Revisiting Congestion Control for Lossless Ethernet
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Revisits lossless Ethernet congestion control and proposes ACK-driven congestion control (ACC) that uses receiver ACK time series and packet conservation to compute pipe capacity and the exact number of excessive packets, then temporarily halts congested flows with a new shun state to drain queues. Implements ACC in SoftRoCE, evaluates on a 5-host 10GbE testbed and 20-rack fat‑tree simulations, shows no deadlocks in 50 random runs, cuts average small-flow FCT by 1.3–3.3× and 99th-percentile by 1.4–11.5× while improving latency and throughput.
AI-written summary · about
chapters (10)
- 0:00Congestion control intro
- 1:04PFC and existing controls
- 4:18Decoupling detection and control
- 5:24Packet conservation in lossless Ethernet
- 8:54ACT-driven paradigm
- 10:41Design principles
- 11:28Introducing ACC and rules
- 14:40ACC implementation
- 15:24Evaluation and comparison
- 17:36Conclusion and future
related talks
17mRe architecting Congestion Management in Lossless Ethernet
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
23mACK to the Future: TCP in 2025
16mFRCC: Towards Provably Fair and Robust Congestion Control
19mTackling DC Congestion and Bursts
22mAdapting TCP for Reconfigurable Datacenter Networks
21mAccelTCP: Accelerating Network Applications with Stateful TCP Offloading
16mBackpressure Flow Control
