Abstract

Cloud Computing is becoming the mainstream paradigm, as organizations, both large and small, begin to harness its benefits. Cloud computing gained its success for giving IT exactly what it needed: The ability to grow and shrink computing resources, on the go, in a cost-effective manner, without the anguish of infrastructure design and setup. The ability to adapt computing demands to market fluctuations is just one of the many benefits that cloud computing has to offer, this is why this new paradigm is rising rapidly. According to a Gartner report, the total sales of the various cloud services will be worth 204 billion dollars worldwide in 2016. With this massive growth, the performance of the underlying infrastructure is crucial to its success and sustainability. Currently, cloud computing heavily depends on data centers for its daily business needs. In fact, it is through the virtualization of data centers that the concept of "computing as a utility" emerged. However, data center virtualization is still in its infancy; and there exists a plethora of open research issues and challenges related to data center virtualization, including but not limited to, optimized topologies and protocols, embedding design methods and online algorithms, resource provisioning and allocation, data center energy efficiency, fault tolerance issues and fault tolerant design, improving service availability under failure conditions, enabling network programmability, etc.
This dissertation will attempt to elaborate and address key research challenges and problems related to the design and operation of efficient virtualized data centers and data center infrastructure for cloud services. In particular, we investigate the problem of scalable traffic management and traffic engineering methods in data center networks and present a decomposition method to exactly solve the problem with considerable runtime improvement over mathematical-based formulations. To maximize the network's admissibility and increase its revenue, cloud providers must make efficient use of their's network resources. This goal is highly correlated with the employed resource allocation/placement schemes; formally known as the virtual network embedding problem. This thesis looks at multi-facets of this latter problem; in particular, we study the embedding problem for services with one-to-many communication mode; or what we denote as the multicast virtual network embedding problem. Then, we tackle the survivable virtual network embedding problem by proposing a fault-tolerance design that provides guaranteed service continuity in the event of server failure. Furthermore, we consider the embedding problem for elastic services in the event of heterogeneous node failures. Finally, in the effort to enable and support data center network programmability, we study the placement problem of softwarized network functions (e.g., load balancers, firewalls, etc.), formally known as the virtual network function assignment problem. Owing to its combinatorial complexity, we propose a novel decomposition method, and we numerically show that it is hundred times faster than mathematical formulations from recent existing literature.