Stanford professor pushes Homa protocol to replace TCP for AI workloads

Stanford professor emeritus John Ousterhout is promoting Homa, a message-based protocol meant to replace TCP in AI datacenters. Homa shifts congestion control to the receiver and prioritizes short messages: p99 latency for short messages is 92 microseconds versus 1.2 milliseconds for TCP, a 13x improvement. The protocol has been backported to RHEL 8 and 9.5, with an IETF standardization draft and Linux kernel upstreaming underway.

• Homa p99 latency for short messages is 92 microseconds vs 1.2 milliseconds for TCP
• Homa runs alongside TCP, allowing gradual application migration
• Protocol backported to Red Hat Enterprise Linux 8 and 9.5
• IETF standardization draft and Linux kernel upstreaming are in progress