Global arrays is normally used over MPI anyhow. I guess there's SHMEM, but that's integrated with at least OpenMPI (and others, I think). CHARM++ has been used at scale, but it's semi-proprietary.