Adaptively Fitting Network Topologies to Traffic Locality in Clos-Type Data Center Networks

Localized traffic is ubiquitous in today's data centers. In this context, we propose to fit the topologies of the underlying network into the traffic locality, so that we can improve the efficiency of network resource utilization. We make our network infrastructure to be versatile, in the sense that its topology can be fitted into the profile of the traffic. We describe our network's topological architecture, design its addressing and routing schemes, and validate its adaptively fitting capability. We also evaluate its performance and compare it with fat-tree, a representative Clos-type data center network architecture. The evaluation results demonstrate that the our network can deliver the same throughput as fat-tree, but use significantly reduced network resources.