Energy efficiency of Zipf traffic distributions within Facebook's data center fabric architecture

Open architectures like the one recently unveiled by Facebook allow a detailed assessment of the energy efficiency of commercial data centers. This paper explores the fit of Zipf-like distributions typical of network traffic, to updates of user pages and the entity graph, for the new Facebook data center network architecture. We find that network resource consumption could be reduced by as much as 40-50% through several changes, either to the software, or to the data center design. Of these, employing a connected hub-and-spoke subgraph representation for each popular node, with each pod operating locally on its node of the subgraph, appears to hold the most energy savings potential. This work is part of a larger effort to more completely characterize the efficiency of data center computer-and network architectures beyond the normal reporting of facility power utilization efficiency (PUE), which is blind to energy proportionality and other aspects of the efficiency within the computer- and network architecture, or IT portion, of the data center.