Energy-aware routing in data center network

TLDR

Data center networks interconnect vast numbers of servers and require efficient, fault‑tolerant routing; however, tree‑based designs and newer 1:1 oversubscription architectures such as Fat‑Tree and BCube deploy many devices that are often underutilized when traffic is below peak. Idle network devices in these architectures consume excess energy, posing a significant challenge for data center operators.

Abstract

The goal of data center network is to interconnect the massive number of data center servers, and provide efficient and fault-tolerant routing service to upper-layer applications. To overcome the problem of tree architecture in current practice, many new network architectures are proposed, represented by Fat-Tree, BCube, and etc. A consistent theme in these new architectures is that a large number of network devices are used to achieve 1:1 oversubscription ratio. However, at most time, data center traffic is far below the peak value. The idle network devices will waste significant amount of energy, which is now a headache for many data center owners.