NUMA balancing
The technique of placing a thread and its data near each other in a NUMA system (Non-Uniform Memory Access), where a CPU reaches its "own" local memory faster than memory attached to another processor/socket. Wrong placement causes unnecessary latency.
Background: in a server with several CPU sockets (or a chiplet CPU) the memory is divided so each processor has "its" memory near it. Reading the local memory is fast; reading another socket's memory (over the interconnect) is slower — that's what "non-uniform" means. NUMA balancing ensures a thread runs on a core near the memory it uses: the OS can move either the thread to the data or the data to the thread. Without it, a thread can constantly fetch over the slow link → a performance loss on memory-intensive loads. Important in large servers, virtualized environments, and even modern chiplet CPUs (AMD EPYC/Ryzen where different CCDs have different memory distances). Linux has automatic NUMA balancing. A central optimization for memory-intensive multithreaded performance. Related to memory controller (IMC) and cache hierarchy.