Introducing Service-level Awareness in the Cloud

Managing the resources of a virtualized data-center is a key issue in cloud computing. Existing research mostly assumes that applications are either allocated the required resources or fail. Combined with the fact that most cloud applications have dynamic resource requirements, this imposes a fundamental limitation to cloud computing: To guarantee on-demand resource allocations, the data-center needs large spare capacity, leading to inefficient resource utilization. This is especially problematic when dealing with unexpected events, such as flash crowds, hardware failures and performance inference among applications. These phenomena are well-known and software is readily written to cope with them, as long as resource provisioning is sufficient. However, given the short duration and large magnitude of such events, provisioning enough capacity is often economically unfeasible. Hence, the data-center may become overloaded, rendering hosted applications unresponsive. To efficiently and robustly deal with unexpected events, we introduce service-level awareness in the cloud. Application are augmented with a dynamic parameter, the service level, that monotonically affects both their resource requirements and their delivered end-user experience. For example, online shops offer end-users recommendations of similar products. No doubt, such recommender engines greatly increase usability, however, they are highly resource demanding. By selectively deactivating the corresponding code, resource requirements can be controlled at the expense of end-user experience. In case of unexpected events, the infrastructure can simply ask applications to temporarily reduce their requirements. Consequently, end-user experience is downgraded, but the user is at least provided with partial content in a timely manner. We built the necessary software to add service level-awareness to clouds, with contributions both on application-side as well as infrastructure-side