In my testing, it pairs more naturally with MPI, as each MPI process can be bound to a particular node (-N x) and then have all memory allocations be local (-l); on the Nehalem — and likely any NUMA — systems this can be a big win by ensuring processes stay close to their memory. As MPI has much more explicit ‘my’ memory vs. ‘your’ memory than OpenMP, this type of binding is easy to implement.