Umeå University's logo

The overall goal of the work presented in this thesis was to find ways of managing power-performance tradeoffs in cloud data centers. To this end, the relationships between the power consumption of data center servers and the performance of applications hosted in data centers are analyzed, models that capture these relationships are developed, and controllers to optimize the use of data center infrastructures are proposed.

The studies were motivated by the massive power consumption of modern data centers, which is a matter of significant financial and environmental concern. Various strategies for improving the power efficiency of data centers have been proposed, including server consolidation, server throttling, and power budgeting. However, no matter what strategy is used to enhance data center power efficiency, substantial reductions in the power consumption of data center servers can easily degrade the performance of hosted applications, causing customer dissatisfaction. It is therefore crucial for data center operators to understand and control power-performance tradeoffs.

The research methods used in this work include experiments on real testbeds, the application of statistical methods to create power-performance models, development of various optimization techniques to improve the power efficiency of servers, and simulations to evaluate the proposed solutions at scale.

This thesis makes multiple contributions. First, it introduces taxonomies for various aspects of data center configuration, events, management actions, and monitored metrics. We discuss the relationships between these elements and support our analysis with results from a set of testbed experiments. We demonstrate limitations on the usefulness of various data center management actions for controlling power consumption, including Dynamic Voltage Frequency Scaling (DVFS) and Running Average Power Limit (RAPL). We also demonstrate similar limitations on common measures for controlling application performance, including variation of operating system scheduling parameters, CPU pinning, and horizontal and vertical scaling. Finally, we propose a set of power budgeting controllers that act at the application, server, and cluster levels to minimize performance degradation while enforcing power limits.

The results and analysis presented in this thesis can be used by data center operators to improve the power-efficiency of servers and reduce overall operational costs while minimizing performance degradation. All of the software generated during this work, including controller source code, virtual machine images, scripts, and simulators, has been open-sourced.