Homa: The End of TCP for AI Clusters — John Ousterhout, Stanford

This post introduces Homa, a novel transport protocol designed for AI clusters. It contrasts Homa's message-based, receiver-driven approach with TCP's limitations for AI workloads. Key technical contributions include: 1. Message-based communication: Homa treats data as discrete messages, enabling prioritization. 2. Receiver-driven congestion control: The receiver manages flow by issuing grants, preventing congestion at the sender. 3. SRPT (Shortest Remaining Processing Time) prioritization: Homa uses SRPT to prioritize short messages. 4. Leverage of switch hardware queues: Homa utilizes multiple hardware queues in switches to bypass long-queued traffic with low-latency messages. Performance benchmarks show over 10x reduction in tail latency for short messages compared to TCP.