Automated cloud reconfiguration with HTN planning : engineering a domain model and performance insights

Abstract

With the emergence of cloud technologies, how applications are developed, deployed, and maintained has drastically changed. Despite the many advantages that Cloud Computing has brought, today's software services are more complex due to their dependencies and entanglements. Managing these entanglements between software services makes reconfiguration processes complicated. Many state-of-the-art tools have been developed that allow for reconfigurations of software services using declarative or imperative approaches. Concerning imperative approaches, it is hard to keep track of every change in reconfiguration scripts, and understanding the general reconfiguration strategy requires much effort. Regarding tools following a declarative approach, reconfiguration files are much clearer and more maintainable because they are easier to declare software requirements. Nevertheless, many of these tools fall short when cyclic dependencies exist among software components. Moreover, although end-users define their requirements declaratively, there is much tight coupling at these tools' core logic implementation. Using an HTN planner introduces a layer of abstraction around the core logic of reconfigurations, removing tight couplings. To use an HTN planner, one must define the domain model and problem cases separately by using formats suitable for planners. Although defining a problem domain is a complex task that must be carried out by domain experts, HTN planners enable us to take advantage of problem domains' hierarchical structure. Thus, this task can be done much more simply. In this thesis, instead of implementing a reconfiguration logic, we define our reconfiguration domain based on a cloud component model called Aeolus and problem cases in HPDL format. We use an existing HTN planner to produce a plan, i.e., a sequence of actions, for achieving the desired reconfiguration goals. Furthermore, we perform our experiments to test the effectiveness of using HTN planners with various problem sizes.