This paper presents a scalable, commodity data center network architecture that leverages commodity Ethernet switches to support the full aggregate bandwidth of large clusters. The authors argue that appropriately architected and interconnected commodity switches can deliver more performance at lower cost than current high-end solutions. The proposed architecture is fully backward compatible with Ethernet, IP, and TCP, and requires no modifications to end hosts, operating systems, or applications. The paper discusses current data center network topologies, including two- and three-tiered switch architectures. It highlights the problem of oversubscription, where the total available bandwidth is limited by the bandwidth of the communication hierarchy's root. Multi-path routing techniques like ECMP are discussed, but they have limitations in terms of path diversity and routing table size. The authors propose a fat-tree architecture, which is a type of Clos network, to achieve full bisection bandwidth for large clusters. This architecture uses commodity switches and allows for efficient traffic distribution. The paper describes the use of two-level routing tables to enable multi-path routing and efficient traffic diffusion. It also discusses flow classification and scheduling techniques to manage traffic and avoid packet reordering. The paper also addresses fault tolerance, power, and heat dissipation issues in data center networks. It presents a prototype implementation using NetFPGAs and Click, and evaluates the performance of the proposed architecture. The results show that the proposed architecture achieves higher bandwidth and lower power consumption compared to traditional data center designs. The architecture is scalable, cost-effective, and compatible with existing Ethernet and IP standards.This paper presents a scalable, commodity data center network architecture that leverages commodity Ethernet switches to support the full aggregate bandwidth of large clusters. The authors argue that appropriately architected and interconnected commodity switches can deliver more performance at lower cost than current high-end solutions. The proposed architecture is fully backward compatible with Ethernet, IP, and TCP, and requires no modifications to end hosts, operating systems, or applications. The paper discusses current data center network topologies, including two- and three-tiered switch architectures. It highlights the problem of oversubscription, where the total available bandwidth is limited by the bandwidth of the communication hierarchy's root. Multi-path routing techniques like ECMP are discussed, but they have limitations in terms of path diversity and routing table size. The authors propose a fat-tree architecture, which is a type of Clos network, to achieve full bisection bandwidth for large clusters. This architecture uses commodity switches and allows for efficient traffic distribution. The paper describes the use of two-level routing tables to enable multi-path routing and efficient traffic diffusion. It also discusses flow classification and scheduling techniques to manage traffic and avoid packet reordering. The paper also addresses fault tolerance, power, and heat dissipation issues in data center networks. It presents a prototype implementation using NetFPGAs and Click, and evaluates the performance of the proposed architecture. The results show that the proposed architecture achieves higher bandwidth and lower power consumption compared to traditional data center designs. The architecture is scalable, cost-effective, and compatible with existing Ethernet and IP standards.