marin silic, goran delac and sinisa srbljic prediction of atomic web services reliability based on...
TRANSCRIPT
Marin Silic, Goran Delac and Sinisa Srbljic
Prediction of Atomic Web Services Reliability Based on K-means Clustering
Consumer Computing Laboratory
Faculty of Electrical
Engineering and Computing
University of Zagreb, Croatia
http://ccl.fer.hr/
http://www.fer.hr/
http://www.unizg.hr/
ESEC/FSE, Saint Petersburg, Russia, 2013.
Outline
Motivation
Reliability in SOA
State-of-the-art
CLUS Approach
Evaluation
Conclusion
ESEC/FSE, Saint Petersburg, Russia, 2013.
Motivation
Contemporary web applications - SOA
ESEC/FSE, Saint Petersburg, Russia, 2013.
A 2
A 1A 3
A 4
A 5
Web Application
A 2
A 1
A 3
A 4
Composite service
Process of candidates selection
ESEC/FSE, Saint Petersburg, Russia, 2013.
A 2A 1
A 3A 4
A 5 A 6
Functional properties Nonfunctional properties
Ensure the desired functionality Reliability Availabilityβ¦
Impact Qos & QoE
Repository
A 2
A 1
A 3
A 4
A 5
A 6
Service Oriented System
βReliability on demandβ definition
ESEC/FSE, Saint Petersburg, Russia, 2013.
REQ
RES
π‘π πΈππ‘π πΈπ
π‘π πΈπβπ‘π πΈπ<β π‘
ππ=π π
ππ
The ratio of successful against
total number of invocations
Application
PastInvocation
Sample
Drawbacks/Obstacles
Clientβs vs. providerβs perspective Service invocation context Depends on the quality of the sample Acquiring a sample proves to be a difficult task
ESEC/FSE, Saint Petersburg, Russia, 2013.
A 2A 1 A 3
A 4A 5 A 6
QoS QoS QoS
QoS QoS QoS
A 1
QoS1
QoS2
QoS QoS1 QoS2β β
Service Provider
Client
Client
Insight to the Solution
To overcome the drawbacks and obstacles
Collect partial, but relevant past invocation sample
Utilize prediction methods to estimate the reliability for the missing records
ESEC/FSE, Saint Petersburg, Russia, 2013.
State-of-the-art
Collaborative filtering
ESEC/FSE, Saint Petersburg, Russia, 2013.
p1n?β¦p11 ? β¦ p1i ??β¦? p22 β¦ ? β¦β¦β¦β¦ β¦ β¦ β¦ pun?β¦pu1 ? β¦ pui β¦β¦β¦β¦ β¦ β¦ β¦ pmn?β¦pm1 ? β¦ pmi
m users
n services ??β¦?β¦?
ui matrix m,n >>matrix is extremely sparsenumber of values to predict
Collaborative filtering
Computes the similarity using PCC Matrix can be employed in two different ways
ESEC/FSE, Saint Petersburg, Russia, 2013.
p1n?β¦p11 ? β¦ p1i ??β¦? p22 β¦ ? β¦β¦β¦β¦ β¦ β¦ β¦ pun?β¦pu1 ? β¦ pui β¦β¦β¦β¦ β¦ β¦ β¦ pmn?β¦pm1 ? β¦ pmi
UPCC approach
IPCC approachHybrid approach
Disadvantages of Collaborative Filtering
Scalability
Having millions of users and services β these approaches do not scale
Accuracy in dynamic environments
Internet is a highly dynamic systemDo not consider environment conditions
ESEC/FSE, Saint Petersburg, Russia, 2013.
CLUStering
To address scalabilityApplies the principle of aggregationReduces the redundant data by clustering users
and services using K-means
To improve the accuracy Introduces environment-specific parametersDisperses the collected data across the
additional dimension
ESEC/FSE, Saint Petersburg, Russia, 2013.
CLUS Overview
ESEC/FSE, Saint Petersburg, Russia, 2013.
(1c)
(2c)
(5c)
Data Clustering Phase
r(u, s, t) p(r)
RawData
ClusteredData
Environment Clustering
UsersClustering
ServicesClustering
Creationof D
(3c)
(4c)
Prediction
Prediction Phase
Environment-specific Clustering
Set of environment conditions
ESEC/FSE, Saint Petersburg, Russia, 2013.
πΈ={π1 ,π2 ,β¦,ππ ,β¦,ππ}
t0 tct1 ti-1 ti tc-1w1 w2 wi wc
e1 e2 ei β¦β¦ en
β¦ β¦ππ€ π
=1
ΒΏπ πβ¨ΒΏ βπ βπ π
ππ ΒΏ
ππ€1ππ€2
ππ€ πππ€π
K-meansclusteringA day
User-specific Clustering
Set of users clusters
ESEC/FSE, Saint Petersburg, Russia, 2013.
π={π’1 ,π’2 ,β¦,π’π ,β¦,π’π}
u1 u2 ui β¦β¦ um
β¦β¦
ππ={ππ1 ,ππ2 ,β¦,πππ ,β¦,πππ }
e1 e2 ei β¦β¦ enππ1ππ2
ππππππ
ππ ππ ππππ
ππ ππ ππ ππ ππ
K-meansclustering
Service-specific Clustering
Set of services clusters
ESEC/FSE, Saint Petersburg, Russia, 2013.
π={π 1 ,π 2 ,β¦, π π ,β¦,π π }
s1 s2 si β¦β¦ sl
β¦β¦e1 e2 ei β¦β¦ enππ1ππ2
ππππππ K-meansclustering
ππ ππ ππππ
ππ ππ ππ ππ ππ
ππ={ππ1 ,ππ2 ,β¦,πππ ,β¦,πππ }
Creation of Space D Each record, r(u, s, t), is associated to the
belonging clusters uk , sj , ei Each entry in D is computed as follows:
R contains all the records that belong to clusters uk , sj , ei
ESEC/FSE, Saint Petersburg, Russia, 2013.
π· [π’π ,π π ,ππ ]=1
ΒΏπ β¨ΒΏ βπ βπ
ππ ΒΏ
Prediction
Assuming an ongoing rc=(uc, sc, tc) First, it checks the collected sample:
If H is not empty
Otherwise,
ESEC/FSE, Saint Petersburg, Russia, 2013.
π»={π hβ¨π’π=π’hβ π π=π hβ π‘π , π‘hβπ€π }
ππ=1
ΒΏπ»β¨ΒΏ βπ βπ»
ππ ΒΏ
ππ=π· [π’π , π π ,ππ ] ,π’πβπ’πβ§π πβπ πβ§π‘πβππ
Evaluation
Comparison with the state-of-the-artUPCC IPCCHybrid
Evaluation measuresPrediction accuracy
oMAE, RMSEPrediction performance
oAggregated prediction time
ESEC/FSE, Saint Petersburg, Russia, 2013.
Evaluation
Experiment setupAmazon EC2 Cloud
ESEC/FSE, Saint Petersburg, Russia, 2013.
Data
Evaluation
Results β Impact of data densityPrediction accuracy β with load intensity
ESEC/FSE, Saint Petersburg, Russia, 2013.
Evaluation
Results β Impact of data densityPrediction performance β with load intensity
ESEC/FSE, Saint Petersburg, Russia, 2013.
Evaluation
Results β Impact of number of clustersPrediction accuracy, Data density = 20%
ESEC/FSE, Saint Petersburg, Russia, 2013.
Evaluation
Results β Impact of number of clustersPrediction performance, Data density = 20%
ESEC/FSE, Saint Petersburg, Russia, 2013.
Conclusion
Proposed a CLUS approach Improved the prediction accuracy
By introducing environment-specific parameters At least 56% lower RMSE value than the state-of-the-art
Improved the prediction performance By applying principle of aggregation Execution time reduced for two orders of magnitude when
compared to the state-of-the-art Flexibility of approach
Trade-off between accuracy and scalability Can be applied in different environments
ESEC/FSE, Saint Petersburg, Russia, 2013.
Q&A
Thanks the audience for listening.
ESEC/FSE, Saint Petersburg, Russia, 2013.