Homa: The end of TCP for AI clusters [video]

<cite index="11-2,14-2">Stanford professor emeritus John Ousterhout used a talk at the AI Engineer World's Fair to argue that TCP suits datacenters poorly, and he is now promoting Homa, a protocol he describes as his life's mission.</cite> <cite index="14-1">Homa began as a PhD dissertation by Behnam Montazeri, now a Google staff engineer.</cite> <cite index="6-2">In the benchmark presented, Homa cut P99 latency for short messages roughly 13x versus TCP and about halved latency for the longest messages.</cite> <cite index="6-6">It controls congestion from the receiver and uses switch priority queues.</cite> <cite index="11-5,11-6">Homa installs as a Linux kernel module built from GitHub source, with no reboot required.</cite> <cite index="11-8">Network architect Ivan Pepelnjak has called it a solution looking for a problem.</cite> This matters mainly for teams running their own clusters. <cite index="9-4">Agentic workflows and fine-grained coordination emit floods of small messages whose tail latency dictates end-to-end throughput.</cite> <cite index="11-10">A millisecond of latency can leave an expensive GPU idle.</cite> Builders on hosted APIs are unaffected. <cite index="2-6,2-7">Equivalent gains for AI applications, and comparisons with RoCE, still need separate measurement.</cite> Profile your short-message share before changing anything.