A Stanford University computer science professor has introduced a new network protocol called Homa designed to address limitations of the Transmission Control Protocol (TCP) in artificial intelligence workloads. John Ousterhout, professor emeritus at Stanford, presented the protocol during a recent talk at the AI Engineer World's Fair, stating that TCP was not optimally suited for modern datacenter environments handling AI computations.

Homa, originally developed as a PhD dissertation by Behnam Montazeri in 2019, operates differently from TCP by using a message-based approach rather than a stream-based model. In this design, the receiver manages congestion control by receiving explicit information about incoming data volume in the first packet, allowing it to schedule subsequent transmissions. This method prioritizes shorter messages using a shortest-remaining-processing-time algorithm, reducing their latency by up to 13 times compared to TCP under similar network conditions.

According to Ousterhout, Homa achieves a 99th percentile latency of 92 microseconds for short messages in a 100 Gbps network at 80% utilization, compared to 1.2 milliseconds for TCP. Even for larger messages, Homa demonstrates approximately double the performance. The protocol can be integrated into Linux kernels on both client and server systems through a straightforward installation process without requiring system reboots.

Ousterhout emphasized that Homa does not replace TCP but can coexist with it, enabling a gradual transition for applications. He is currently working on an IETF standardization document and has facilitated the protocol’s backporting to Red Hat Enterprise Linux versions 8 and 9.5. Additionally, he is collaborating with a major financial services firm on a prototype implementation.

The development of Homa comes amid broader industry efforts to optimize network performance for latency-sensitive applications. Other approaches, such as DPDK in database systems, NVMe over Fabrics in storage networks, and QUIC in web services, have similarly sought to circumvent TCP’s limitations. High-performance computing communities, including those in artificial intelligence and financial trading, have also highlighted the impact of network latency on system efficiency.

Criticism of Homa’s claims has emerged from network architect Ivan Pepelnjak, who questioned the protocol’s performance assertions and suggested it may address a problem that is not universally prioritized. Despite these concerns, Ousterhout maintains that Homa offers a meaningful improvement for specific workloads where millisecond-level latency reductions can significantly affect computational efficiency, particularly in GPU-intensive AI training environments.

The protocol remains under active development, with ongoing efforts to integrate it into mainstream networking stacks and expand its adoption across industries.