eTran: Extensible Kernel Transport with eBPF
summary
eTran extends eBPF to create an extensible in-kernel transport that adds new XDP hooks and a packet Q BPF map for queuing, plus XDP egress and XDP gym to enable proactive sends and buffering; a root privileged control daemon handles complex control logic and AFXDP provides fast packet I/O. Implementations of TCP+DCTCP and Homa total ~12–24k LOC (2.6k LOC for eBPF extensions) and show up to 4.8×/1.8× higher throughput and 3.7×/7.5× lower latency versus existing kernel stacks.
AI-written summary · about
chapters (11)
- 0:00eTran introduction
- 1:41Extending kernel transport challenges
- 3:27eBPF for safe kernel extensibility
- 5:20eTran split transport design
- 7:47eTran architecture overview
- 9:11Programming eBPF hooks
- 10:36Multi-tenancy traffic management
- 11:39Implementing DCTCP and Homa
- 12:23Evaluation vs Linux TCP and Homa
- 13:48Performance results
- 14:10Summary
