UC San Diego

Today's data centers consist of several thousand PCs that provide massive amounts of computational power and storage capacity in a cost-effective manner. Due to the highly distributed nature of the applications running in these clusters, intra-node communication bandwidth is key to the performance of the data center. Unfortunately, communication bandwidth is in many cases a significant bottleneck in large-scale clusters. Although current data center networks leverage high-end switching elements with massive switching capacity, current architectures are heavily oversubscribed and fail to meet with the requirements in intra-node communication bandwidth. What is worse, performance greatly degrades with larger cluster sizes, while the cost increases exponentially with cluster size. This thesis presents the implementation of a network architecture that leverages commodity Ethernet switches and supports full bisection bandwidth in clusters consisting of tens of thousands of hosts. Together with Mohammad Al-Fares and Amin Vahdat, we have shown that by appropriately interconnecting switches and using a novel routing algorithm, we can achieve significantly better performance than current high-end solutions at a fraction of the cost. This architecture is fully compatible with Ethernet and applications running on TCP/IP, and requires no end-host modifications. The focus of this thesis is the implementation of our architecture in hardware. More specifically, I show that our architecture can be implemented in commodity switches in a straightforward manner, requiring only slight modifications in their hardware. The thesis also focuses on the implementation and evaluation of the architecture in a network simulator