fattrees

Fattrees, or fat-trees, refer to a class of network topologies designed for large-scale data centers to provide scalable, high-bandwidth interconnections among servers. Rooted in Clos network principles, fat-tree fabrics use multiple parallel paths and uniform switch port counts to achieve high bisection bandwidth and non-blocking behavior under typical loads. The topology emphasizes that bandwidth grows toward the root, making the network effectively fatter at higher levels, hence the name.

In a conventional k-port fat-tree, each switch has k ports. About k/2 ports connect to servers through edge switches, and the remaining k/2 ports provide uplinks to the aggregation layer. The aggregation layer connects upward to a layer of core switches, typically arranged so that each core switch connects to a subset of pods. The result is a three-layer or multi-layer fabric with many equal-cost paths between any pair of servers. For even k, a common formula yields roughly (k^3)/4 servers supported by the fabric.

Routing and traffic management rely on equal-cost multipath routing to distribute flows across multiple available paths, improving throughput and resilience to failures. Fat-tree designs can scale by adding more pods or upgrading to higher-port switches, preserving symmetry and path diversity. They are commonly built with commodity Ethernet switches and aim to minimize oversubscription.

While fat-tree fabrics offer high bandwidth and scalability, they introduce design and management complexity, including careful cable management and load balancing to avoid bottlenecks. They influenced later data-center fabrics and are closely related to spine-leaf architectures, which adopt similar principles with modern switch ecosystems. The concept originated from research in the late 2000s, notably as a practical realization of Clos networks for data centers.