22
Toward Optimal and Efficient Self-Adaptation in Large Web Processes Prashant Doshi Assistant Professor LSDIS Lab, Dept of Computer Science, University of Georgia Joint work with: Kunal Verma, Yunzhou Wu, and Amit Sheth

Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Embed Size (px)

DESCRIPTION

Toward Optimal and Efficient Self-Adaptation in Large Web Processes. Prashant Doshi Assistant Professor LSDIS Lab, Dept of Computer Science, University of Georgia Joint work with: Kunal Verma, Yunzhou Wu, and Amit Sheth. Outline of the Talk. Understanding Volatility Two characterizations - PowerPoint PPT Presentation

Citation preview

Page 1: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Prashant DoshiAssistant Professor

LSDIS Lab, Dept of Computer Science, University of Georgia

Joint work with:

Kunal Verma, Yunzhou Wu, and Amit Sheth

Page 2: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Outline of the Talk• Understanding Volatility

– Two characterizations

• Our Approach– Abstract Processes and Service Managers– Adaptation as a Decision-Making Problem

• A Framework for Studying Adaptation– Evaluation criteria

• Optimality• Computational Efficiency

• Some Experimental Results

• Value of Changed Information– Definition– Experimental Results

• Discussion and Future Work

Page 3: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Understanding Volatility• Data Volatility

– Atypical input and execution data• E.g.. delay in satisfying order

adverse drug reaction– New knowledge

• E.g.. New drug alert

Component Volatility– Change in the state of the process

participants• E.g.. Web service failure or abnormal behavior

• Expected Volatility– Events known to occur with some chance

• E.g.. delay in satisfying order Worsening of patient symptoms

Unexpected Volatility– E.g.. New drug alert

New co-morbidity

data volatility

component volatility

expected(with some chance)

unexpected

Page 4: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Abstract Processes and Service Managers• Pre-specified abstract processes

– Ordering of activities– Inter-activity constraints: E.g. Coordination constraints

• Process and Service Managers

Heart FailureClinical Pathway

Page 5: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Abstract Processes and Service Managers• Our architecture

– Two tiers• Resources Layer• Control Layer

Page 6: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

A Framework for Studying Adaptation• Two criteria for evaluating approaches

– Cost-based optimality

– Computational efficiency

• Formalize adaptation as a decision problem– Two general choices

• Ignore the change• React to the change

– Example methodology: Markov decision processes (MDP)

Decreasing Computational EfficiencyDecreasing Optimality

Centralized Adaptation

DecentralizedAdaptationHybrid approaches

Page 7: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

A Framework for Studying Adaptation• Centralized Approaches

– PM is responsible for adaptation• Global oversight

• Decentralized Approaches– SMs are responsible for local adaptation

• Local oversight

• Difficult to manage inter-activity constraints

• Hybrid Approaches– Both PM and SMs share the responsibility of adaptation

• Global and local oversight

Page 8: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Establishing the Ends of the Spectrum• Centralized adaptation to

expected data volatility• Example: M-MDP method (Verma, Doshi et al. ICWS 06)

Properties:

Theorem: M-MDP adapts the process optimally

to exogenous events expected with some chance

and with coordination constraints

• PM has global oversight and controls the SMs• Does not scale well: Complexity exponential in the number of SMs

Computer assembly

Page 9: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Establishing the Ends of the Spectrum• Decentralized adaptation to

expected data volatility• Example: MDP-CoM method (Verma, Doshi et al. ICWS 06)

• Challenge: Satisfying

coordination constraints

Properties:• Scalable to multiple SMs• Not optimal

Computer assembly

Coordination Mechanism

Page 10: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Research Challenge: Hybrid Approaches• Idea #1: Least-commitment

– PM steps in only when needed• E.g. when deciding on a coordinating action

• Idea #2: Inter-SM communication– Motivation for communication: Regret

Page 11: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Penalty of delay = $400

900

1300

1700

2100

2500

0.1 0.2 0.3 0.4 0.5 0.6 0.7

Probability of delay

Av

era

ge

Co

st(

$) M-MDP

Random

Hyb. MDP

MDP-CoM

Penalty of delay = $200

900

1300

1700

2100

2500

0.1 0.2 0.3 0.4 0.5 0.6 0.7

Probability of delayA

vera

ge

Co

st($

)

M-MDP

Random

Hyb. MDP

MDP-CoM

Some Experimental ResultsAdapting to delay in supply chain• Choices

•Wait out the delay•Change the supplier

M-MDP incurs the least average costMDP-CoM the most

Runtime for MDP-CoM remains fixedas number of activities increases•Decentralized adaptation is parallelizable

Page 12: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Related Work• Verification of correctness of manual changes to control flow

– Adept (Reichert&Dadam98), Workflow inheritance (Aalst&Basten02), inter-task dependencies (Attie et al.93)

• Event Condition Action (ECA) rules for adaptation– Agentwork (Muller et al.04)

• Change of service providers based on migration rules in E-Flow (Casati et al.00)

• We complement previous work in this area by emphasizing:– Cost based optimality – Computational efficiency

Page 13: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Unexpected Data Volatility• Example

– Rate of order satisfaction may change arbitrarily– Cost of service may change arbitrarily

• Research Challenges1. How to be cognizant of the change

2. When to adapt to the change

• Our approach– Query the service providers for revised information

• Cost of querying!

– Adapt when information is useful

Page 14: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Possible Approaches• Query a random provider for relevant information

– Advantages• Up-to-date knowledge of queried service provider• Performs no worse than “do nothing” strategy

– Disadvantages• Querying for information not free • Paying for information that may not be useful

– Information may not change Web process

• Value of Changed Information (VOC) (Harney&Doshi,ICSOC06)– Decides if obtaining information is expected to be:

• Useful– Will it induce a change in optimality of Web process?

• Cost-efficient– Is the information worth the cost of obtaining it?

• Extension of VOI (Value of Information)

Page 15: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Value of Changed Information• VOC

– Measures how “badly” the current process is expected to perform in changed environment

– Defined as the difference between:• Expected performance of the old process in the changed environment• Expected performance of the best process in the changed environment

• Formalizing VOC– Actual service parameters are not known

• Must average over all possible revised parameters

– We use a belief of revised values • Could be learned over time

Page 16: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Manufacturer’s Beliefs For Supply Chain

Example - Beliefs of Order Satisfaction

Page 17: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Adaptive Web Process Composition

…Prov 1 Prov 2 Prov n

VOC VOC VOC

Keep current process

Query Provider Re-compute process if

needed

1. SM calculates VOC for each service provider involved in Web process

2. PM finds provider whose changed parameter induces the greatest change in process (VOC*)

3. Compare VOC* to cost of querying

VOC* < Cost of Querying

VOC* > Cost of Querying

*

Page 18: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Empirical ResultsMeasured the average process cost over a range of query cost values

– Query random strategy cost grows at a larger rate than VOC– VOC queries selectively– VOC performs no worse than the do nothing strategy

Supply Chain Web Process Patient Transfer Web Process

Page 19: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Discussion• Understanding dynamic environments is crucial

– Categorizations needed• Data and component volatility

• Expected (with probabilities known a’priori) and unexpected events

• Other taxonomies?

• A framework for studying adaptation– Criteria for evaluation

• Cost-based optimality

• Computational efficiency

– We established the ends of the spectrum• Centralized (M-MDP) and decentralized approaches (MDP-CoM)

• Research on hybrid approaches needed

Page 20: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Discussion

• Value of changed information (VOC)– Unexpected and arbitrary data volatility– Query for revised information

• Obtains revised information expected to be useful

• Avoids unnecessary queries

• VOC calculations are computationally expensive– Knowledge of service parameter guarantees may be used to

eliminate unnecessary VOC calculations (Harney&Doshi, WWW 07)

– Other approaches needed

Page 21: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Future Work

• More study and characterization of volatility

• Research on hybrid approaches

• Handle component volatility– Candidate approaches: A-WSCE architecture (Chafle et

al.06)

– k-service redundancy and k-process redundancy

Page 22: Toward Optimal and Efficient Self-Adaptation in Large Web Processes

Thank You

Questions