Architecting Efficient Data Centers.

Abstract

Data center power consumption has become a key constraint in continuing to scale Internet services. As our society’s reliance on “the Cloud” continues to grow, companies require an ever-increasing amount of computational capacity to support their customers. Massive warehouse-scale data centers have emerged, requiring 30MW or more of total power capacity. Over the lifetime of a typical high-scale data center, power-related costs make up 50% of the total cost of ownership (TCO). Furthermore, the aggregate effect of data center power consumption across the country cannot be ignored. In total, data center energy usage has reached approximately 2% of aggregate consumption in the United States and continues to grow.

This thesis addresses the need to increase computational efficiency to address this grow- ing problem. It proposes a new classes of power management techniques: coordinated full-system idle low-power modes to increase the energy proportionality of modern servers. First, we introduce the PowerNap server architecture, a coordinated full-system idle low- power mode which transitions in and out of an ultra-low power nap state to save power during brief idle periods. While effective for uniprocessor systems, PowerNap relies on full-system idleness and we show that such idleness disappears as the number of cores per processor continues to increase. We expose this problem in a case study of Google Web search in which we demonstrate that coordinated full-system active power modes are necessary to reach energy proportionality and that PowerNap is ineffective because of a lack of idleness. To recover full-system idleness, we introduce DreamWeaver, architectural support for deep sleep. DreamWeaver allows a server to exchange latency for full-system idleness, allowing PowerNap-enabled servers to be effective and provides a better latency- power savings tradeoff than existing approaches. Finally, this thesis investigates workloads which achieve efficiency through methodical cluster provisioning techniques. Using the popular memcached workload, this thesis provides examples of provisioning clusters for cost-efficiency given latency, throughput, and data set size targets.