Hard
Allocating on one NUMA node, working from another
On a dual-socket NUMA system, a setup thread pinned to socket 1 allocates and zero-initializes a large array (the initializing writes are its 'first touch'). Afterward, the array is handed off to a worker thread pinned to socket 0, which repeatedly reads and writes it in a hot loop, and the array is never touched again by anything on socket 1. What determines the memory latency the worker experiences, and does it improve on its own over time?
19 of 53 in Operating Systems