On Peer-to-Peer Media Streaming

by Dongyan Xu, Mohamed Hefeeda, Susanne Hambrusch, Bharat Bhargava

Dept. of Computer Science, Purdue University, West Lafayette

Contents Introduction Streaming Model Media Data Assignment Admission Control Protocol Simulation Results Conclusion

Introduction General P2P System (File)

‘open-after-downloading’

P2P Media Streaming System ‘play-while-downloading’

Characteristics shared by both categories Self-growing (capacity amplification) Server-less (no server-like behavior) Heterogeneity (bandwidth) (authors omitted storage capacity heterogeneity)

Introduction Characteristic owned by P2P Media Streaming

System Multiple supplying peers

Introduction Two problems addressed

Media data assignment Fast amplification of streaming capacity

Two solutions proposed OTSp2p – optimal media data assignment

DACp2p – distributed differentiated admission control protocol

Streaming Model Assumptions:

CBR Video bitrate R0,

Can be partitioned into equal size segments of playback time

Roles of peers: each supplying peers join at most one session at any time

Bandwidth of peers: Out-bound bandwidth of supplying peer Ps:

0 0 0( ) , , ,2 4 2out s N

R R RR P

t

This set of values prevents the assignment problem from becoming the NP-hard binpacking-like problem.

Streaming Model Assumptions:

Classes of peers: N classes according to N values of their out-bound bandwidth,

System capacity: Sum of out-bound bandwidth

( )

0

( )

( ) s s

out sP P t

sys

R P

C tR

0class-n2nR

Optimal Media Data Assignment Goals:

Continuous playback Minimum buffering delay at Pr

To determine: Media segments being transmitted by Playback start time

Example: Supplying peers are with out-bound

bandwidth of

(1 )isP i m

1 2 3 4, , ,s s s sP P P P

0 0 0 02, 4, 8, 8R R R R

Optimal Media Data Assignment Different assignments lead to different buffering

delay Assignment 1: buffering delay = 5 t

Optimal Media Data Assignment Different assignments lead to different buffering

delay Assignment 2: buffering delay = 4 t

Optimal Media Data Assignment Algorithm OTSp2p

m supplying peers sorted in descending order in out-bound bandwidth,

Lowest class among them is class-n Alogrithm:

0 r1

(P ) ( )m

iin out s

i

R R R P

Optimal Media Data Assignment Theorem

Given m supplying peers

OTSp2p will compute an optimal data assignment Achieves the minimum buffering delay

(1 )isP i m

0 r1

(P ) ( )m

iin out s

i

R R R P

tmTbuf min

Admission Control Protocol Requirements:

Should not starve the lower-class peers Purely distributed fashion Differentiation – the higher the outbound bandwidth,

the greater probability being admitted, with shorter waiting time and buffering delay

DACp2p Characteristics: Each supplying peer operates individually with

requesting peer Operate in a probabilistic fashion

Admission Control Protocol DACp2p – Supplying Peers

Probabilistic vector For For

If being idle for Tout, ‘relaxes’ the admission preference

After serving peer, If no ‘reminder’ received, ‘relaxes’ the admission preference If certain ‘reminder’ received before, ‘tightens’ the

admission preference

]Pr[,],2Pr[],1Pr[ N

0.1]Pr[,1 ikikiiNik 21]Pr[,

Admission Control Protocol DACp2p – Requesting Peers

Randomly select M supplying peers via some peer-to-peer lookup mechanism

Pr will be admitted if obtains enough permissions among the M peers such that

they are neither down nor busy willing to provide the service their aggregated out-bound bandwidth is enough

then execute OTSp2p to compute the data assignment

Admission Control Protocol DACp2p – Requesting Peers

Pr will be rejected not enough permissions from these M peers leaves a ‘reminder’ to a subset W W is chosen from busy peers as follows:

currently favors the class of Pr

the aggregated out-bound bandwidth offered by W is equal to

Backoff for at least a period of Tbkf before another request xth rejection, backoff period =

sumRR 0

1 xbkfbkf ET

Note that the rejected peer may not in the future being served by the exactly the same set of W.

Simulation Results Performance Metrics:

System capacity amplification Request admission rate Average buffering delay Average waiting time (before admission)

Simulation Results Simulation Environment

Total 50,100 peers (50,000 requesting + 100 ‘seed’) Video length = 60mins Supplying peer are class-1 peer Requesting peers: class(1, 2, 3, 4) = (0.1, 0.1, 0.4,

0.4) M = 8, probes 8 randomly selected supplying peers Tout = 20mins, Tbkf = 10mins, Ebkf = 2

Simulation time = 144 hrs, first request in first 72 hrs Comparison situation of non-differentiated admission

control protocol (NDACp2p):

0.1,,0.1,0.1

Simulation Results System Capacity Amplification

Simulation Results Request Admission Rate

Simulation Results Average buffering delay

Simulation Results Average Waiting Time

Given average number of rejections x, average waiting time can be computed as 1 x

bkfbkf ET

Conclusion Problems in Peer-to-Peer Media Streaming

Media data assignment Fast capacity amplification

Solutions Proposed Algorithm OTSp2p

Distributed DACp2p protocol

DACp2p Features Fast system capacity amplification Benefits all requesting peers in

admission rate waiting time buffering delay

Create an incentive of peers to offer truly available out-bound bandwidth

End of Presentation

Thank you!