A paper from Google describing how to make data center network energy proportional, i.e. the energy consumption in watts is proportional to the utilization. The current status is that, servers are more power-hungry than network because of the over-subscription. However, server is also easier to be energy proportional. Network equipments, because they are always-on, they use roughly the same amount of energy regardless the utilization.

In the paper, the energy proportionality in network equipments is provided by increasing the dynamic range of links. However, it also proposed to use flattened butterfly network so that the energy proportionality can be more efficient.

The flattened butterfly network (FBFLY, see references in paper) is modeled as “\(k\)-ary \(n\)-flat” network. Where \(k\) is the number of nodes in a dimension and \(n\) is the dimension. It provides a low diameter network compared to Clos, and few physical links are used.

The FBFLY network baseline is a \(k\)-ary 2-flat network, which there are \(k\) switches formed a mesh. Each switch connects to \(c\) hosts (concentration \(c\)). To make a 2-flat network (2-dimensional) into 3-flat network (3-dimensional), we replicate the network for \(k\) times and connect each switch with its counterpart in the \(k-1\) replicates. So in general, there are \(k^{n-1}\) switches in \(k\)-ary \(n\)-flat network, each switch connects to \((n-1)(k-1)\) other switches and \(c\) hosts. The routing in FBFLY network is the following: Each move within the same dimension is one-hop.

Bibliographic data

   title = "Energy Proportional Datacenter Networks",
   author = "Dennis Abts and Michael R. Marty and Philip M. Wells and Peter Klausler and Hong Liu",
   booktitle = "Proc. ISCA'10",
   page = "338--347",
   month = "June 19-23",
   year = "2010",
   address = "Saint-Malo, France",