Cluster Optimization: Harnessing DPUs for Ambient Mesh Workloads
summary
Prototype offloads Ambient Service Mesh data plane onto an OCTEON10 DPU, moving CNI, zTunnel and Waypoint processing off servers; engineers run the model on a virtual cluster and plan an open source POC. Measured DPU line processing reaches 50 Gbps in <15W and current node costs show zTunnel at 0.35–0.85 CPU and Waypoint at 0.69–0.74 CPU depending on packet size. Remaining work is integrating hardware accelerators and maturing the software toolchain.
AI-written summary · about
chapters (12)
- 0:00Cloud-optimized silicon intro
- 1:05What is a DPU
- 3:58Service mesh importance
- 5:48Ambient mesh overview
- 7:11CPU use in ambient mesh
- 9:15Offload CNI to DPUs with Cilium
- 11:27Ambient mesh offload proposal
- 12:51Full offload architecture
- 13:33Key takeaways
- 15:28Current state and future
- 16:52Performance comparison challenges
- 18:18Testing without DPU
related talks
06mMaximizing Programmable DPUs in the Network
05mLightning Talk: Optimizing Cluster Workloads: Cilium Envoy on DPU
39mFunction Dedicated Network Deployment Model in DPU/IPU-Like Device Acro...
30mZXDH DPU Adapter and It's Application
13mLightning Talk: Accelerating Cilium: Our Journey on Offloading Cilium To DPU
08mdevelops the Distributed UPF with NVIDIA Bluefield-3 DPU
25mDPDK performance optimization for edge computing scenario
18mSmartNIC DPU Enablement in OpenStack with OVN