Integrating gRPC with the Homa Transport Protocol - John Ousterhout, Stanford University
summary
Homa is integrated into gRPC by a kernel driver plus a user library so C++ apps switch transports with a one-line code change (channel credentials on client and server); Java is in progress, Go and security are not done. Homa shows 10–100× lower tail latency for small messages in prior papers and in experiments here yields 73 µs vs 116 µs for a short RPC (≈40% faster) and 7–83× gains across workloads. Integration is tested on a 40-node cluster but requires a C++ only, kernel module required setup; gRPC adds a 4–5× overhead versus raw sockets and its internal complexity slows optimization.
AI-written summary · about
chapters (10)
- 0:00gRPC and Homa integration
- 2:28TCP struggles in data centers
- 9:16Homa transport design
- 13:48Homa vs TCP performance
- 15:55Homa for widespread use
- 17:58Homa with gRPC integration
- 21:49gRPC complexity challenges
- 25:43gRPC performance overhead
- 29:16Future of gRPC and Homa
- 30:00Q&A
related talks
24mEnhancing Homa Linux for Efficient RPC Transportation
15mA Linux Kernel Implementation of the Homa Transport Protocol
46mIntroduction to gRPC
37mgRPC Intro
33mIntro to gRPC
24mHow We Built Protosockets to Go Beyond HTTP and gRPC
24mHow Low Can You Go? A Comparative Dive Into gRPC and TtRPC for...
30mBringing HTTP/3 To gRPC at Cloudflare Scale