What it is
On a NUMA machine memory is attached to a node, and a CPU reaches its own node's memory faster than another's. With the default policy a page is placed on the node of the thread that first touches it, so a program that initialises its data on one thread and then works on every node does most of its accesses remotely, and the init node's memory serves everything. Interleaving spreads the load; initialising each thread's part from that thread keeps accesses local; automatic NUMA balancing migrates pages after the fact. For Kubernetes, the Topology Manager decides whether exclusive CPUs and memory are aligned to one NUMA node.
When to use it
Use it to explain a latency or bandwidth problem on a multi-socket host, to read numactl -H and numastat, to decide between interleave and per-thread initialisation, or to check whether a pod's CPU request can be aligned to one NUMA node.
Gotchas & limits
The placement model assumes a symmetric machine and uniform access inside each thread's slice. Relative latency weights accesses by node distance and is not a time. Automatic balancing is modelled only for partitioned data and is not verified on a multi-node reference machine yet. Topology Manager behaviour depends on the Kubernetes version and is to be verified for yours.