This chapter owns three concepts that every later chapter builds on:
The AllReduce barrier -- the synchronisation point that makes every GPU dependent on the worst network path in the cluster. JCT = slowest participant, not the average. This is why a single bad switch port can cost hundreds of GPU-hours per day.
Tail latency at scale -- a per-operation failure rate that is acceptable in a single-server context becomes a near-certain stall at cluster scale. The math requires zero packet loss, not low packet loss.
Why TCP fails -- TCP's loss recovery mechanism (congestion window backoff + retransmit) introduces milliseconds of latency per event. AllReduce barriers cannot tolerate this. RDMA with RoCEv2 replaces TCP, but requires the fabric to be lossless. Achieving lossless Ethernet is the subject of Chapters 5-6.