Stanford Professor Emeritus Proposes Homa Protocol To Replace TCP In AI Data Centers
TCP's connection state, byte-stream ordering and slow congestion inference are exactly the costs that mount in clusters where large model transfers and short KV cache lookups share one network, so a message-oriented protocol that prioritizes short messages is a measure of where AI data center networking is heading.
Reporting from 1 source: GIGAZINE.
John Ousterhout, a Stanford University professor emeritus, is advocating for Homa, a transport protocol he argues is better suited to AI-oriented data centers than TCP. Homa processes short messages preferentially to hold down latency, and it can run alongside existing TCP. Its research traces to a 2018 paper by Behnam Montazeri, Ousterhout and others, and it assumes RPC-style communication.
TCP manages state per connection and treats data as one continuous byte stream, which is what makes it reliable on the open internet and costly inside a data center. There, many nearby machines exchange large numbers of small requests, and the bookkeeping plus transmission-order guarantees eat a large share of communication time.
Homa takes the opposite approach. It communicates in messages rather than streams, assumes remote procedure calls, and delivers short messages ahead of long ones. Because it can be introduced alongside TCP, the argument is about which traffic gets which protocol, not about turning TCP off.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.