LIVE Q&A #3
summary
Team uses eBPF to tune transport in-kernel without rewriting TCP, packaging programs as a single binary golden image for fast rollout and minute-scale rollback. A centralized controller aggregates per-host telemetry then sets one service-wide threshold derived from BDP and flow count — ECN threshold tuned to BDP — to balance DCTCP and Cubic across shared queues. RX offload coalescing with ECN-marked packets increases CPU, and using an SK local map trades memory for fewer callbacks.
AI-written summary · about
chapters (10)
- 0:00Live Q&A
- 1:50Kernel vs user-space transport
- 5:21DCTCP with hardware offloads
- 6:48Mixed data center traffic
- 8:50Dynamic window tuning control
- 12:24BPF deployment challenges
- 14:57BPF policy conflicts
- 17:24BPF tuning overhead
- 19:57BPF future vs kernel
- 22:01Closing remarks
related talks
28mSafe sharing of the network with eBPF
32meBPF: BPF kernel infrastructure
23mLive Q&A with Speakers – James Zeng
30mA Beginner's Guide to eBPF Programming for Networking
14mA large-scale deployment of DCTCP
16mElectrode: Accelerating Distributed Protocols with eBPF
23mACK to the Future: TCP in 2025
15meTran: Extensible Kernel Transport with eBPF
