Scaling of capacity and reliability in data center networks

The traditional connectivity model within the data center is that of a hierarchical tree with redundant connections (\fat tree") and with a top node consisting of one or more routers that bring in (and send out completed) requests for processing. In this paper we examine alternative connectivity models for large-scale data centers. In the first model, we examine hypergrids as the structure connecting switches and routers to the edge server racks. In the second model, we examine random graphs as the interconnecting network. We compare and contrast the capacity, congestion and reliability requirements for these relative to fat-trees. We show that, as the system size increases and for uniform switch-end-to- switch-end demand, the fat-tree configuration emerges as an expensive option demanding higher port density switches but has low congestion and high reliability. In contrast, the random graph model shows the same low level of congestion, lower cost due to fewer ports and reasonable reliability, whereas the hypergrid model does not require scaling of switch ports, provides high reliability but exhibits higher congestion.