Vinicius Tavares Petrucci

Modern software applications are increasingly being deployed in highly dynamic computing environments, leading to frequent changes in their execution environment, such as workload variation, changing resource availability, and component faults. In order to e ectively cope with dynamic scenarios and take advantage of available resources, applications need to be able to adapt their con guration dynamically at run-time in response to changes in their execution environment, preferably without requiring any external intervention from administrators and developers. For instance, web applications in a cluster experience large periods of low utilization and present an opportunity for using dynamic adaptation techniques to reduce energy consumption with little impact on performance and timeliness properties. In recent years, many adaptive policies for web server cluster power management have been investigated. The development of these adaptive policies may be complex in itself, and the required support mechanisms for these policies are implemented in an ad-hoc fashion, or by means of low level (built-in) operating system modules. As a result, the adaptation logic and mechanisms are hard coded (and mixed) within the application code, making it di cult to reuse the basic implementation to experiment with di erent adaptation requirements. In the medium term, when the application goes to a real operational environment, using these ad-hoc approaches substantially hinders application maintenance and evolution activities. In this work, we present a framework-based approach for dynamic adaptation of distributed applications. The proposed framework is used to introduce dynamic adaptation capabilities, intended to address power and performance management, into a server cluster infrastructure. Our approach consists of providing a reusable infrastructure with common elements to monitor and adapt running applications, and an adaptation language to enable one to express high-level adaptation policies, both designed separately from the target application. By using external adaptation mechanisms, our approach enables one to modify and reason about application's adaptation logic with ease. In addition, it allows to reuse the adaptation infrastructure across di erent adaptation requirements, helping to reduce the cost of engineering such adaptive applications.