A Practical Guide to NUMA Affinity in Kubernetes
NUMA effects are one of those problems that don’t show up in dashboards, but will happily show up in your p99 latency and in “why is this box slower than the identical box next to it?” Kubernetes can help—but only if you enable the right node-level managers and verify the result from inside the workload. The Problem — The “Cross-NUMA” tax On multi-socket or multi-NUMA machines, not all CPU cores are equally “close” to all memory and PCIe devices. If a workload ends up with CPUs on one NUMA node and memory (or NIC / GPU) on another, you can pay a real latency / throughput penalty. ...