Author:

Abstract:

Services computing facilitates the creation of large scale applications. Servicesnbsp;relatively small and manageable software units with clearly defined interfaces. Applications then consist of orchestrated invocations of services, the so-called composite services. The services on which a composite service relies - called component services - have various quality of service (QoS) characteristics, such as performance, reliability, availability, accuracy. Such quality parameters can be used by a composite service to select component services when called for. Service selection and composition is particularly challenging when the system is large-scale - consisting of thousands of nodes, components and composite services - and dynamic - where QoS varies. Particularly challenging is the situation where potentially large failures can occur. This thesis ambition is to create a highly resilient system for dynamic service compositions. While traditionally, a resilient system is able to deal with failures, our aim is to conceive a system that considers failures as ‘business as usual’, to which it gracefully molds itself. The large or potentially huge scale of such systems (involving tens of thousands of nodes and services) makes a central selection and composition nbsp;infeasible. In our research, we investigate a decentralized nbsp;and self- organizing approach to dynamic service composition. In particular, we study delegateMAS, a coordination mechanism originally targeted for large-scale coordination and control applications, such as traffic and logistics management, where entities need to coordinate over nbsp;Such coordination and control systems are intrinsically dynamic due to changes in operational (uncertainty of service time, orders, travel demand) or exceptional conditions (vehicle failures, infrastructure problems). nbsp;this thesis we, first, define a decentralized solution for dynamic service composition using delegate MAS. We define TaskAgents that are responsible for enacting composite service instances, and ResourceAgents that manage the usage of component services. Two delegate MASs are defined, nbsp;exploring compositions and for propagating information about intended compositions. Second, we thoroughly investigate the scaling of the system. nbsp;ran experiments that were large and huge in scale (up to tens of thousands of nodes and services). Third, we assess the behavior of the system under failing conditions, including drastic failure scenarios. These experiments show that the approach is effective, efficient, scales linearly, and cannbsp;even with severe failures. We validate our approach by applying it to different domains and by performing a thorough evaluation of it. We conclude this thesis showing that it is possible to create service compositions which can cope with failures, without relying on centralized solutions.