advanced · multicore
When Does NUMA Start to Matter?
What is NUMA, and at what point does it start to affect application performance?
Practical scenario
A batch job scales well to 16 threads on a laptop and poorly to 64 on a large two-socket server. Someone proposes rewriting the parallel decomposition.
What it tests
- • Understanding that memory access cost depends on placement
- • Knowledge of first-touch allocation
- • Whether the candidate knows when it can be safely ignored