ROPE: Reducing the Omni-kernel Power Expenses

Abstract

Over the last decade, power consumption and energy efficiency have arisen as important performance metrics for data center computing systems hosting cloud services. The incentives for reducing power consumption are several, and often span economic, technological, and environmental dimensions.

Because of the vast number of computers currently employed in data centers, the economy of scale dictates that even small reductions in power expenditure on machine level can amount to large energy savings on data center scale.

Clouds commonly employ hardware virtualization technologies to allow for higher degrees of utilization of the physical hardware. The workloads encapsulated by virtual machines constantly compete for the available physical hardware resources of their host machines. To prevent execution of one workload from seizing resources that are intended for another, absolute visibility and opportunity for resource allocation is necessary. The Omni-kernel architecture is a novel operating system architecture especially designed for pervasive monitoring and scheduling. Vortex is an experimental implementation this architecture.

This thesis describes ROPE (Reducing the Omni-kernel Power Expenses), which introduces power management functionality to the Vortex implementation of the Omni-kernel architecture. ROPE reduces the power consumption of Vortex, and does so while limiting performance degradation. We evaluate the energy savings and performance impacts of deploying ROPE using both custom tailored and industry standard benchmarks. We also discuss the implications of the Omni-kernel architecture with regards to power management, and how energy efficiency can be accommodated in this architecture.