John Ousterhout, professor emeritus at Stanford, argues that TCP is a poor fit for datacenter AI workloads and is promoting Homa, a protocol he describes as a clean-slate rethink of congestion control. According to The Register, Homa is message-based rather than stream-based, with the receiver managing congestion by scheduling packets based on how much data is incoming. It prioritizes shorter messages using a shortest-remaining-processing-time algorithm.
Ousterhout says Homa is easy to add: compile it from GitHub, install the module into the Linux kernels on clients and servers, and no reboot is required. It can run alongside TCP, allowing gradual migration, and he claims it even makes remaining TCP applications run faster. He is drafting an IETF standardization document, working to upstream Homa into the Linux kernel, and helping a large financial services company build a prototype. The protocol was backported to Red Hat Enterprise Linux 8 and 9.5 in March.
The latency gains are significant in Ousterhout's tests: at 80% utilization on a 100 Gbps network, p99 latency for short messages was 92 microseconds, about 13 times faster than TCP's 1.2 milliseconds; even long messages were twice as fast. But not everyone is convinced. Network architect Ivan Pepelnjak published a critical 2023 position paper questioning Ousterhout's performance characterizations of TCP and describing Homa as a solution looking for a problem.