Introduction to the experiment of service composition based on (mixed) Integer Programming (1)

Zeng04, "QoS-aware middleware for Web Services Composition"

Experiment settings
Hardware and basic software: a cluster of PC with the same configuration: Pentium III 933 MHz with 512 Ram, Windows 2000, Java 2 Enterprise Edition v1.3.0, Oracle XML developer kit, connected by LAN through 100 Mbits/sec Ethernet cards
Application Software: IBM's Web Services Toolkit (Develop Web Service), agflow, which uses IBM's optimization solutions and Library (OSL) as IP Solver
Scenario
A travel planning application
(1) Environment type
Static environment:
No change in the QoS of any component service during a given composite Service Execution
Dynamic Environment:
QoS of component services may undergo changes during the execution of a composite service: (a) existing services become unavailable (B) New services with better QoS appear (c) component Service not able to complete the task (d) Component Service may complete the task but without meeting the expected QoS
Simulated by randomly changing the QoS of the component services
Test Data
(1) composite service generation
"Randomly repeating exsiting states in the composite service shown in Fig. 3"
Number of States: 10, 20, 30 ,..., 80
Number of candidate component services per task: 10, 20, 30, 40
(2) QoS data collection
Execution duration, execution price: defined in the underlying service ontology (?)
Reliability, reputation: calculated by the service composition Manager (which logs QoS information during task executions)
Availability: calculated by the Service Broker (it records the up and down time of each Service)
Experiments
(1) Measuring computation cost
Compare the computing costs of three service selection methods (local optimization, global planning by exhaustive search, global planning by Integer Programming) in static and dynamic environments.
The purpose is to provide some evidence to determine which selection method should be used under different conditions.
(1.1) in the static environment. the global method only needs to be executed once, and the local method needs to be executed n times (n is the number of tasks ). experiment results: Global> global> Local
(1.2) in the dynamic environment, global executes each task and compares the computing cost of the global (IP) method in both dynamic and static environments.
(2) Measuring QoS of composite services
Composite service has 20 or 80 tasks and one execution path, in static and dynamic environment.
The result of the experiment is naturally that the global optimization method is better than the local optimization method.

Alrifai09, "combining global optimization with local selection for efficient QoS-aware service composition" Experiment settings

Hardware and basic software: HP ProLiant dl380 G3 machine with 2 Intel Xeon 2.80 GHz processors and 6 gb ram, Linux (centos Release 5), Java 1.6
APP: lpsolve Version 5.5
Scenario
(1) create several test cases of the QoS-based service composition problem, each test case: n service classes, l service candidates per class, M global QoS Constraints
(2) first use the global optimization method to obtain the optimal solution, then use the method proposed in this article to solve the same case, and compare the results.
(3) number of quality levels: 10, 20, 30, 40, 50
Test Data
(1) first dataset: qws real dataset (this is a Web Service set, but there is no combined service set. I don't know how the author uses it in this experiment)
(2) second dataset: 20000 artificial web services by assigning arbitrary QoS values (normally distributed in the range between 1 and 100)
Experiments
Experiments (1) and (2) Compared the computation cost of the global and hybrid methods (10, 20, 30, 40, 50). The results show that the latter is better than the former.
Experiments (3) and (4) compare the effectiveness of the two methods. The results show that the solution obtained by the method in this article can obtain the approximate global method.
(1) Performance Comparison W. R. T. The number of Web Service candidates
Experiment parameters:

	Qws	Random Dataset
Number of Service Classes	5	10
Candidate per class	50,100 ,..., 500	100,200 ,..., 2000

(2) Performance Comparison W. R. T. The number of Web Service Classes
Experiment parameters:

	Qws	Random Dataset
Number of Service Classes	5, 10 ,..., 25	10, 20 ,..., 100
Candidate per class	100	500

(3) optimality W. R. T. The number of Web Service candidates
(4) optimality W. R. T. The number of Web Service Classes