The Query-Cycle Simulator for Simulating P2P Networks

Mario T. Schlosser

Tyson E. Condie

Sepandar D. Kamvar

Stanford University

2

Problem: Accurately Simulate

Real-World P2P Networks.

Motivation: Testing P2P

Algorithms.

Problem

For each peer i {

-Repeat until convergence {

-Compute. . .

-Send . . .

}

}

3

Goals P2P Simulator

Descriptive Simple Easily Extensible Make it available on the web so that people

can test and compare their algorithms on a standard platform.

4

Query Cycle Model

Query Cycle 1

5

Query Cycle Model

Query Cycle 2

6

Query Cycle Model

Query Cycle 3

7

Properties to Model Peer Content Network Parameters Peer Behavior

8

Properties to Model Peer Content

How Much? What Type?

Network Parameters Peer Behavior

9

Data Volume Observations

Model

Simulator assigns # of files owned by peer i according to

distribution.

Saroiu,Gummandi,and Gribble. A Measurement Study of Peer-to-Peer File Sharing Systems, 2002.

10

Content Type: Observations Content Categories

Zipf distribution on file popularityCrespo and Garcia-Molina. Semantic Overlay Networks, 2002.

Korfhage, Information Storage and Retrieval, 1997.

Punk Rock Hip-

HopJazz

0

0.2

0.4

0.6

0.8

1

1.2

1 2 3 4 5

Files

Po

pu

lari

ty

11

Content Type: Model Modeling Content Categories:

Assume n content categories. C={c1,c2,…,cn} A peer i is assigned content categories according to the Zipf

distribution:

It is then assigned an interest level p(c|i) to each of the assigned content categories by a uniform random distribution.

n

i

i

ccp

1

/1

/1)(

12

Content Type: Model Modeling Files:

Each distinct file f may be uniquely identified by {c,r} A peer is assigned files by:

cF

i

rc

i

rcfp

1

,

/1

/1)|(

)|()|()|( ,, cfpicpifp rcrc

13

Recap on Content Assignment

14

Recap on Content Assignment

Assign Data Volume

15

Recap on Content Assignment

{c1, c3, c4}

Assign Content Categories

16

Recap on Content Assignment

{c1=.5, c3=.3, c4=.2}

Assign Interest Level to Content Categories

17

Recap on Content Assignment

{c1=.5, c3=.3, c4=.2}

Assign Files

{c,r}={c1,f1} {c,r}={c1,f7} . . .

18

Properties to Model Peer Content Network Parameters

Topology Bandwidth

Peer Behavior

19

Network Parameters Topology:

Observation: Power Law Topology Model: probability of connecting to a peer is

proportional to the degree of that peer. Bandwidth

Simple Bandwidth Model Can be easily extended.

20

Properties to Model Peer Content Network Parameters Peer Behavior

21

Query-Cycle Model At each cycle, peer i may be:

active inactive or down

22

At each cycle, peer i may be: active inactive or down

Query-Cycle Model

Issues a single query.

Waits for incoming responses.

Selects a source and downloads file.

Also:

Responds to queries.

Forwards query messages.

23

At each cycle, peer i may be: active inactive or down

Query-Cycle Model

Responds to queries.

Forwards Query Messages.

24

At each cycle, peer i may be: active inactive or down

Query-Cycle Model

Does nothing.

25

Properties to Model Peer Content Network Parameters Peer Behavior

Uptime and Session Duration Query Activity Queries Query Responses Downloads

26

Uptime Observations

ModelAt each query cycle, probability of being up is drawn from distribution in Saroiu et al.

Saroiu,Gummandi,and Gribble. A Measurement Study of Peer-to-Peer File Sharing Systems, 2002.

27

Queries Observations

None

Model Based on the idea that peers query for files in

the same categories that they own.

)|()|()|( ,, cfpicpiqp rcrc

cF

i

rc

i

rcfp

1

,

/1

/1)|(

28

Responses and Downloads Responses

If a peer receives a query for which it owns the file, it responds.

Source Selection Random

29

Extensions Different Types of Peers

i.e., Malicious Peers Different Models for Different Situations

Reputation-based source selection. Edutella: model distribution over markups

rather than content categories. Web Services: Change models for content

distribution, query activity, etc. However, parameters are the same.

30

Samples

31

Future Work Test predictions against observations in

P2P networks “in the wild”. Observations, observations,

observations. Model other networks.

32

The End Code, demos will be available at

http://www.stanford.edu/~sdkamvar/research.html next monday.

33

Motivation

Network or peer property Affected algorithms

TopologyContent distribution

Bandwidth, uptime of peers

Structuring algorithmsWhatever

Stability of trust algorithms

34

Query Activity Observations

ModelAt each query cycle, . . .

Saroiu,Gummandi,and Gribble. A Measurement Study of Peer-to-Peer File Sharing Systems, 2002.