Infrastructure Cost Comparison of Intra-Data Centre Network Architectures
Ronald Romero Reyes, Thomas Bauschert · 2018
This paper presents an evaluation and comparison of the infrastructure costs of five intra-data centre network architectures, namely, leaf-spine, fat-tree, hybrid fat-tree, Facebook 4-Post and a new Facebook fabric. A bottom-up cost calculation approach is used whereby all network architectures are designed so as to meet target oversubscription ratios. The resulting designs provide a list of materials that specify both the switching and cabling components required to build the networks. With this information the costs are evaluated from price books of network components available on the market. Different data centre sizes - defined by the number of physical servers housed in the data centre - are considered to investigate how the costs scale and compare across all network architectures. The results reveal that the economic savings brought in by each architecture depend on the data centre size. The most cost-efficient architecture is leaf-spine. The hybrid fat-tree also provides low costs for large-size data centres. In all considered scenarios it has been observed that the leaf (or top of the rack) switches are the dominant cost drivers.