I need to capture multicast traffic inside Kubernetes clusters and am comparing hostNetwork, macvlan, and ipvlan. Are there other networking approaches that can provide near host-level latency? I'm also concerned about a shared platform where users can launch workloads dynamically, since assigning hostNetwork to every multicast-capable Pod may not be practical or desirable.
2 Answers
hostNetwork is usually the simplest option and gives you the closest behavior to the node’s network stack, with minimal added latency. A host-device setup through the CNI can also provide direct access to a physical interface, but it generally assigns that device to only one Pod at a time, so it doesn’t scale well for arbitrary user workloads.
Take a look at Multus with a secondary network interface. In our setup, Kube-OVN provides the network and Multus attaches it as an additional CNI interface, allowing multicast-capable workloads to use a dedicated interface without putting the entire Pod on the host network.
That’s the main issue for us. We’re building a platform where users can start workloads dynamically, so we can’t reliably require or allocate hostNetwork whenever a workload needs multicast.