Power-efficient resource allocation in Cloud data centers

Cloud Data Centers (DCs) provide computational resources and services for a huge number of customers. A great variety of applications, such as online storage, scientific computing, and web services, requires a set of resources located in DCs, which should be available when needed in a flexible and dynamic way. Therefore, Cloud providers need to manage DC resources carefully in order to satisfy as many service requests as possible to increase their own revenues. Indeed, the major challenge in DCs deployment consists in facing the rapidly increasing demand of Cloud services, which asks for a flexible and dynamic design of the Cloud. Another important aspect that should be considered is the interaction between the Cloud and the 5G network. The 5G strict requirements can be matched only integrating the Cloud functionality with the next-generation mobile networks applying the Mobile Edge Computing (MEG) and Mobile Edge Cloud (MEC) paradigms. This thesis offers a detailed overview on power-efficient techniques for DC management, and it mainly provides several contributions: first, it tackles the resource allocation problem introducing novel power-aware placement strategies for tasks and Virtual Machines (VMs) to be placed within the DC; next, it considers the DC networking part showing several load balancing and resilience mechanisms for Software Defined DCs; then, it outlines a topology-aware scheduler for distributed applications using the Message Passing Interface (MPI) library, that is used to avoid network congestion and minimize the overall communication length among all the processes. Finally, the thesis focuses on the Internet of Things (IoT) scenario, where devices produce data that can be consumed by softwares (i.e., tasks or VMs) typically running in the DC. More in detail, this thesis deals with the communication among IoT devices using the Time Synchronous Channel Hopping (TSCH), and it presents a centralized scheduler able to meet both real-time constraints and communication deadlines. The results show that the developed techniques improve physical resource utilization in DCs reducing the power costs and increasing the return on investments (RoI): resource allocation is performed using multi-objective techniques which optimize both power consumption and DC performance by computing the allocation pattern of thousands of VMs and tasks in less than ten seconds. Load balancing and resilience techniques provide robustness and high resource availability for Software Defined DCs while the MPI scheduler performs the communication between processes avoiding network congestion and exploiting the communication parallelism. Conclusively, it will be shown that in the IoT scenario results are promising: the implemented scheduler significantly reduces the number of violated deadlines with respect to other classical schedulers.