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

If you run AI workloads, your network latency for short messages could drop by 13x.

A retired Stanford University professor, John Ousterhout, is advocating for a new internet protocol called Homa to replace TCP, the current standard for web and cloud computing. Ousterhout argues that TCP's design, based on byte streams and its approach to congestion control, is ill-suited for the low-latency demands of modern AI workloads. Homa, which is message-based and allows receivers to manage congestion control, prioritizes shorter messages using a shortest-remaining-processing-time algorithm. This results in significantly lower latency for short messages, with a 99th percentile latency of 92 microseconds compared to TCP's 1.2 milliseconds in certain network conditions. Ousterhout is working to standardize Homa and integrate it into the Linux kernel, with initial backporting to Red Hat Enterprise Linux versions 8 and 9.5 already completed. He is also collaborating with large companies to explore Homa's practical applications.

Homa's message-based architecture and receiver-driven congestion control represent a fundamental shift from TCP's stream-based, sender-regulated model. This change directly addresses the critical latency requirements of AI applications, where even millisecond delays can lead to expensive GPU idling. By prioritizing short, time-sensitive messages, Homa aims to optimize network performance for AI-specific traffic patterns, a capability that TCP, designed for more general-purpose data transfer, struggles to provide efficiently.

The successful adoption of Homa could signal a broader trend towards specialized network protocols tailored for emerging AI infrastructure. While TCP has been a foundational element of the internet for decades, its limitations in high-performance computing environments are becoming increasingly apparent. The development and potential standardization of Homa, alongside existing efforts like QUIC and DPDK, indicate a growing industry recognition of the need for network innovations that can keep pace with the demands of advanced computing and artificial intelligence.