Safe sharing of the network with eBPF
summary
Meta uses eBPF to push concrete transport fixes: a receiver-driven rwnd rewrite at TC that cuts packet discards, lowers buffers and packets-in-flight; a struct_ops-based CC that samples one‑way delay plus ECN to limit queuing; pacing and max-rate caps that halve cross-region P95/P99 latencies; deployment uses a control plane managing 10+ tunings, tcp_iter to update live connections and pinned maps with a BPF versioning manager to handle restarts and map schema changes.
AI-written summary · about
chapters (11)
- 0:00Introduction to eBPF network sharing
- 1:06Microbursts in data center networks
- 6:22TCP flow control with eBPF
- 9:13Network tuning with eBPF
- 11:00Motivation for custom congestion control
- 13:06Building delay-based congestion control
- 15:34eBPF control plane at Meta
- 18:24Deploying custom congestion control
- 21:12Handling long-lived connections
- 23:26Control plane resilience
- 26:39Future plans and open source
related talks
22mLIVE Q&A #3
32mOverview of the BPF networking hooks and user experience in Meta
32mMPTCP: Extending kernel functionality with eBPF and Netlink
38mPacket Metadata - Where Are Thee?
19mTackling DC Congestion and Bursts
32meBPF: BPF kernel infrastructure
30mA Beginner's Guide to eBPF Programming for Networking
39mThe evolution of service mesh with eBPF