Stanford professor champions Homa protocol for faster AI datacentre networks
A Stanford professor emeritus is promoting Homa, a new network protocol designed to replace TCP for artificial intelligence datacentres. John Ousterhout argues that TCP, the foundational protocol underlying the internet and cloud computing, is poorly suited to modern AI workloads and has made championing Homa his life's mission since retiring from teaching.
Homa is message-based rather than stream-based, with receivers instead of senders managing congestion control using an algorithm that prioritises shorter messages. Performance benchmarks show Homa delivering latency of 92 microseconds versus TCP's 1.2 milliseconds at the 99th percentile for shorter messages—13 times faster—and roughly twice as fast for longer messages on 100 Gbps networks at 80 per cent utilisation. Ousterhout is currently drafting IETF standardisation documentation and working to upstream Homa into the Linux kernel, with Red Hat Enterprise Linux versions 8 and 9.5 already receiving backports.
- Stanford professor promotes Homa, a new protocol to replace TCP for AI workloads
- Homa delivers 13x faster latency for shorter messages, can coexist with TCP
- Standardisation efforts underway with some industry scepticism from critics
New here? Start with this
The internet and cloud computing rely on a communication system called TCP, created decades ago. It has worked well for many applications, but computer scientists now argue it is poorly suited to modern artificial intelligence systems that must process vast amounts of data at extreme speed.
Homa is an alternative communication system created specifically for AI datacentres. Instead of the sending computer managing data flow, Homa lets the receiving computer take control. This approach proves significantly faster, especially for the short messages that AI systems frequently send.
A Stanford professor named John Ousterhout is championing this new system. His efforts include getting Homa recognised as an official internet standard and integrating it into Linux, the operating system that runs much of today's cloud infrastructure. If widely adopted, Homa could meaningfully improve how quickly AI datacentres can operate.
Read the full article at the source →
Originally published by The Register as “Stanford prof is beating the drum for a new protocol to replace TCP”.