Exploiting processor groups to extend scalability of the GA shared memory programming model