Lightning Talk: Accelerating Cilium: Our Journey on Offloading Cilium To DPU
summary
Demonstrates a near-complete offload of Cilium’s primary CNI to a DPU with minimal Cilium changes, implemented as an offload CNI shim that picks VFs, maps them to pods, and signals a gRPC DPU agent on Marvell Octeon 10. eBPF dataplane and XFRM/crypto run on the DPU, endpoint and IP cache mapping keeps host pods reachable when some nodes lack DPUs, and a vendor-plug layer preserves neutrality. Open-source repo includes perf-per-watt and security-acceleration numbers.
AI-written summary · about
chapters (11)
- 0:00Offloading Cilium to DPUs
- 1:07Kubernetes-DPU integration challenges
- 2:30Offloading CNI to DPUs
- 3:54Primary CNI to DPU data plane
- 5:41Offload solution architecture
- 7:06Pod interface flow to DPU
- 8:12Patching Cilium for DPU
- 9:17Data plane benefits and offload
- 10:00LLM inference use case
- 11:22Open-source status and future
- 12:30Q&A
related talks
05mLightning Talk: Optimizing Cluster Workloads: Cilium Envoy on DPU
44mCilium - Container Security and Networking Using BPF and XDP
05mLightning Talk: Unveiling 5 Essential Insights from Cilium Implementation in a...
07mLightning Talk: Taming Noisy Network Neigbours with Cilium and eBPF
19mCluster Optimization: Harnessing DPUs for Ambient Mesh Workloads
30mZXDH DPU Adapter and It's Application
06mMaximizing Programmable DPUs in the Network
24mThe Cilium Story - Why We Created Cilium