Under the first-touch policy common on Linux NUMA systems (a page is placed in the memory of the node where the thread first writing it runs), what should programmers watch out for?

Answer: C. If one thread zero-initializes a large array, all of it lands on one node, and threads on other nodes end up using far memory

malloc alone assigns no physical pages; the page fault on first write places each page in the memory of the node where the writing thread runs. So if one thread does the initialization, the whole array gathers on that node, and in the parallel computation much of each thread's portion is in far memory. Initializing in parallel with the same division of work as the computation puts each portion near its thread.

• A. It is not random; the first thread to touch decides.
• B. Placement happens at first write, not at malloc.
• C. Correct. Initialization should be split among the threads that later use the data.