Skip to content

2026

SAIC Motor's Kubernetes-Based Multi-Tenant Networking Practice: Building a Unified Network Foundation with OVN-Kubernetes

SAIC Motor is a Chinese automobile company. When we chose Kubernetes as a unified, multi-tenant infrastructure for containers, virtual machines, and AI agents, its network had to evolve beyond basic Pod connectivity into a unified multi-tenant network foundation capable of supporting heterogeneous workloads. Through its OVN-based software-defined networking capabilities, OVN-Kubernetes provides a consistent network model for containers, KubeVirt virtual machines, and agent runtimes. This is particularly important for agents: each tenant can have a group of agents that communicate with one another, while agents belonging to different tenants remain isolated at the network layer.

Today, SAIC Motor operates two generations of Kubernetes clusters:

  1. The existing fleet, which runs OVN-Kubernetes in central mode and relies on an in-house multi-tenant networking solution built on capabilities that have since been deprecated upstream.
  2. The new AI agent fleet, which runs OVN-Kubernetes in interconnect mode and is currently being evaluated as the foundation for our next-generation multi-tenant infrastructure.

This blog post outlines our journey toward designing and building the new fleet, including the architectural decisions and capabilities we are evaluating along the way. Looking ahead, once the required feature set has been fully implemented and validated, we plan to gradually migrate workloads from the existing fleet to the new architecture.

Scaling User Defined Networks

User Defined Networks (UDNs) provide isolated networks for workloads as an alternative to a single cluster-default network. Tenants and applications can use separate address spaces and topologies, with traffic that does not mix across networks. See the UDN feature page for the model and API.

Debugging NetworkPolicies with NetObserv

Observability is a key part of any system, and Kubernetes networking is no exception. But even if you see WHAT happens in your network, you still don’t know WHY it happens. Especially if you have a number of network policies in the cluster, it is always difficult to tell why exactly a given connection is allowed or denied.

To solve this problem, OVN-Kubernetes has introduced a new observability feature together with the NetObserv project that adds explicit communication between observability and networking.

Accelerating and Offloading Kubernetes Networking: Leveraging DPUs with OVN-Kubernetes

Introduction

This blog post provides a comprehensive guide on deploying OVN-Kubernetes in an environment leveraging Data Processing Units (DPUs). This setup is crucial for high-performance networking in cloud-native applications, offloading network processing from the host CPU to the DPU, and providing better security for the networking control-plane. Additionally, OVN-Kubernetes brings in robust features like User Defined Networks (UDNs) that enable per tenant network isolation into the Kubernetes environment and integrates with the DPU solution.