Tuning Linux TCP for data-center networks
summary
Data-center tuning focuses on keeping queues on hosts not switches: enable ECN-aware congestion control (e.g., ECTCP or BBRv2), disable idle cwnd shrink, use TCP Fast Open and pacing to avoid million-flow bursts, and size TCP buffers for long-haul links (London–Singapore ~200 ms). Implement dynamic IPv6 flow-label load balancing (protective load balancing) to shift hotspots, and add TCP telemetry tying RPC messages to per-socket stats—measurements show network is the culprit less than one third of the time; much delay comes from queuing and CPU/copy hotspots.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:47Data center network overview
- 5:14Optimizing network throughput
- 9:05Dynamic load balancing
- 12:12WAN performance tuning
- 14:28Host CPU and memory efficiency
- 21:04TCP telemetry and profiling
- 25:52Advanced TCP features
- 28:47Future challenges and conclusion
related talks
22mAdapting TCP for Reconfigurable Datacenter Networks
26mTuning Your Cloud: Improve Global Network Performance for Applications
23mACK to the Future: TCP in 2025
26mTuning Your Cloud: Improve Global Network Performance for Application
19mTackling DC Congestion and Bursts
14mA large-scale deployment of DCTCP
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
22mLIVE Q&A #3
