1. A Method for P2P Streaming System Join Latency Reduction Using Preferred Peer Lists
Jeonghun Noh
Sachin Deshpande*
Information Systems Laboratory
Stanford University
* Sharp Laboratories of America

2. Toward Low Latency P2P Video Streaming
Real-time video streaming is time-sensitive.
Users prefer quick video play-back.
Channel surfing benefits from a short join latency.
Join latency in P2P systems
During join procedure, new peers connect to peers in the system to receive video.
Existing P2P networks: new peers blindly probe for parents at random.
Rapid channel switching, channel surfing.
GOAL: Design a P2P network that provides
a small join latency

3. Overview Preferred Peers List System Experimental Results Conclusions
Preferred Peer Cache/List
Parent switching
Soft Handover
Experimental Results
Conclusions

4. Overview: Preferred Peer List System
Preferred Peer Cache (PPC)
Caches information about available peers (preferred peers) in the overlay.
Maintained by the source peer
Preferred peers have a high probability of being available to other peers
Using PPC for peer recommendations
New incoming peers contact the source peer.
The source peer sends a Preferred Peer List (PPL) containing preferred peers.
The new peers directly contact the preferred peers to connect to.

5. Partial Learning of the Overlay
Source peer partially learns about the overlay from the peer’s join / leave reports S PX PY X Y
JOIN Report
LEAVE Report

6. Using the Knowledge about the Overlay
A list of peers, PPL, is created when a recommendation is requested from new peers.
Among the most recently updated peers, peers recommended less often are chosen.
Peer address | Use-count | Update time (in secs) 3
24.0 12 2
23.0 7
20.0
PPC
(at the source)
Selected peers

7. Join Procedure Using PPL
Initial join request
Contact video source
Receive number of trees, video rate
Receive flat list, preferred peer list (PPL) …
Video stream
Direct connection and peer probe
Contact preferred peers on PPL
Probe peers on flat list
Use probe replies;
if preferred peer attempt fails.
Selects best parent from probe replies
Parent selection factor
Available bandwidth
Minimize tree height
Combine the prev slide to this. Instead,
Note that on probing, send upstream hosts list to avoid loop.
Over provisioning 1.11 for Salesman, 1.5 for Foreman

8. Parent Switching
Preferred peer recommendations do not guarantee the best parents
One method to choose better parents
Utilize collected probe replies
Perform the probe-based join process

9. Distribution Tree at time 71s
No Parent Switch
: video source
: peer in the session
: peer that left the session
Can we allign slides? Also, get some values for tree height.
Average Tree Height : 3.76
Average PSNR : 39.71dB

10. Distribution Tree at time 71s
Parent Switch
: video source
: peer in the session
: peer that left the session
Can we allign slides? Also, get some values for tree height.
Average Tree Height : 2.67
Average PSNR : 40.13dB

11. Soft Handover Parent switching may cause video discontinuity.
Using Soft Handover
Keep old and new parents simultaneously so that both parents forward video packets.
As soon as duplicate packets are detected, send leave message to old parent.

12. Experimental Setup Network simulation in ns-2 Video streaming
75 / 300 active peers
Random peer arrival/departure average: ON (90 sec) / OFF (10 sec)
Peer uplink: typical access bandwidth / uniform bandwidth
Video streaming
Mother & Daughter sequence encoded with 281 Kb/s (CIF quality)
15-minute live multicast
System constructs 4 complimentary multicast trees.

13. Distribution of Join Latency
Comparison of join time distribution
PPL system versus a probe-based system
Influence of peer uplink bandwidth: heterogeneous / uniform distributions
Legend is blocking the curves.
One plot on one slide.
First tree join time
Full connection time

14. Effects of Parent Switching
Probe-
based
PPL
PPL with Parent Switch
PSNR (dB)
40.67
40.33
40.85
Average play-out latency (sec)
1.97
1.40
1.38
Bar graph. PSNR difference is so insignificant
Simulation setup: 500 secs, uniform bandwidth distribution for peer uplink
PSNR averaged over 10 simulations and 75 peers

15. Effects of Soft Handover
Observations
(+) Up to 55% of redundant packet forwarding is reduced
(-) Slight PSNR drop may occur
Ignorant
Handover
Soft
Number of duplicate packets
7,677
3,448
Average video quality (dB)
39.89
39.72
3448/7677 => 55% reduction , 278/618 => same. 55% reduction
75peers. Session time : 200secs

16. Conclusions
PPC keeps partial information about the P2P overlay system.
Preferred peers are selected from incomplete PPC.
One step closer to low-latency P2P video streaming:
New peers can join faster.
By switching parent after PPL-based direct attach, overlay quality is improved.
By Soft Handover, up to 55% of redundant video packet transmission can be avoided.