On the Energy Efficiency of Networked Systems

Energy is a first-class resource for datacenter operators since its cost is the biggest limiting factor in scaling a large computing facility. The solution embraced by major operators is to build their facilities in strategic geographical locations and to abandon expensive specialized hardware for cheap commodity systems. However, such systems are not efficient when it comes to energy and a considerable amount of research effort has been put in finding a solution to this problem. Furthermore, the need for more programmable and flexible networking devices is pushing the need for hardware commoditization also within the datacenter network. In this thesis we propose two solutions aimed at improving the overall energy efficiency of a datacenter facility. The first address efficiency in computing, by proposing a different hardware architecture for server systems. We propose a hybrid architecture that blends traditional server processors with very-low-power processors from the mobile devices world. The second solution envisions the usage of current server platforms as network switches or routers and provides guidelines for the implementation of power saving algorithms that do not affect peak performance while saving up to 50% power. This work is based on both theoretical modeling and simulation and experimentation with real-world prototypes.