streaming – Demonstrated the benefits of considering congestion as a metric– Congestion-distortion optimized packet scheduling

• Reduces latency vs. ARQ by up to 30%• Reduces congestion vs. RaDiO by up to 60%

– Evaluated RaDiO and CoDiO in a realistic network simulator– Developed real-time low-complexity CoDiO light scheduler

• Optimized packet scheduling for peer-to-peer networks– Optimized retransmission scheduling scheme– Prioritization from a sender to its receivers– CoDiO light reduces latency by up to 50%


Main PublicationsMain Publications

• Distortion model for throughput-limited streaming– Setton, Zhu & Girod, Proceedings ICME 2004

– Setton, Zhu & Girod, Proceedings ICIP 2004

– Zhu, Setton & Girod, Proceedings PCS 2004

– Zhu, Setton & Girod, Signal Processing: Image Communications, 2005

• Packet scheduling for low-latency video streaming– Setton & Girod, Proceedings MMSP 2004

– Setton, Zhu & Girod, Proceedings ISCAS 2005

– Setton, Yoo, Zhu, Goldsmith & Girod, Wireless Communications Magazine, 2005

• Optimized packet scheduling for peer-to-peer networks– Setton, Noh & Girod, Proceedings ACM Multimedia, 2005

– Setton, Noh & Girod, to appear Proceedings ICIP 2006


Other ContributionsOther Contributions(not covered in this presentation)(not covered in this presentation)

• Encoding of H.264/AVC SP and SI pictures– Rate-Distortion model for SP and SI pictures

– Optimal SP encoder settings for streaming with packet losses

– Contributed SP and SI encoder to H.264/AVC standard reference software

• Adaptive streaming of SP and SI pictures

• Streaming of SP and SI pictures over peer-to-peer networks

Setton & Girod, IEEE Transactions CSVT, accepted 2006

Setton, Ramanathan & Girod, Proceedings VCIP 2006

Setton & Girod, Proceeding VCIP 2005

Setton, Noh & Girod, Proceedings ACM Multimedia, 2005

congestion-aware video streaming over peer-to-peer networks eric setton information systems...

Documents

peer video

distributed peer

peer packet scheduling

video distortion

losses video stream

rate kbs self congestion

provisioning slide

seconds slide