Network Traffic Quality of Service (QoS) Classes for Containers & Pods Via EBPF/XDP
summary
Mizar CNI implements pod-level QoS using eBPF/XDP and Geneve overlays, classifying traffic into premium / expedited / best-effort via an on-node transit daemon that programs eBPF maps. XDP redirects or passes packets, sets DSCP for end-to-end recognition, and enforces EDT timestamp-based rate limiting; a dynamic per-second bandwidth monitor clamps best-effort when premium appears. Demo shows premium taking ~1.4 Gbps while best-effort falls to ~100 Mbps (~10%), with zero kernel code changes; prototype, working but needing TX hook and API polish.
AI-written summary · about
chapters (11)
- 0:00Introductions and agenda
- 2:44Defining network QoS
- 4:48eBPF and XDP overview
- 7:00Mizar architecture
- 7:53QoS design and implementation
- 19:39Traffic classification API
- 20:41Demo setup
- 22:41Live demo of QoS performance
- 25:47Demo results summary
- 27:26Future work and conclusions
- 29:31Q&A session
related talks
07mLightning Talk: Taming Noisy Network Neigbours with Cilium and eBPF
22mLIVE Q&A #3
33mBringing packet queueing to XDP
37mEfficient and Flexible Virtual Machine Networking Through eBPF
29mSimplifying the Networking and Secur...
24mIntroduction to eBPF Qdisc and Use Cases
22mRole of eBPF in Kubernetes Networking and Security
32mExtending Network Quality of Service support in OpenStack