Here is how to think about NUMA. In a modern AMD epyc box, you have 2 CPUs. Each cpu has 350GB/s of memory bandwidth and has PCI devices connected to it. Memory speed between CPU sockets is some fraction, 70GB/s?
if you run computation on cpu0 and talk to nic on cpu1, you burn a lot more cpu cycles, than if you move your computation to cpu1.
So in theory, if you partition the box using containers such that everything on cpu0 runs in container0 and has own nics, and same thing on cpu1. You end up with 2 'virtual' boxes that might actually perform better within your container than outside of it.
Note on modern CPUs, each cpu is further broken down into numa nodes(numactl -H and numa stuff n lscpu)...perf degradation isn't as great going between chiplets, but it's measurable(2x reduction in ram bandwidth?).