Stanford prof is beating the drum for a new protocol to replace TCP

John Ousterhout, a retired Stanford University professor of computer science, is advocating for a new data transport protocol called Homa to replace the widely used TCP protocol. Ousterhout argues that TCP, despite its historical significance, is not an ideal match for modern AI workloads in datacenters. He suggests that Homa, which can be easily integrated into existing networks, offers significant improvements over TCP.

Homa operates on a message-based system rather than TCP's stream-based approach, allowing for explicit congestion control. The protocol prioritizes shorter messages using a shortest-remaining-processing-time (SRPT) algorithm, resulting in drastically reduced latency for shorter messages. Ousterhout claims that the 99th percentile (p99) latency for Homa's shorter messages is 92 microseconds, which is 13 times faster than TCP's 1.2 milliseconds p99 latency in the same network conditions.

The protocol is designed to run parallel to TCP, enabling a gradual transition from TCP to Homa for applications. With Homa, TCP applications may even experience performance improvements. Ousterhout is currently working on standardizing Homa as an IETF protocol and integrating it into the Linux kernel. He is also helping large companies evaluate Homa's potential applications.

Despite the benefits, Homa faces skepticism from some network architects and experts. Ivan Pepelnjak, a prominent network architect, has criticized Ousterhout's performance claims and questioned Homa's solution to existing problems. However, other industries, such as AI, high-performance databases, and financial services, are actively exploring Homa's potential to improve network performance in latency-sensitive workloads.