Barre: Empowering Simplified and Versatile Programmable Congestion Control in...
summary
Barre is a practical PCC-style congestion controller that runs on BlueField-3 SmartNICs, uses RTD probe packets plus high-frequency CMP signals, and implements adaptive AMD increases, a dual-lock limiter, and an inflight monitor to bound packet bloat. Deployed on a 400 Gbps RoCE cluster for over a year supporting 10,000 GPUs, Barre improves AI training throughput by about 9.6%, reduces queue length up to 19%, handles RT probe loss up to 8.9%, and adapts DCQCN parameters for a 19% normalized throughput gain.
AI-written summary · about
chapters (9)
- 0:00Barre introduction
- 1:25Modern congestion control challenges
- 3:13Practical design principles
- 6:26Barre’s algorithm
- 11:12Barre on BlueField-3
- 13:06Evaluation vs DCQCN & InfiniBand
- 14:33Production results
- 15:16Optimizing DCQCN with Barre
- 16:20Conclusion
related talks
17mWhite-Boxing RDMA with Packet-Granular Software Control
17mTowards provably performant congestion control
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
16mBolt: Sub-RTT Congestion Control for Ultra-Low Latency
23mACK to the Future: TCP in 2025
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches
22mLIVE Q&A #3
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
