As an emerging area gaining prevalence in the industry, Web Services was established to satisfy the needs for better flexibility and higher reliability in web applications. However, due to the lack of reliable frameworks and difficulties in constructing versatile service composition platform, web developers encountered major obstacles in large-scale deployment of web services. Meanwhile, performance has been one of the major concerns and a largely unexplored area in Web Services research. There is high demand for researchers to conceive and develop feasible solutions to design, monitor, and deploy web service systems that can adapt to failures, especially performance failures. Though many techniques have been proposed to solve this problem, none of them offers a comprehensive solution to overcome the difficulties that challenge practitioners.

Central to the performance-engineering studies, performance analysis and performance adaptation are of paramount importance to the success of a software project. The industry learned through many hard lessons the significance of well-founded and well-executed performance engineering plans. An important fact is that it is too expensive to tackle performance evaluation, mostly through performance testing, after the software is developed. This is especially true in recent decades when software complexity has risen sharply. After the system is deployed, performance adaptation is essential to maintaining and improving software system reliability. Performance adaptation provides techniques to mitigate the consequence of performance failures and therefore is an important research issue. Performance adaptation is particularly meaningful for mission-critical software systems and software systems with inevitable frequent performance failures, such as Web Services.

This dissertation focuses on Web Services framework and proposes a performance-driven service composition scheme, called WS-Pro, to support both performance analysis and performance adaptation. A formalism of transformation from WS-BPEL to Petri net is first defined to enable the analysis of system properties and facilitate quality prediction. A state-transition based proof is presented to show that the transformed Petri net model correctly simulates the behavior of the WS-BPEL process. The generated Petri net model was augmented using performance data supplied by both historical data and runtime data. Results of executing the Petri nets suggest that optimal composition plans can be achieved based on the proposed method.

The performance of service composition procedure is an important research issue which has not been sufficiently treated by researchers. However, such an issue is critical for dynamic service composition, where re-planning must be done in a timely manner. In order to improve the performance of service composition procedure and enhance performance adaptation, this dissertation presents an algorithm to remove loops in the reachability graphs so that a large portion of the computation time of service composition can be moved to a pre-processing unit; hence the response time is shortened during runtime. We also extended the WS-Pro to the ubiquitous computing area to improve fault-tolerance.