Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

35
Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar Vivek Shrivastava, Suman Bane rjee University of Wisconsin-Madis on, USA ACM NOSSDAV’05

description

Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar. Vivek Shrivastava, Suman Banerjee University of Wisconsin-Madison, USA ACM NOSSDAV’05. imposed incentive/rule. natural incentive. Cathedral and Bazaar. Cathedral. Bazaar. Introduction. - PowerPoint PPT Presentation

Transcript of Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

Page 1: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

Natural Selection in Peer-to-Peer Streaming: From the Cathedral to

the Bazaar

Vivek Shrivastava, Suman BanerjeeUniversity of Wisconsin-Madison, USA

ACM NOSSDAV’05

Page 2: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

2

Cathedral and Bazaar

Cathedral Bazaar

imposed incentive/rule natural incentive

Page 3: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

3

Introduction

Peer altruism (resource contribution) is a key factor of p2p applications

BitTorrent employs a tit-for-tat rule This work exploits scenarios in p2p stream

ing media applications where resource sharing is a natural behavior without external rules and incentives

Page 4: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

4

Outlines

BackgroundProposed Bazaar frameworkEvaluation and simulationSummary

Page 5: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

5

Background

not altruistic limited asymmetrical upstream bandwidth of

DSL and cable modems uploading may reduce its access bandwidth and

degrade the network access performance

Page 6: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

6

This paper

p2p streaming environment has inherent natural incentives for peers to contribute

form efficient overlay tree for data streaming

shift from Cathedral style to Bazaar style, where no rules are imposed on peers and resources sharing takes place naturally as peers try to maintain their perceived data utility

Page 7: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

7

Utility

Utility benefit(incoming bandwidth, latency) – cost (out

going bandwidth) maybe different in different peers

Page 8: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

8

An example

all selfish

Page 9: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

9

An example

Case (a) incoming bandwidth to peer A is 80kbps

Case (b) incoming bandwidth to peer A is 200kbps outgoing bandwidth from peer A is 400kbps

If the increase in A’s perceived data utility offsets the loss, A will do so

Page 10: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

10

An example

final overlay depends on the mix of utility functions of peers

peer B also has natural incentive to join under A as the streaming bandwidth increases from 80kbps to 200kbps

Page 11: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

11

This paper

provide a platform to facilitate the formation of such a mutually beneficial overlay without introducing any rules or incentives in the system

based on natural incentive of peer to conserve own incoming bandwidth by attracting the new entrant to join under itself rather than its parent

Page 12: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

12

Bazaar framework

regular peers strategic, maximize incoming and minimize

outgoing bandwidthBSE (Boot Strap Entity)

~tracker, provide overlay information to new node

root (publisher) altruistic, allocates fixed outgoing bandwidth

during the entire streaming session

Page 13: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

13

Market Quote (M-Quote)

Each node provides a M-Quote of its services to attract peers to join under it rather its parent, so as to preserve its own incoming bandwidth

has the following advertised components: bandwidth latency from root to the peer

kept in BSE

Page 14: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

14

Bazaar framework

peers can perform join/leave the overlay dynamically advertise their services participate in shuffle operation to improve

overlay structure

Page 15: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

15

Bazaar in action

1. root enters the system

2. A interested in the contentjoins the system

3. revise M-Quote (250 = 500/2)

4. if A is selfish, it can sendM-Quote of 0 or do notsend at all. However, a newnode will definitely choosesroot as peer which reduce the bandwidth shared by A. This motivates A to offer competitive quote

Page 16: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

16

Bazaar in action

B choose A if bandwidth outweighs latency

B choose A if latency outweighs bandwidth

Advertising a lower outgoing bandwidth which leads to competition of parent bandwidth, A may participatein a local shuffling operation

Page 17: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

17

Bazaar in action

shuffle is a periodic operation

if B is attracted by the new quote,the overlay changes accordingly

Page 18: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

18

Evaluation

Utility function

Y. hua Chum J. Chuang, and H. Zhang, “A case for taxation in peer-to-peer streaming broadcast,” Workshop on Practice and theory of incentives in networked systems (PINS), 2004

Page 19: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

19

Simulation environment

peer-to-peer network simulator myns, developed at University of Maryland

use Transit Stub topology generated by GT-ITM topology generator

50 peers, all results are averaged over 1000 permutations of peer join order

Page 20: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

20

Evaluate under modes of operation

Page 21: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

21

Performance metrics

throughput incoming bandwidth of each peer

total system utility sum of perceived data utility of all peers

Page 22: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

22

Bimodal simulations

all peers are categorized as high or low capacity peers

high capacity: outgoing bandwidth randomly selected from 500Kbps to 1Mbps

low capacity: random [50Kbps, 450Kbps]simulate heterogeneous peer environments

by varying fraction of high capacity peers from 0 to 1

max streaming rate is 500Kbps

Page 23: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

23

Bimodal simulations – system utility

Page 24: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

24

Bimodal simulations – system utility

Observations utility increases with fraction of high capacity,

as cost of forward (fraction of forward bandwidth over max outgoing bandwidth) reduces

strategic is better than random, as random my degrade it own and others utility

Page 25: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

25

Bimodal simulations – throughput

Page 26: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

26

Bimodal simulations – throughput

Page 27: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

27

Bimodal simulations – throughput

Observation gap between altruistic and strategic mode

decreases with decrease in fraction of high capacity peers

The fraction of high capacity peers in real life is low, so strategic peers can achieve good throughput in real life

Page 28: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

28

Trace based simulations

outgoing capacity distribution of peers are based on traces collected from Sigcomm, Slashdot and Gnutella

max streaming rate is 500Kbps

A. Bharambe, S. Rao, V. Padmanabhan, S. Seshan, and H. Zhang, “The impact of heterogeneousbandwidth constraints on DHT-based multicast protocols,” International Workshop on P2P Systems, 2005

Page 29: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

29

Trace based simulations

varying the fraction of strategic peers

Page 30: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

30

Trace based simulations

Page 31: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

31

Trace based simulations

Observations Sigcomm: performance of Bazaar degrade subs

tantially as the fraction of strategic peers increases

Slashdot and Gnutella: performance degrade gracefully

infer Bazaar framework is particularly well suited for many p2p streaming scenarios, in which peers are mostly resource poor

Page 32: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

32

Summary

Bazaar framework make use of the natural inherent incentive facilitate formation of efficient overlay structure improve performance in p2p streaming

applications involving strategic peers optimize by shuffle-k operations works well in environments with low fraction of

high capacity peers

Page 33: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

THE END

Page 34: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

34

Cathedral approach

most existing p2p streaming impose rules and incentives to motivate contribution

each peer is expected to follow – Cathedral approach

Page 35: Natural Selection in Peer-to-Peer Streaming: From the Cathedral to the Bazaar

35

Trace-based

Sigcomm: streaming Sigcomm conferences or workshops; most audience were interested in the contents but could not attain in person

Slashdot: a popular web-based discussion forum; audience is either interested in the contents or curious about the system

Gnutella: hosts in the Gnutella systemCHU ET. AL, “Early deployment experience with an overlay based internet broadcasting system,” USENIX Annual Technical Conference, June 2004