Non-overlapping rings: A new architecture for designing switch clusters in data centers

Abstract

The data center network is an important component of the modular cloud data center. Currently, a common architecture involves connecting servers in the data center by switches arranged in a generalized tree topology. However, cloud computing service providers are considering adding “east-west” horizontal links between top of rack (TOR) switches to improve latency performance and/or reducing cost between groups of TOR switches and to off-load traffic from the tree network. One possible method to achieve this involves connecting TOR switches into clusters. Some clusters would be designated as high performance clusters providing low latency, while other clusters would be optimized for cost, while still providing adequate latency performance. In this paper, we present an algorithm for designing lower-cost network clusters using non-overlapping rings (NOR). We describe some of the performance characteristics of these NOR clusters such as coverage vs. number of rings. We show that cost savings can be achieved with minimal increase in end-to-end delay, and that good performance can also be maintained under failure conditions. We also describe how incremental network growth can be easily accommodated using a smooth growth design through our algorithm. As cabling is a major practical consideration for cluster based architectures, we also define a simple cabling architecture for NOR networks.