https://hgpu.org/?p=8104
Extending MPI to Accelerators