Architecture Design and Management for Data Center Networks

Kai Chen
EECS Department at Northwestern University


Abstract

Driven by technology advances and economic forces, data centers are being built around the world to support various infrastructure services (e.g., GFS, Map-reduce and Dryad) and Internet/cloud applications (e.g., web search, e-commerce, social networking). How to design a sustainable data center network architecture and how to manage it are two fundamental challenges for data center research, and form the basis of my PhD thesis work. In this talk, I will focus on introducing two coupled systems called OSA and DAC. Specifically, OSA is a novel optical switching architecture for data center networks. At its core, by exploiting runtime reconfigurable optical devices, OSA flexibly adapts its topology and link capacities to various traffic patterns and thus achieves high network bisection bandwidth. The key idea I hope to deliver via OSA is that: a thin but flexible architecture, instead of a static fat one, can better support data center networking in a simple, cost-effective and sustainable way. On the other hand, DAC is a management system that performs address autoconfiguration for OSA and other generic data center networks. DAC is motivated by the key observation that data center networks encode locality and topology information into their server/switch addresses for performance and routing purposes which makes the traditional address autconfiguration protocols, like DHCP, not applicable any more in the new data center environment. The core of DAC is to reduce the address autoconfiguration problem to the graph isomorphism (GI) problem and solve it efficiently by leveraging the data center topology properties. In both designs, we have evaluated and validated our systems via extensive analysis, simulations and prototype implementations. At the end of my talk, I will summarize the other research in my PhD study and foresee the future plans.