Enhancing Homa Linux for Efficient RPC Transportation
summary
Homa gets a dynamic per-peer window: an RTT probe every 1 ms measures real-time RTT and classifies RTT into low/mid/high to shrink or grow unscheduled and scheduled windows, avoiding the static RTT_bytes tuning. Tests on 25Gb and 100Gb links show 5–18% throughput gain and improved P99 latency for large messages, and optimized Homa survives a six-client incast without buffer overrun. gRPC work reuses IDs to implement stream RPCs without creating seven separate Homa RPCs, reducing creation/reclaim overhead.
AI-written summary · about
chapters (10)
- 0:00Introduction to Homa
- 1:26Homa protocol overview
- 5:39Limitations of Homa
- 9:53Enhancing congestion control
- 13:49Performance test results
- 16:38Testing incast handling
- 17:45Coexistence with TCP
- 18:51Stream RPC improvements
- 20:19Future improvements
- 21:01Conclusion and Q&A
related talks
15mA Linux Kernel Implementation of the Homa Transport Protocol
39mIntegrating gRPC with the Homa Transport Protocol - John Ousterhout, Stanford University
13mRakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs
17mBackdraft: a Lossless Virtual Switch that Prevents the Slow Receiver Problem
35mWhy OCP networking performances on VMs are suboptimal and how to improve them by 5x
18mHydraRPC: RPC in the CXL Era
15mHostping: Diagnosing Intra-host Network Bottlenecks in RDMA Servers
24mHow We Built Protosockets to Go Beyond HTTP and gRPC