Multi theaded program sharing variables on NUMA system

Question

I am doing some experiment with NUMA systems.Where in I have a multithread/2-threads c program sharing one cache line (int64) variable between them. When I run both threads on the same node the program takes nearly 50% more time to finish than if I run/bind these threads on two different NUMA nodes. I was thinking it should be other way round because they share the data, they should finish faster if they run on the same node.

Am I missing something here?

  pthread_mutex_lock (&shared_mutex);
  if (thinfo->thread_num == 1)
    shared_var++;
  else
    shared_var--;
  pthread_mutex_unlock (&shared_mutex);