How do racks become a cluster?
A cluster's topology is how its servers and switches connect. The shared links determine how much traffic can move between groups.
The shared uplinks set a boundary
- 16 server links × 100 Gb/s
1,600 Gb/s of downstream capacity
- Leaf switch
Local traffic may stay here
- 8 uplinks × 100 Gb/s
800 Gb/s for traffic leaving the leaf
- Spine layer
Connects onward to other leaves
Picture several roads joining
Servers may connect to leaf switches, which connect to spine switches. Not every server-facing link has an independent full-speed path through the next layer.
In an invented example, 16 links at 100 Gb/s feed eight uplinks at 100 Gb/s. That is 1,600 Gb/s downstream and 800 Gb/s upstream: 2:1 oversubscription for traffic leaving the leaf.
Local traffic may not cross that boundary. The ratio does not predict every job’s slowdown.
Surviving is not the same as performing
A failed uplink may leave every server reachable but reduce available bandwidth. Test the degraded condition against the workload’s requirements.
Placement matters too. A job spread over constrained paths may behave differently from one kept local.
Try it yourself
Use 24 downstream and 12 upstream links, all 200 Gb/s. What happens when six uplinks remain?
Check your answer
The ratio changes from 2:1 to 4:1. That is a shared-capacity calculation, not proof that an application becomes twice as slow.