Application Deployment at Distributed Clouds

A current trend in Networking and Cloud Computing is to provide compute resources not only at large data centres but also at many smaller centres or within the network itself, e.g., nearby backbone routers. These resources are often closer to customers and, hence, have a better connectivity to the customers than resources of few large data centres farther away. What are the benefits of deploying applications at many sites? What are the challenges? This thesis answers these questions and focuses (i) on optimisation problems capturing the relationship between network-related service quality (response time and data rate) and incurred resource costs, (ii) on application architectures comprising multiple components and their interdependencies, (iii) on design and implementation of an automated management system that deploys, scales, and configures an application at runtime. More concretely, a Template graph describes the application components and scaling properties. The presented Template Embedding Problem uses this Template and allocates resources, assigns customers to resources, and scales the application in a single, compact step yielding optimal solutions. Another investigation integrates queuing systems into a Facility Location Problem in order to improve accuracy of response time optimised resource allocation. The resulting problem is convex and different approximations using the linearisation technique trade off solving time and solution quality. While the two problems have to be solved centrally, a distributed algorithm solving a Facility Location Problem finds solutions with good quality. Finally, a system design comprising architecture, interfaces, and mechanisms, its proof of concept, and a testbed that emulates a Distributed Cloud Computing scenario are presented. The prototype and testbed served as demonstrators at public events.