Teaching_Code

NUMA

This code snippet should demonstrate the problem that can arise with NUMA architectures. Likwid can be used to pin the program and ensure the expected results. Be sure to adjust the core ids to the specific architecture the program is running on

---

make
likwid-pin -c 0,10 ./numa

If you run the numa program and pin it to different cores it should print out different runtimes for the two threads. This results in the fact that both vectors are allocated on one thread.

possible output:

Thread: 1 on CPU: 10; Time (ms): 21581
Thread: 0 on CPU: 0; Time (ms): 34825

---

make
likwid-pin -c 0,1 ./numa

However if you pin the program on the same socket you should see the same runtimes

possible output:

Thread: 1 on CPU: 1; Time (ms): 21617
Thread: 0 on CPU: 0; Time (ms): 21675

---

numa_fixed.cpp shows one possible solution to the NUMA problem. By resizing the vector it will be newly allocated and is therefore located in the domain of the corresponding thread