Experiences with a Large-Scale DeploymentExperiences with a Large-Scale Deploymentof Stanford Peer-to-Peer Multicastof Stanford Peer-to-Peer Multicast

Experiences with a Large-Scale DeploymentExperiences with a Large-Scale Deploymentof Stanford Peer-to-Peer Multicastof Stanford Peer-to-Peer Multicast

JeongJeonghhun Nohun Noh, Pierpaolo Baccichet, Pierpaolo Baccichet**, Bernd Girod, Bernd Girod

Image, Video and Multimedia Systems GroupImage, Video and Multimedia Systems GroupInformation Systems LaboratoryInformation Systems Laboratory

Stanford UniversityStanford University

**Now with Dyyno IncNow with Dyyno Inc..

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 2

Outline

Review of Stanford P2P Multicast (SPPM) Experimental setup– System configuration– P2P multicast of ESWC

Analysis– Peer statistics– System performance

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 3

Stanford P2P Multicast (SPPM)

P2P-based low-latency and video-aware live streaming Multiple-multicast-tree overlay– Pushing video: reduces end-to-end delay and

duplicate video packets– Path diversity: improves resilience– Fine-grained bandwidth aggregation

Video/network aware packet handling Tested on PlanetLab

[Setton et al., P2PMMS 2005]

[Setton et al., Proc. IEEE 2008]

[Baccichet et al., ICME 2007]

[CoopNet 2002, SplitStream 2003]

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 4

Multiple Complementary Multicast Trees

Tree 1

Tree 2

… …Video stream

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 5

Packet Scheduling

Parent

PI B P B P B

7 1 6 1 4 1 2

[Setton et al., ICIP 2006]

Child 1

Child 2

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 6

Local Retransmission

P

PI B P ……

Parent 1

Parent 2 fails

Video Buffer

Child

[Setton et al., ICME 2006]

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 7

Outline

Overview of Stanford P2P Multicast Experimental setup– System configuration– P2P multicast of ESWC

Analysis– Peer statistics– System performance

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 8

System Configuration

Video source

System used for data collection

RetransmissionServer

Tree 1

Tree 2

Super node(s)

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 9

Technical Details

P2P streaming of ESWC 2008 Video– CG 3D game contents– Encoded using H.264/AVC– Bitrate: 600 kbps(video: 560kbps) – Resolution: 640 by 480 pixels

End-to-end delay < 10 seconds Data logging– Peer statistics– Packet loss ratio

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 10

Outline

Overview of Stanford P2P Multicast Experimental setup– System configuration– P2P multicast of ESWC

Analysis– Peer statistics– System performance

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 11

Audience Size

Day 1 Day 2 Day 3

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 12

Peer Uplink Bandwidth

20% of users supply 70% of aggregate uplink bandwidth Multiple trees allow more peers to contribute

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 13

Peer Lifetime

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 14

Departure Rate

30 minutes later…Will it stay in system?

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 15

Peer Accessibility

SYMM

ETRIC

STATIC IP

STATIC IP

(Fire

wall)

FULL C

ONE

RESTR. IP

RESTR. PORT

SYMM

ETRIC

STATIC IP

STATIC IP

(Fire

wall)

FULL C

ONE

RESTR. IP

RESTR. PORT

Access Type Distribution

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 16

Average Uplink Bandwidth

Data from Day 3

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 17

Packet Loss Ratio (PLR)

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 18

PLR over Time

Data from Day 3

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 19

Conclusions

Analysis of rich peer statistics 80/20 rule– 20% of users supply 70% of uplink bandwidth– 85% of users come from 20% of countries

Older peers are more reliable than younger peers Server assistance– 2.3x aggregate bandwidth w/o assistance– Extra bandwidth from super nodes

SPPM achieves low-latency transmission delay w/ low packet loss

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 20

Thank you!

Email: jhnoh@stanford.edu

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 21

Uplink Contribution – 80/20 Rule

Noh et al.: Experiences with a large-scale deployment of the SPPM May 11, 2009 22

Active Overlay Management

[Noh et al., ICME 2008]

[Noh et al., Globecom 2008]