NUMA and CPU pinning for UPF

Why NUMA Affinity Matters

Modern multi-socket servers — two or four CPU sockets, each with its own attached memory — are NUMA (Non-Uniform Memory Access) systems. Each CPU has a local memory region with low-latency access; accessing memory attached to another socket goes over the inter-socket interconnect (UPI on Intel, IFOP on AMD) at substantially higher latency and lower bandwidth. For a CPU-bound workload like a UPF processing millions of packets per second, NUMA crossings are catastrophic. A worker thread on socket 0 polling a NIC on socket 1: every packet fetched from the NIC ring is a remote read; every memory…