On Peer-to-Peer Media Streaming Dongyan Xu Mohamed Heffeda Susanne Hamrusch Bharat Bhargava 2002 International Conference on Distributed Computing Systems.

On Peer-to-Peer Media Streaming

Dongyan Xu

Mohamed Heffeda

Susanne Hamrusch

Bharat Bhargava

2002 International Conference on Distributed Computing Systems

Outline

Introduction P2P Media Streaming Model Optimal Media Data Assignment Fast System Capacity Amplification Simulation Conclusion

IntroductionCategory of P2P system

Main difference between a general P2P system and a P2P media system is the data sharing mode

– Open-after-downloading mode– Play-while-downloading mode

IntroductionCharacteristics of a P2P media system

Self-growingThe more peers it serves ， the larger the capacity it will have

Server-lessSuch as opening a large number of simultaneous connection

HeterogeneousDifferent out-bound bandwidth contribution to the system

IntroductionCharacteristics of a P2P media system

Many-to-oneMultiple supplying peers in one real-time streaming session

IntroductionProblems of P2P media system

Media data assignment for a multi-supplier peer-to-peer streaming session

Fast amplification of the P2P streaming capacity

P2P Media Streaming Model

Roles of peersEach supplying peer participates in at most one P2P streaming session at any time

Bandwidth of peers

– R0 : denote the playback rate of the media data

– Rin(Pr) = R0

– Rout(Ps) = R0/2n (R0/2 , R0/4 , …. R0/2N)

P2P Media Streaming Model

Classes of peers– Classify the peers into N classes according to their

out-bound bandwidth offer– Class-n peer : offer out-bound bandwidth R0/2n (1

n N )≦ ≦ Capacity of the P2P streaming system

Segments of media data– Media data be partitioned into small sequential

segments of equal sizes– δt of each segment is the same

Optimal Media Data Assignment

Bad CaseRequesting peer : Pr

Supplying peers : P1s , P2

s , P3s , P4

s ( R0/2 , R0/4 , R0/8 , R0/8) P1

s : 8k+1 , 8k+2 , 8k+3 , 8k+4

P2s : 8k+4 , 8k+5

P3s : 8k+6

P4s : 8k+7

( k = 0, 1, 2, 3, …. )

Optimal Media Data Assignment

Optimal CaseRequesting peer : Pr

Supplying peers : P1s , P2

s , P3s , P4

s ( R0/2 , R0/4 , R0/8 , R0/8)

(1) The lowest class among supplying peer is class-n

(2) Computes the assignment of the first 2n segments

Optimal Media Data Assignment

The algorithm OTSp2p compute an optimal media data assignment achieves the minimum buffering delay

The minimum buffering delay

Fast System Capacity Amplification

Waiting time : interval between requesting peer first streaming request and the earliest time it can be admitted

T : duration of the P2P streaming session

Class-1 : P3s , P4

s , P3r

Class-2 : P1s , P2

s , P1r, P2

r

Average waiting time

(0+T+2T)/3 = T

Fast System Capacity Amplification

Average waiting time

(T+T+0)/3 = 2T/3

•Different admission decisions lead to different growth of streaming capacity

•Higher-class requesting peers will lead to a faster amplification of the system capacity

Fast System Capacity AmplificationDistributed admission control protocol (DACp2p)

Key features– Supplying peer can decides whether or not

to participate in a streaming session by probability

– Requesting peer may send a reminder to a busy supplying peer Ps

Fast System Capacity AmplificationDACp2p – Supplying Peers

Each Ps maintains an admission probability vector <Pr[1] , Pr[2], ..Pr[N]>

How to determine probability vector

1. Suppose Ps is class-k peerPr[i] = 1.0 when 1 i k≦ ≦Pr[i] = 1/2i-k when k<i N≦class i is favored class of Ps , if Pr[i] =1.0

2. If Ps idle ， then probability vector will be updated after a period of Tout

k < i N, Pr[i] = Pr[i]*2≦

Fast System Capacity AmplificationDACp2p – Supplying Peers

(3) If Ps finished serving a streaming ， will update its probability vector

During the streaming session ， did not receive any request of its favored classk < i N, Pr[i] = Pr[i]*2≦

If received one request of its favored class ， request peer left a reminder to Ps ， if k is the highest favored class of requesting peer which left a reminder Pr[i] = 1.0 when 1 i k≦ ≦Pr[i] = 1/2i-k when k<i N≦

Fast System Capacity AmplificationDACp2p – Requesting Peers

Pr obtains a list of M randomly supplying peers ， and directly contact the candidate from high to low classes

Pr will be admitted– Pass the probabilistic admission test– Rsum = R0

Pr will be rejected – Pr will leave a reminder to a busy Ps who currently

favors the class of Pr

– Backoff for at least a period of Tbkf befor making the request again

Simulation

Number of requesting peers : 50000 Number of seed supplying peers : 100 Each seed peers is a class-1 peer Show time of video : 60 min Class distribution of requesting peers

class-1 : 10%class-2 : 10%class-3 : 40%class-4 : 40%

M = 8 Tout = 20 min Tbkf = 10 min Simulate period : 144 hours During the first 72 hours ， the 50000 peers make their first

streaming requests

Simulation

System capacity amplification

Simulation

Request admission rate

DACp2pNDACp2p

Simulation

Average buffering delay

DACp2p NDACp2p

Conclusion

Algorithm OTSp2p which computes optimal media data assignments for P2P media streaming

Algorithm DACp2p which achieves fast system capacity amplification and creates an incentive for peers to offer their truly available out-bound bandwidth