Speedrun Through Splicing Sockets with Sockmap
summary
Shows how kernel-side socket splicing using eBPF actually cuts RPC latency from about 5.8 μs → 4.7 μs by bypassing the full network stack, with working examples and code on GitHub. Explains required pieces: a BPF map of socket references, eBPF SK_MSG/SK_SKB programs attached to the right hooks, and how to auto-populate maps via cgroup so_ops. Covers supported socket kinds (TCP, vSOCK), four redirect modes, limitations, and production uses in Cilium transparent L7 filtering and a TikTok proxy bridge.
AI-written summary · about
chapters (9)
- 0:00Welcome
- 1:47Sockmap for container networking
- 6:02Live demo: latency improvement
- 6:24Sockmap as BPF map
- 7:27Setting up Sockmap
- 11:04Inserting sockets
- 13:58Four redirect scenarios
- 19:43Real-world use cases
- 23:19Learn more
related talks
47mSpeedrun through Splicing Sockets with Sockmap
32mOverview of the BPF networking hooks and user experience in Meta
30mA Beginner's Guide to eBPF Programming for Networking
37mEfficient and Flexible Virtual Machine Networking Through eBPF
32meBPF: BPF kernel infrastructure
26mContainer Networking: The Play of BPF and Network NS with different...
23mSYN Proxy at Scale with BPF
01h 21mTutorial: Developing High-performance Networking Applications