August 17–22, 2008, Seattle, Washington, USA | Mohammad Al-Fares, Alexander Loukissas, Amin Vahdat
This paper presents a scalable and cost-effective data center network architecture that leverages commodity Ethernet switches to support the full aggregate bandwidth of large clusters. The authors argue that while current network architectures, which often use specialized hardware and routing protocols, can scale to thousands of nodes but suffer from significant bandwidth limitations and high costs, a fat-tree topology can achieve full bisection bandwidth with lower costs and better performance. The fat-tree architecture uses a hierarchical structure of interconnected switches, where each switch has multiple ports for both intra-pod and inter-pod communication. The paper introduces a two-level routing table to ensure even distribution of traffic across the network, and proposes flow classification and scheduling techniques to optimize resource utilization. The authors also discuss fault tolerance and power/heat management, showing that their design can reduce power consumption and heat dissipation by 56.6% and 56.5%, respectively, compared to traditional hierarchical designs. A prototype implementation using Click, a software router framework, is described, and performance evaluations demonstrate the effectiveness of the proposed architecture.This paper presents a scalable and cost-effective data center network architecture that leverages commodity Ethernet switches to support the full aggregate bandwidth of large clusters. The authors argue that while current network architectures, which often use specialized hardware and routing protocols, can scale to thousands of nodes but suffer from significant bandwidth limitations and high costs, a fat-tree topology can achieve full bisection bandwidth with lower costs and better performance. The fat-tree architecture uses a hierarchical structure of interconnected switches, where each switch has multiple ports for both intra-pod and inter-pod communication. The paper introduces a two-level routing table to ensure even distribution of traffic across the network, and proposes flow classification and scheduling techniques to optimize resource utilization. The authors also discuss fault tolerance and power/heat management, showing that their design can reduce power consumption and heat dissipation by 56.6% and 56.5%, respectively, compared to traditional hierarchical designs. A prototype implementation using Click, a software router framework, is described, and performance evaluations demonstrate the effectiveness of the proposed architecture.