1. Fat-Tree

In 1985, Charles E. Leiserson of MIT invented the Fat-Tree network. As shown in the figure below, a fat-tree network is a binary tree in which bandwidth gradually increases from the root nodes down to the leaf nodes.

In August 2008, a group of computer scientists at UC San Diego published a scalable network architecture design that adopted a topology inspired by the fat-tree topology, achieving better scalability than previous hierarchical networks.

Advantages of Fat-Tree networks:

• Load balancing
• Low latency

Disadvantages of Fat-Tree networks:

• Scale is limited by the number of ports on the core switches
• Poor fault tolerance at the lower layers
• Does not adapt well to one-to-all and all-to-all communication patterns

2. Spine-Leaf

The Spine-Leaf network is a common data center network architecture. Unlike traditional hierarchical network structures, the Spine-Leaf network achieves a balanced topology at two levels: the Spine layer and the Leaf layer. Switches at the Leaf layer connect to servers, while switches at the Spine layer are responsible for connecting the Leaf-layer switches. Every Leaf switch connects to every Spine switch, thereby providing high-bandwidth and low-latency communication paths.

Advantages of Spine-Leaf networks:

• Good horizontal scalability: network capacity can be increased by adding more Spine and Leaf switches
• High reliability: multiple paths are supported, effectively avoiding single points of failure
• Low latency: equidistant multipath connections keep data transmission paths simple and stable

Disadvantages of Spine-Leaf networks:

• High cost: every Leaf switch must connect to all Spine switches, resulting in a large number of connections
• Complex configuration: the connections between switches must be carefully planned to maintain network symmetry
• Limited adaptability: in certain application scenarios the Leaf-layer connection scheme may need to be adjusted

3. Dragonfly

The Dragonfly network is an efficient network topology proposed by John Kim and others in 2008, first applied to high-performance computing (HPC) clusters. This structure mainly resolves the tension between bandwidth and cost in traditional fat-tree architectures, making it particularly suitable for large-scale parallel computing and data-intensive tasks. Through hierarchical switches and links, the Dragonfly network divides the network into multiple groups, each containing multiple routers or switches, with groups interconnected by relatively few inter-group links.

Advantages of Dragonfly networks:

• High bandwidth utilization: effective data transmission through local and global connections
• Small network diameter: fewer hops improve data transmission speed and reduce latency
• Low cost: efficient connection schemes reduce overall cabling and hardware costs

Disadvantages of Dragonfly networks:

• High routing complexity: global connections require relatively complex routing algorithms
• Average fault tolerance: a failure may affect a fairly wide range of connections
• Unsuitable for small-scale networks: in small deployments, the Dragonfly structure may lead to wasted resources

4. Torus

Torus (the ring network) is a topology commonly seen in high-performance computing (HPC) and distributed systems, first proposed by computer scientists in the 1980s and applied to parallel computing architectures.

The Torus network is a multi-dimensional mesh structure in which boundary nodes form a closed loop through extra connections, making the connections of each node more symmetric.

Advantages of Torus networks

• Good load balancing: nodes are evenly connected, effectively spreading the communication load
• Low network latency: thanks to the closed-loop connections, the distance from each node to any other is short, so communication latency is low
• High fault tolerance: multipath connections let the network maintain communication despite a single point of failure

Disadvantages of Torus networks

• Limited scalability: in high-dimensional cases, the mesh boundary and node count grow substantially, increasing hardware complexity
• Complex cabling: especially in three dimensions and higher, cabling and node connections become difficult
• Unsuitable for broadcast communication: for one-to-all and all-to-all communication patterns, the Torus topology performs poorly and broadcast efficiency is low