a link-based ranking model for servicesle2i.cnrs.fr › img › publications ›...

A Link-Based Ranking Model for Services

Camelia Constantin1 Bernd Amann1 David Gross-Amblard2

1LIP6Universite Pierre et Marie Curie Paris

2Laboratoire CEDRICC.N.A.M. Paris

Cooperative Information Systems(CoopIS 2006)

C.Constantin, B.Amann, D.Gross-Amblard (Universit e Paris VI et C.N.A.M)A Link-Based Ranking Model for Services CoopIS 2006 1 / 21

IntroductionModelImportance ComputationConclusion

ContextMotivating ExampleRelated WorkApproach

Distributed Applications

Observation

Service-oriented computing and P2P architectures are more andmore popular for building web applications

Many web applications are based on distributed data andprocessing

Central Issue

Controlling (estimating and improving) the quality of manydistributed applications

Our Approach

Rank services according to the estimated influence of their qualityon the global quality of applications

Outline

Introduction

Quality ContributionService UsageEffective Contribution

Importance Computation

Service ImportanceExperiments

Conclusion and Future Work

Example: Data-Intensive Services

Web Service:Parameterized Query (SQL, XQuery) over local and remote data sending aset of data as response

Important services : those which contribute the most

Example: Ranking of ServicesRanking Model

S6 > S4 > S3 >

S5 >S2=S1

Applications:

Search engines (services, data)

Service configuration → replicate S6

Pricing: should be more expensive thanC.Constantin, B.Amann, D.Gross-Amblard (Universit e Paris VI et C.N.A.M)A Link-Based Ranking Model for Services CoopIS 2006 5 / 21

State of the artRanking of Web Services

Sampling-based ranking of data-intensive services with respect to aspecific client [Caverlee et al, ’04]

Recommendation-based techniques (user feedback and voting services)[Abbadi et al, ’04], [Kalepu et al, ’04]

Conformance of Services’QoS published features during given periods[Aberer et al, ’05]

−→ Do not consider the collaboration links between services

Link-based Ranking of Web Pages

Consider the web pages and the links between them as a graph (eg.Google’s PageRank[Page et al, 98], HITS [Kleinberg, 99])

Adapting PageRank for Ranking Services

Web Page graph −→ Service graph:

Web page −→ service web and hyperlink −→ service call?

Some important differences:

Service is generally a ”black-box” for the clientsService does a computation on behalf of its caller, returns a resultMore complex quality criteria for services (eg. response time)Hyperlinks=static follow-up links, calls=dynamic usage of services

Our model

Considers both static (quality contribution) and dynamic (serviceusage) links between services

Our model

How to compute the contribution scores for each service?

Quality Contribution Score

The quality of Si depends on the quality of all services Sj that it uses

Semantic and static measure, similar to hyperlinks

Computed independently on the calls between services

Estimated locally by si for all services sj that it calls

Score cji ∈ [0, 1] comparing the contribution of all services Sj to Si

cji = 1

Quality Contribution Score

The quality of Si depends on the quality of all services Sj that it uses

Semantic and static measure, similar to hyperlinks

Computed independently on the calls between services

Estimated locally by si for all services sj that it calls

Score cji ∈ [0, 1] comparing the contribution of all services Sj to Si

cji = 1

Example: Static Query Analysis

S3 estimates the static contribution of S4, S5, S6 to its quality

−→ S4 contributes the most, S5 contributes the least

Example: Static Query Analysis

S3 estimates the static contribution of S4, S5, S6 to its quality

−→ S4 contributes the most, S5 contributes the least

Service Usage Score

Si uses Sj

Dynamic measure expressing the way in which Sj is used by Si

Usage function uji(t) ∈ [0, 1] computed locally by Si for each Sj

Based on information registered in the log (Logi ) of Si during aperiod

Example

Age/freshness of the results of Sj

The part of the results of Sj missed by Si

The percentage of the results of Sj that are useful to Si

Example: Percentage of Useful Results

S3 computes the service usage of S4, S5, S6 by observing the usageof their results

−→ S5 is not used, S6 is more used than S4

Effective Contribution ScoreServices effectively contribute to the quality of others only if they are called

−→ Static contribution of S5 is smaller than the one of S6

Effective Contribution ScoreServices effectively contribute to the quality of others only if they are called

−→ Service usage of S5 is greater than the one of S6

−→ Service usage of S5 is 0

Effective Contribution ScoreEffective contribution=quality contribution * service usage(ceff

ji (t) = cji ∗ uji(t))

0.5*(50%)

0.1*0%

0.4*(100%)

−→ effective contribution of s5 is greater than the one of s6

The contribution of sj to si is maximized when service usage is 1

Summary

Contribution scores computed by Si for services Sj that it uses:

Score Description Si → Sj Sk → Si → Sj

(See Paper)

static distribution cji —contributionservice usage weight uji(t) ∈ [0, 1] ujik(t) ∈ [0, 1]

effective weighted cji ∗ uji(t) cji ∗ ujik(t)contribution distribution

Contribution links have a static and a dynamic part

Service ImportanceExperimets

Total contribution

effective contribution

S3 has a local knowledge on the effective contribution of S4,S5,S6

Total contribution

effective contributioneffective contribution

S1 and S2 have local knowledge on the effective contribution of S3

Total contribution

effective contributioneffective contribution

Ex: How to compute the effective contribution of S4 to all others?

Importance

Importance of a Service

Total effective contribution to the quality of other services

Importance Computation

Static contributions −→ C (Contribution matrix)

Log+C −→service usage E (Effective contribution matrix)

The importance is the solution of a linear system with matrix E

AlgorithmsCompute the importance by iterative algorithms:

E →Algorithms Service Importance

Computation stops at convergence

Algorithms

Centralized Iterative Algorithm

Single service that gets the contribution values from all others

Builds matrix E and computes the importance

→ Computing service must get possibly secret information from others

Distributed Iterative Algorithms

Each service computes its importance by exchanging messages

Synchronous: all services compute the same iterationAsynchronous:

Computation messages are sent asynchronously, at the same timeas the call parameters during normal service callsEach service computes different iterations and updates theirimportance at each call

Number of Messages for Different Graph ModelsGraph configurations generated by four methods:

Max, Linear-Copying, Small-World, Client-Server.

Number of Iterations for Different Usage Functions

Converged Services w.r.t. Number of Iterations

(a) synchronous algorithm (b) asynchronous algorithm

Contributions:

Link-based ranking model for servicesCan be applied to different quality measuresExpresses the quality contribution and usage between services

Importance Computation:Distributed computation, each service computes its own importanceTwo algorithms: a synchronous and an asynchronous oneSimulation of importance computation on four service graphs

Future Work:

Ongoing implementation on ActiveXML (WebContent, lot 5 - P2P).Apply importance scores for P2P optimisation:

Optimize the global response time of distributed queriesP2P data refresh strategies

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Contributions:

Future Work:

Thank you for your attention

a link-based ranking model for servicesle2i.cnrs.fr › img › publications ›...

Documents