HyperOXN: A Novel Data Center Topology Driven by Machine Learning

Abstract

Driven by emerging applications such as machine learning, cloud computing, and big data, modern data center network architecture has been evolving to meet the challenging requirements, like scalability, agility, energy efficiency, and high performance. In the meantime, artificial intelligent applications are expediting the convergence of high-performance computing and data centers. To address the challenges noted above, we investigate communication patterns for emerging applications and find that the dynamic and diverse *-cast traffic play a significant impact on the performance of new applications. Inspired by Hypermeshes topology, this paper presents HyperOXN, a novel cost-efficient topology for exascale DCNs. HyperOXN takes advantage of passive optical wavelength division multiplexing technologies. We show that HyperOXN outperforms other topologies in latency, cost, and scalability.