Exploiting Content Localities for Efficient Search in P2P Systems

Lei Guo1 Song Jiang2 Li Xiao3 andXiaodong Zhang1

1College of William and Mary, USA2Los Alamos National Laboratory, USA

3Michigan State University, USA

Peer-to-Peer Search

• Two Performance Objectives – Individual peer: improve the search quality– Internet management: minimize the search cost

Fast, fast, fast, andthe more the better!

P2P user

Don’t be so greedy, the Internet is shared by all the people!

Network manager

Existing Solutions

• Generally aim to one of the two objectives and have performance limits to the other

• Flooding:– Most effective for user’s experience– Least efficient for network resource utilization

• Random walk:– Traffic efficient, but– Long response time and limited number of search

results

Super-Node Architecture

• Super-node– Index server for its leaf nodes

• Problems– Index based search has limits

• Hard for full-text search• Impossible for encrypted content search

– Not responsible for the content quality of its leaf nodes – The structure becomes large and inefficient.

• A leaf node has to connect to multiple super-nodes to avoid single point failure

• Generating an increasingly large number of super-nodes

Gnutella Population in One Day (2003)

number of peers

number of super peers

One super node only connects to 3-4 peers in average!

Outline

• Our Measurement Study

• CAC: Constructing Content Abundant Cluster

• SPIRP: Selectively Prefetching Indices from Responding Peers

• CAC-SPIRP: Combining CAC and SPIRP

• Performance Evaluation

• Conclusion

Our Measurement Study

• Existing measurement studies– A small percentage of popular files account for most

shared storage and transmissions in P2P systems– A small amount of peers contribute majority number

of files in P2P. – They are only the indirect evidence of content locality

• Some files may be never accessed, or accessed rarely

• Our purpose– Fully understand the localities in the peer community

and individual peers– Get first-hand traces for our simulation study

Trace Collection

• Four-day crawling on the Gnutella network– Open source code of LimeWire Gnutella– Session based collection (for the whole life time of

peers)

• Query sending traces by different peers– 25,764 peers– 409,129 queries

• Content indices of different peers– Full indices of 18,255 peers– 37% free riders

Top Content Providers (in percentage)Que

ries

Rep

lied

by T

op Q

uery

Res

pond

ers

(%)

Res

ults

Rep

lied

by T

op R

esul

t Pro

vide

rs (

%)

100

80

60

40

20

0

100

80

60

40

20

00 20 40 60 80 100

Content Locality in the Peer Community

A small group of peers can reply nearly all queries and provide most of resultsA small group of peers can reply nearly all queries and provide most of results

Number of Queries100 101 102 103 104

Pe

rce

nta

ge

of

Pe

ers

(%

)100

80

60

40

20

0

Pe

rce

nta

ge

of

Pe

ers

(%

)

Number of Results100 102 104 106

100

80

60

40

20

0

The Localities of Search Interests of Individual Peers

• A peer can get search results from a small number of its top query responders: they share the same search interests

• Similar to the idea in Locality of Interest scheme, but our conclusion is based on real P2P systems

Top Query Responders Top Result Providerstop 1 top 10 top 5% top 10% top 20% top 1 top 10 top 5% top 10% top 20%

Que

ry C

ontr

ibut

ions

(%

)

Res

ult

Con

trib

utio

ns (

%)

100

80

60

40

20

0

50

40

30

20

10

0

60

Reorganizing the P2P Management Structure

• Clustering those small number of content abundant peers

• Prefetching indices from those top query responders

CAC: Constructing Content Abundant Cluster

• Objectives– Clustering those small number of content abundant

peers in P2P overlay– Providing high quality and fast service

• Content Abundant Cluster– An overlay on top of P2P network– Self-evaluate, self-identify, and self-organize– Persistent public service for all peers in the system– Strong content-based (not index-based)

ClusteringLeveling

CAC: System Structure

C A CC A C

0

00 0

0

0 0

11

1

1 1

1

1

1

22

22

2

2

2

2

2

2 2

2

3

3

3

3

3

3

4

X

3

3

2

Dynamic Update

CAC: Search Operations

• Queries are sent to CAC first– Up-flowing operation– Flooding in CAC

• Unsatisfied queries are propagated from CAC to the whole system– Down-flooding operation– Propagated from low levels to high levels

Up-flowing

C A CC A C

0

00 0

0

0 0

11

1

1 1

1

1

1

22

22

2

2

2

2

2

2 2

2

3

3

3

3

3

3

4

Down-flooding

C A CC A C

0

00 0

0

0 0

11

1

1 1

1

1

1

22

22

2

2

2

2

2

2 2

2

3

3

3

3

3

3

4

Unused links

SPIRP: Selectively Prefetching Indices from Responding Peers

• Basic operations– Peer I initiates a query q

• Query hits: displays the results• Misses: sends q

– Peer R responds query q• sends query results as well as• piggybacks indices of all shared files

– Peer I receives response• Display the searching results as well as • stores piggybacked indices

• Indices updating– Active updating indices by responding peers– Updating indices demanded by requesting peers

• Replacement of file indices

Where are these files?

Pop music

Classic music

SPIRP Technique

♫♫

R1

R2

Query = “Beethoven mp3”

I

SPIRP Technique

pop

classic

NULLNULL

R1

R2

Query = “Beetle mp3”

Where are these files?I

SPIRP Techniqueclassic

pop

R1

R2

Query = “Beetle mp3”

I

SPIRP Techniqueclassic

pop

R1

R2

Query = “Beetle mp3”

No enough space tosave indicesI

SPIRP Techniqueclassic

pop

♫♫

R1

R2

Replace completeI

Query = “Beetle mp3”

CAC-SPIRP

• CAC: application level infrastructure– Significantly reducing bandwidth consumption– Good response time when queries success in CAC– Long response time when queries fail in CAC

• SPIRP: client-oriented and overlay independent– Significantly reducing response time– Small traffic when queries can be satisfied in cache– Same traffic as flooding when cache misses

• CAC-SPIRP– Easy to combine the two techniques– Consider the trade-off between the two performance objectives– Has both merits of search quality and search cost

Simulation Environment

• Content trace and query trace– 4 day Gnutella crawling in our measurement

• Overlay topology– Traces by Clip2 Distributed Search Solutions

• Session duration– Pareto distribution fitted from measurement

results

P(x) = 14.5311 * x -1.8598

Evaluation Metrics

• Query success rate– CAC: success rate in CAC (normalized to flooding)– SPIRP: success rate in local cache (normalized to flooding)

• Overall network traffic– accumulated communication traffics for all queries, responses,

and index transferring (normalized to flooding)

• Average response time– use the number of routing hops (normalized to flooding)

Evaluate for different query satisfactions– 1, 10, 50 results, representing different user demands

Performance Evaluation for CAC

0 10 20 30 40 50Cluster Size (In Percentage of P2P Network Size)

5% top content abundant peers are good enough for cluster construction

Ove

rall

Tra

ffic

(Nor

mal

ized

)

1

0.8

0.6

0.4

0.2

0

0 10 20 30 40 50Cluster Size (In Percentage of P2P Network Size)

0 10 20 30 40 50Cluster Size (In Percentage of P2P Network Size)

Su

cce

ss R

ate

in C

AC

(n

orm

aliz

ed

)

1

0.8

0.6

0.4

0.2

0

Avg

Re

spo

nse

Tim

e

(No

rma

lize

d)

2

1.5

1

0.5

0

Minimum Results = 1Minimum Results = 10Minimum Results = 50

Minimum Results = 1Minimum Results = 10Minimum Results = 50

Minimum Results = 1Minimum Results = 10Minimum Results = 50

0 10 20 30 40 50Cluster Size (In Percentage of P2P Network Size)

CAC Member Selection

0 0.01 0.02 0.03 0.04Success Response Rate of Content-Abundant Peers

Suc

cess

Rat

e in

CA

C (

norm

aliz

ed) 1

0.8

0.6

0.4

0.2

0

Minimum Results = 1Minimum Results = 10Minimum Results = 50

Avg

Re

spo

nse

Tim

e (

No

rma

lize

d)

Ove

rall

Tra

ffic

(N

orm

aliz

ed

)

0 0.01 0.02 0.03 0.04Success response rate of CAC Peers

1

0.8

0.6

0.4

0.2

0

0 0.01 0.02 0.03 0.04Success Response Rate of CAC Peers

2

1.5

1

0.5

0

Minimum Results = 1Minimum Results = 10Minimum Results = 50

Minimum Results = 1Minimum Results = 10Minimum Results = 50

• Overall traffic is not sensitive to CAC member quality• Traffic can be significantly reduced even for

randomly selected CAC members CAC down flooding is very efficient

CAC-SPIRP Overall PerformancePeers having 1 to 5 queries satisfiedPeers having 10 to 20 queries satisfiedPeers having 30 to 40 queries satisfiedPeers having at least 50 queries satisfied

Peers having 1 to 5 queries satisfiedPeers having 10 to 20 queries satisfiedPeers having 30 to 40 queries satisfiedPeers having at least 50 queries satisfied

Query Satisfaction = 1Query Satisfaction = 10Query Satisfaction = 50

0 2 4 6 8 10Size of Incoming Index Set Buffer (in M Bytes)

Ave

rage

Res

pons

e T

ime

(Nor

mal

ized

)

2

1.6

1.2

0.8

0.4

0

Su

cce

ss R

ate

in L

oca

l Ca

che

1

0.8

0.6

0.4

0.2

Ove

rall

Tra

ffic

(N

orm

aliz

ed

)

1

0.8

0.6

0.4

0.2

0 2 4 6 8 10Size of Incoming Index Set Buffer (in M Bytes)

0

0 2 4 6 8 10Size of Incoming Index Set Buffer (in M Bytes)

0

CAC-SPIRP reduces both the overall trafficand response time significantly

Conclusion

• CAC-SPIRP fundamentally addresses the P2P search problem by a re-organization.– Exploiting organizational content locality

• CAC: a content abundant cluster provides high quality and fast services.

– Exploiting user content locality• SPIRP: a client prefetching technique to speed up

search by avoiding unnecessary queries