Skip to content

L7 · Network Fabric#

conceptstub

What it is#

The physical network that connects nodes: InfiniBand, RoCE v2 over Ethernet, Ultra Ethernet, AWS EFA or HPE Slingshot, with its switches, topology (leaf-spine, rail-optimized) and congestion control.

Why it matters#

Once a job spans nodes, the slowest link sets the pace of every collective. Bandwidth, topology and congestion behaviour (PFC, ECN, adaptive routing) decide whether hundreds of GPUs stay busy or wait on the network.

In this module#

Physical fabrics and how pods get attached to them.

Planned content

  • InfiniBand, RoCE v2, EFA, Slingshot
  • Kubernetes networking: primary CNI vs secondary networks
  • Congestion control for collective traffic