On Data Center Network Architectures for Interconnecting Dual-Port Servers

During the past decade, various novel data center network (DCN) architectures have been proposed to meet various requirements of large scale data centers. In existing works that consider server-centric DCN architectures, the lengths of a server-to-server-direct hop and a server-to-server-via-a-switch hop are assumed to be equal. We propose the concept of Normalized Switch Delay (NSD) to distinguish a server-to-server-direct hop and a server-to-server-via-a-switch hop, to unify the design and analysis of server-centric DCN architectures for interconnecting servers with two network interface cards. In [1], the authors claim that BCN is the largest known architecture to interconnect dual-port servers, with diameter 7, given a switch port number. We notice that the existing DPillar [2] architecture accommodates more servers than BCN does under the same configurations. Motivated by this fact, we consider a fundamental problem: maximizing the number of dual-port servers, given network diameter and switch port number; and give an upper bound on this maximum number. Then, we propose three novel architectures that try to achieve this upper bound: SWCube, SWKautz, and SWdBruijn, based on the generalized hypercube, Kautz graph, and de Bruijn graph, respectively. The number of servers that SWCube can accommodate is comparable to that of DPillar. The number of servers that SWKautz and SWdBruijn can accommodate is generally greater than that of DPillar. Compared with three existing architectures, the three proposed architectures, SWCube, SWKautz and SWdBruijn demonstrate various advantages. Analysis and simulations on the newly proposed architectures also show that they have nice properties for DCNs, such as low diameter, high bisection width, good fault-tolerance, and the capability of efficiently handling network congestion.