A Framework for Quality of Service and Fault Management in Service-Oriented Architecture

Service-Oriented Architecture (SOA) provides a powerful yet flexible paradigm for integrating distributed services into business processes to perform complex functionalities. However, the flexibility and environmental uncertainties bring difficulties to system performance management. In this dissertation, a quality of service (QoS) management framework is designed and implemented to support reliable service delivery in SOA. The QoS management framework covers runtime process performance monitoring, faulty service diagnosis and process recovery. During runtime, the QoS management system provides a mechanism to detect performance issues, identify root cause(s) of problems, and repair a process by replacing faulty services.;To reduce the burden from monitoring all services, only a set of the most informative services are monitored at runtime. Several monitor selection algorithms are designed for wisely selecting monitoring locations. Three diagnosis algorithms, including Bayesian network (BN) diagnosis, dependency matrix based (DM) diagnosis, and a hybrid diagnosis, are designed for root cause identification. DM diagnosis does not require process execution history and has a lower time complexity than BN. However, BN diagnosis usually achieves a better diagnosis accuracy. The hybrid diagnosis integrates DM and BN diagnosis to get a good diagnosis result while reduces a large portion of the diagnosis cost in BN diagnosis. Moreover, heuristic strategies can be used in hybrid diagnosis to further improve its diagnosis efficiency.;We have implemented a prototype of the QoS and fault management framework in the Llama middleware. The thesis presents the design and implementation of the diagnosis engine, the adaptation manager (for process reconfiguration) in Llama. Diagnosis engine identifies root cause services and triggers the adaptation manager, which decides the solution of service replacement. System performance is studied by using realistic services deployed on networked servers. Both simulation result and system performance study show that our monitoring, diagnosis and recovery approaches are practical and efficient.