Leaf-spine architecture

In short: The leaf-spine architecture is a network design for data centres with two layers of switches: every “leaf” switch is connected to every “spine” switch.

In more detail: The servers attach to the leaf switches. The spine switches (backbone) connect the leaves to one another. Leaves are never connected directly to each other, nor are spines.

In Depth

• Always two hops: between two servers on different leaves the path runs over exactly one spine (leaf → spine → leaf). Latency is even and low.
• Many paths: with 4 spines there are 4 equal paths. Traffic is distributed and a failure only reduces capacity (redundancy).
• Growth: more servers need more leaves, more bandwidth needs more spines (scaling).
• Why: the old three-tier design (access, distribution, core) was a bottleneck for server-to-server (“east-west”) traffic. Modern applications and cloud services generate a lot of it.
• Basis: it is the most common form of a network fabric.

See also: data centre, switch, full mesh topology