A Journey of Accelerating and Offloading OVS with a DPU#

Abdullah Sevincer – Crusoe, Alex Bortok – NVIDIA

• Why: A numbers-backed production account of moving SDN (OVS/OVN) from host kernel to hardware to DPU — datapath internals with real benchmarks, at AI-cloud scale. Squarely the PM's cloud-native/datapath interest.
• What: Three stages: kernel OVS (connection rates topping out in the low thousands/sec, instability growing with load); hardware-accelerated OVS datapath on the host (order-of-magnitude higher connection rate, near-zero failures, but host CPU cores and memory reserved away from customer VMs); then offloading the OVS/OVN control plane itself onto BlueField-3 DPUs, freeing 100% of host CPU — with the lesson that DPU cores aren't host cores, so right-sizing thread count beat adding more.
• How: Slow-path mechanics explained cleanly: PMD threads busy-polling pinned cores instead of IRQ→softirq→netlink upcalls; steering rules programmed by DMA'd descriptor queues the NIC drains itself instead of per-flow netlink/TC calls; per-core parallel connection tracking instead of shared kernel conntrack locks; VF representors programming the eSwitch. Benchmarks via Keysight CyPerf (HTTP over Geneve: overlay, stateful NAT patterns). OVN programs both Geneve east-west and north-south for customer VPCs.
• Where applicable: Crusoe's AI factory (BlueField-3/OVS-DOCA on every node) — co-presented with NVIDIA, so expect product framing; the performance mechanisms and tradeoffs apply to any OVS/OVN fleet, with numbers specific to their hardware and workloads.