1. Fat-tree

1.1. Scalability Problem in Data Centers

• If every host were connected to every other host (full mesh), the number of links would grow as $𝑂\left({𝑁}^2\right)$
• For thousands or millions of servers, this is physically impossible: too many cables, too many switch ports.
• Fat-Tree scales to tens of thousands of hosts while keeping the number of links per switch = $𝑘$ (bounded, manageable).

1.2. Reusing Commodity Switches

• Around 2005, large data centers needed huge bandwidth but high-radix switches (with hundreds of ports) didn't exist.
• The idea: use many small cheap switches (each with $𝑘$ ports) and organize them hierarchically so the aggregate bandwidth grows with the network.

1.3. Balanced Bandwidth (“Fat”)

• In a regular tree topology, links near the root get overloaded (become bottlenecks).

• Fat-Tree "fattens" higher-level links:

  • Instead of one uplink, each switch connects to multiple upper-layer switches.
  • This ensures bisection bandwidth (the minimum capacity between any two halves of the network) grows with the network size.
  • Effect: You can push a lot of traffic across the middle of the network without congestion.

References

1. A Scalable, Commodity Data Center Network Architecture