A Linux Kernel Implementation of the Homa Transport Protocol

Explore a conference talk from USENIX ATC '21 detailing the Linux kernel implementation of the Homa transport protocol. Dive into John Ousterhout's presentation from Stanford University, which showcases Homa/Linux's superior performance compared to TCP and DCTCP. Learn about the benchmark results in a 40-node cluster, where Homa/Linux demonstrated significantly lower latency across all message sizes, with short message tail latency up to 83 times lower than its competitors. Discover how Homa has effectively eliminated network congestion as a major performance bottleneck, shifting focus to software overheads and load balancing challenges. Gain insights into potential future improvements that could yield an additional 5-10x performance boost by addressing these software limitations. The talk covers key aspects such as the Homa API, protocol details, latency comparisons, workload performance, software overheads, load balancing issues, and potential future directions for transport protocols.

Intro
Takeaways
Homa API
Homa Protocol
Homa Latency TCP 40-node cluster, 80% network load
Homa Dominates: All Workloads, All Sizes
Software Overheads
Load Balancing Causes Hot Spots
2-3x Overhead for Load Balancing
Transports in the NIC?
Conclusion