LaCoDa: Layered connected topology for massive data centers

Abstract

One of the fundamental challenges of existing data centers is to design a network that interconnects massive number of servers, and therefore providing an efficient and fault-tolerant routing service to upper-layer applications. Several solutions have been proposed (e.g. FatTree, DCell and BCube), however they either scale too fast (i.e., double exponentially) or too slow. This paper proposes a new data center topology, called LaCoDa, that combines the advantages of previous topologies while avoiding their limitations. LaCoDa uses a small node degree that matches physical restriction for servers, and it also interconnects a large number of servers while reducing the wiring complexity and without affecting the existing servers. LaCoDa can interconnect up to millions of servers with a small number of port switches by increasing the number of directly connected clusters per layer, and therefore avoiding the cluster connections repetitions. The theoretical and experimental results show that the average path length (APL) of a 1000-node LaCoDa is smaller than the APL of a 23-node FlatTree and 11-node BCube and DCell. Moreover, with 6-port switches and a node degree of 4, the bisection bandwidth of 4.19 Million-node LaCoDa is 7% and 14% bigger than the bisection bandwidth of 310-node DCell and BCube, respectively. Consequently, LaCoDa provides better performance than DCel, BCube, FatTree in terms of average path length, throughput, latency and bisection bandwidth.