Your Data Center Is a Router: The Case for Reconfigurable Optical Circuit Switched Paths

The rising tide of data-intensive, massive scale cluster computing is creating new challenges for traditional, hierarchical data center networks. In response to this challenge, the research community has begun exploring novel interconnect topologies to provide high bisection bandwidth—examples include Fat trees [2, 12, 8], DCell [9], and BCube [10], among a rapidly growing set of alternatives, many adapted from earlier solutions from the telecom and supercomputing areas. We argue that these solutions may provide too much— full bisection bandwidth on packet timescales—at too high a cost—literally tons of wiring and thousands of switches. In this work, we suggest that research should take a look back not only at historical topologies, but also historical technologies. More specifically, we suggest that a hybrid packet-switched/circuit-switched network can provide the functions and ease-of-use of today’s allpacket networks, while providing high bandwidth for a large class of applications at lower cost and lower network complexity. Taking advantage of this network requires, however, a philosophical change to the design of data center networks. We propose to augment the electrical switch architecture with an optical circuit-switched network. Implementing this approach requires a network re-design to provide substantial pre-optical queueing at the nodes, treating the entire data center as one large virtually output-queued router. We explain this argument briefly, and expand upon our proposed solution in the sections that follow.