230 likes | 237 Views
CSSSIA Workshop – WWW 2008 Speeding up Web Service Composition with Volatile External Information. John Harney, Prashant Doshi LSDIS Lab, Dept. of Computer Science, University of Georgia Presented By: Haibo Zhao LSDIS Lab, Dept. of Computer Science, University of Georgia.
E N D
CSSSIA Workshop – WWW 2008Speeding up Web Service Composition with Volatile External Information John Harney, Prashant Doshi LSDIS Lab, Dept. of Computer Science, University of Georgia Presented By: Haibo Zhao LSDIS Lab, Dept. of Computer Science, University of Georgia
Adaptation of WS Compositions • Many Web service composition (WSC) techniques assume static environments • Service parameters are assumed fixed • Environments are often dynamic • Examples -- Trip planner • Price of an airline ticket offered by a service (such as Delta’s online booking) increases • Availabilities of rooms at a specific hotel may increase or decrease • Service response time increases due to network outage
Optimal WSC • Optimality – Selecting services so as to minimize cost of trip and optimize other parameters • Underlying objective – maintain WSC optimality • Depends on how accurately the services’ parameters are captured • Requires cognizance of changes in parameter values that may have occurred • Backtracking to previous points in the composition and recomposition may be needed
Motivating Scenario – Travel Planning Backtrack prior to selection of Airline Ticket Service Delta ticket price increases Start Delta Ticketing Service Rental car Service Hotel Service Finish Ticket price expires United Ticketing Service United ticket becomes cheapest available
Greedy Composition • Composition method selects the best service amax at every step (state) without look ahead a1(s0) a1(s1) a1(sm) amax Finish Start Start Start Start Start Start a2(s0) a2(s0) a2(s0) a2(s1) a2(sm) amax … … … amax an(s0) an(s1) an(sm)
Greedy Composition Problem formalization Availability of services Initial State Set of all candidate WSs Set of all states Set of possible WSs that can be invoked at a state Cost of services Goal state
Greedy Composition • A service a is chosen based on its value: Normalized Cost Normalized Availability Importance of Cost Importance of Availability Service selected is the one that yields the greatest value. Policy for choosing the best service at each state
Reactive Query Policies • Introduced by Au and Nau (2006) • Reactive Query Policy • Set of rules that decide when expired parameters of services should be queried and revised values integrated during the composition • Three policies were introduced • Eager • Lazy • Presumptive
Reactive Query Policies • Eager • When a service expires, query that service and halt composition until answer arrives; if needed, backtrack and recompose Query issued Query response received Airline Ticket Service Rental Car Service Hotel Service start HALTS end Ticket Expires Backtrack if needed
Reactive Query Policies • Lazy • When a service expires, query that service and continue composition until goal is reached; if needed, backtrack and recompose Query issued Query response received Airline Ticket Service Rental Car Service Hotel Service start end CONTINUE Ticket Expires Backtrack if needed
Reactive Query Policies • Presumptive • When a service expires, query that service and continue composition until the answer to the query is received; if needed, backtrack and recompose Query issued Query response received Airline Ticket Service Rental Car Service Hotel Service start end CONTINUE Ticket Expires Backtrack if needed
Reactive Query Policies • All previous policies suffer from: • Excessive backtracking (some backtracking is not required) • Checks if the value of the service parameter changes rather than if the policy changes Not all changes in the values of the parameters cause changes in the composition!
Our Approach: Informed-Presumptive • Improves the presumptive approach • Identifies regions of revised parameter values for which the optimal policy does not change • Uses gradient descent to identify the regions Theorem For a WSC, there exists at least one revised parameter vector for which the value of using the current policy with is the same as the value of using the optimal policy with . Typically, there are more parameter vectors for which the above is true!
Informed-Presumptive Gradient Descent • How is the descent performed? • Plot of Cost vs Availability vs Value Difference Value Difference Form a line Cost Start with 2 distinct points on the surface Use gradient to descend down surface Availability Descend until Value Difference is zero
Informed-Presumptive Presumptive vs Informed-Presumptive Informed-presumptive does not backtrack if the revised values fall in region that policy is unchanged Value Difference Presumptive does not backtrack if the revised values are unchanged Cost Value using optimal policy Availability Value using current policy
Informed-Presumptive Gradient Descent Computationally inexpensive to find the region Theorem Given that Va is a linear function of the parameters of a service a, the gradient : is constant.
Empirical Results Experimental Setup Change Likelihood (probability that upon expiration, a service parameter will change) Compared Informed-Presumptive against the previous existing strategies (Lazy, Eager, Presumptive) for the following: Time required to complete the WSC Number of backtracks in the composition Number of compositions that were not completed
Empirical Results • Average Composition Times As likelihood of change increases, the informed-presumptive out performs others by a wide margin For lower change likelihoods, presumptive may outperform informed-presumptive Reason: overstepping in gradient descent leads to slight error in division line
Empirical Results • Average Number of Backtracks As likelihood of change increases, the informed-presumptive backtracksless than the others Note: Lazy approach performs limited backtracking, but the backtracks are more costly than the informed presumptive
Empirical Results • Average Number of Incomplete Compositions As likelihood of change increases, the informed-presumptive has less “incomplete” compositions
Conclusion Many Web service compositions must function in dynamic environments Backtracking may be required Previous composition strategies perform redundant backtracks Our Informed-Presumptive strategy: eliminates unnecessary backtracks implemented without much extra overhead outperforms previous strategies
Future Work Compositions with nonlinear combinations of parameters Challenges: Dependencies between services Gradient descent could be more expensive computationally
Thank you Questions