Towards Power- and Energy-efficient Datacenters

As the Internet evolves, cloud computing is now a dominant form of computation in modern lives. Warehouse-scale computers (WSCs), or datacenters, comprising the foundation of this cloud-centric web have been able to deliver satisfactory performance to both the Internet companies and the customers. With the increased focus and popularity of the cloud, however, datacenter loads rise and grow rapidly, and Internet companies are in need of boosted computing capacity to serve such demand. Unfortunately, power and energy are often the major limiting factors prohibiting datacenter growth: it is often the case that no more servers can be added to datacenters without surpassing the capacity of the existing power infrastructure.

This dissertation aims to investigate the issues of power and energy usage in the modern datacenter environment. We identify the source of power and energy inefficiency at three levels in the modern datacenter environment and provides insights and solutions to address each of these problems, aiming to prepare datacenters for critical future growth. We start at the datacenter-level and find that the peak provisioning and improper service placement in multi-level power delivery infrastructures fragment the power budget inside production datacenters, degrading the compute capacity the existing infrastructure can support. We find that the heterogeneity among datacenter workloads is key to address this issue, and design systematic methods to reduce the fragmentation and improve the utilization of the power budget. This dissertation then narrow the focus to examine the energy usage of individual servers running cloud workloads. Especially, we examine the power management mechanisms employed in these servers and find that the coarse time granularity of these mechanisms is one critical factor that leads to excessive energy consumption. We propose an intelligent and low overhead solution on top of the emerging finer granularity voltage/frequency boosting circuit to effectively pinpoints and boosts queries that are likely to increase the tail distribution and can reap more benefit from the voltage/frequency boost, improving energy efficiency without sacrificing the quality of services. The final focus of this dissertation takes a further step to investigate how using a fundamentally more efficient computing substrate, field programmable gate arrays (FPGAs), benefit datacenter power and energy efficiency. Different from other types of hardware accelerations, FPGAs can be reconfigured on-the-fly to provide fine-grain control over hardware resource allocation, and presents a unique set of challenges for optimal workload scheduling and resource allocation. We aim to design a set coordinated algorithms to manage these two key factors simultaneously and fully explore the benefit of deploying FPGAs in the cloud environment.