Compound graph based hybrid data center topologies

摘要

In large-scale data centers, many servers are interconnected via a dedicated networking structure, so as to satisfy specific design goals, such as the low equipment cost, the high network capacity, and the incremental expansion. The topological properties of a networking structure are critical factors that dominate the performance of the entire data center. The existing networking structures are either fully random or completely structured. Although such networking structures exhibit advantages on given aspects, they suffer obvious shortcomings in other essential fields. In this paper, we aim to design a hybrid topology, called R3, which is the compound graph of structured and random topology. It employs random regular graph as a unit cluster and connects many such clusters by means of a structured topology, i.e., the generalized hypercube. Consequently, the hybrid topology combines the advantages of structured as well as random topologies seamlessly. Meanwhile, a coloring-based algorithm is proposed for R3 to enable fast and accurate routing. R3 possesses many attractive characteristics, such as the modularity and expansibility at the cost of only increasing the degree of any node by one. Comprehensive evaluation results show that our hybrid topology possesses excellent topology properties and network performance.