Energy- efficient and SLA-based management of IaaS Cloud Data Centers

Cloud computing is progressively being adopted in different scenarios by offering on-demand, flexible, and high-scalability access to large-scale distributed resources, with Service Level Agreements-driven management. Virtualization is the basic technology of cloud computing, rendering flexible and scalable system services to cloud systems. As these distributed systems become more widespread, companies and resource providers are building large warehouse-sized data centers to cope with increasing demand for computing resources. However, the amount of electrical energy consumed by data centers increases with the amount of computing power instaled. In the same line, as compute systems grow in size and in complexity, failure events become norm instead of exception, increasing the energy waste even more and affecting the Quality-of-Service of the system perceived by end-users. Moreover, current virtualization technologies do not provide performance isolation, meaning that two applications running in independent virtual machines can interfere in the execution of each other when they share the same physical server, hence violating the Quality-of-Service constraints. This thesis presents two improved mechanisms with the twofold objective of saving electrical costs and respecting the Service Level Agreements stipulated with users. The first objective is achieved by allying an energy optimizing mechanism to detect and mitigate energy inefficiencies, and virtualization tools to provide proactive fault-tolerance and energy efficiency to virtual clusters. Energy inefficiencies are reduced by dynamically consolidating virtual machines and switching off and on physical nodes according to resource demand. Consolidation is implemented based on vertical and horizontal elasticity of resources. The second objective is achieved by articulating a performance estimator mechanism to detect deviation from application Quality-of-Service requirements, and virtualization tools to adapt the map of virtual machines to servers. Two types of workloads are considered, namely CPUand network-bound workloads, with different Qualityof-Service constraints. The analysis of the performance of the proposed mechanisms is done via simulation and experiments in real cloud testbed. The workloads, failures, and performance characteristics used in tests are coherent with the attributes outlined in state-of-the-art studies over large-scale data centers. In the case of the first objective, the results indicate that the proposed strategy improves the work per Joule ratio by approximately 12.9% and the working efficiency by almost 15.9% compared with other state-of-the-art algorithms. For the second objective, the results show that the proposed performance enforcing mechanism is able to fulfil contracted SLAs of real-world environments, while reducing energy costs up to 21%.