Rethinking Fat-Tree Topology Design for Cloud Data Centers

摘要

Data center network (DCN) topologies have recently been the focus of many researchers due to their vital role in achieving high DCN performances in terms of scalability, power consumption, throughput, and traffic load balancing. This paper presents a comprehensive comparison between two most commonly used DCN topologies, Fat-Tree and BCube, with a focus on structure, addressing and routing, and proposes a new DCN topology that is better suited for nowadays data center networks. We show that our proposed topology, termed Circulant Fat-Tree, alleviates traffic congestion at the core switches, improves network latency, and increases robustness against switch and server failures when compared to traditional Fat-Tree DCN topologies.