Umeå University's logo

Energy efficiency has become a fundamental concern in data centers, raising issues to all energy-related costs, including capital costs, operating expenses, and environmental impact. Energy inefficiency is mainly caused by unoptimized use of energy by sub-components of these data centers. For example, energy can be lost due to transport and conversion, cooling, and lightning. Energy can be wasted while running an idle server or when using unoptimized functions to perform a task. Addressing this problem as a whole requires redesigning data centers, rethinking components, and implementing energy-aware algorithms for data center operation. As one step towards achieving this goal, this thesis focuses on the development of resource allocation algorithms to improve the energy efficiency of servers in virtualized data centers. The thesis proposes models, techniques, and algorithms to improve data center resource efficiency for optimized power and performance. We present approaches that takes advantage of horizontal scaling, vertical scaling, CPU frequency scaling, and the scheduling of FPGAs to reduce the power consumption of servers while meeting performance requirements of applications. We design online performance and power models to capture system behaviour while adapting to changes in the underlying infrastructure. Based on these models, we propose controllers that dynamically determine power-efficient resource allocations. We also devise optimization strategies for colocated applications and evaluate their suitability in a number of scenarios. The proposed strategies simplify the handling of trade-offs between power minimization and meeting performance targets. We also consider fluctuations in resource allocation in decision making. Additionally, we propose a scheduling algorithm for the use of custom hardware accelerators, FPGAs, and their integration to data centers for the purpose of increasing processing and energy efficiency. Our evaluation results demonstrate that our proposed approaches provide improved energy-efficient management of resources in virtualized data centers.

Energy efficiency has become an increasingly important concern in data centers because of issues associated with energy consumption, such as capital costs, operating expenses, and environmental impact. While energy loss due to suboptimal use of facilities and non-IT equipment has largely been reduced through the use of best-practice technologies, addressing energy wastage in IT equipment still requires the design and implementation of energy-aware resource management systems. This thesis focuses on the development of resource allocation methods to improve energy efficiency in data centers. The thesis employs three approaches to improve efficiency for optimized power and performance: scaling virtual machine (VM) and server processing capabilities to reduce energy consumption; improving resource usage through workload consolidation; and exploiting resource heterogeneity.

To achieve these goals, the first part of the thesis proposes models, algorithms, and techniques that reduce energy usage through the use of VM scaling, VM sizing for CPU and memory, CPU frequency adaptation, as well as hardware power capping for server-level resource allocation. The proposed online performance and power models capture system behavior while adapting to changes in the underlying infrastructure. Based on these models, the thesis proposes controllers that dynamically determine power-efficient resource allocations while minimizing performance penalty.

These methods are then extended to support resource overbooking and workload consolidation to improve resource utilization and energy efficiency across the cluster or data center. In order to cater for different performance requirements among collocated applications, such as latency-sensitive services and batch jobs, the controllers apply service differentiation among prioritized VMs and performance isolation techniques, including CPU pinning, quota enforcement, and online resource tuning.

This thesis also considers resource heterogeneity and proposes heterogeneousaware scheduling techniques to improve energy efficiency by integrating hardware accelerators (in this case FPGAs) and exploiting differences in energy footprint of different servers. In addition, the thesis provides a comprehensive study of the overheads associated with a number of virtualization platforms in order to understand the trade-offs provided by the latest technological advances and to make the best resource allocation decisions accordingly. The proposed methods in this thesis are evaluated by implementing prototypes on real testbeds and conducting experiments using real workload data taken from production systems and synthetic workload data that we generated. Our evaluation results demonstrate that the proposed approaches provide improved energy management of resources in virtualized data centers.