cuda - Do I really need MPS when running multiple MPI ranks on a single GPU, or Kepler's Hyper-Q itself is enough? -

Actually I would like to run multiple MPI ranks on a single GPU (NVidia K20c), and I know the existence And Kepler

However, my question hyper-q is enough for my need? Or would I have to use MPS? According to the hyper-Q link above, "no additional coding effort is required to enable Hyper-Q. It takes all 20 GUUs of a Tesla, which has the CUDA5 installation and to determine the variables of the environment Many MPI ranks are ready to use GPU-Hyper-Q. "

Does this mean that I do not need an MPS?

ps, I am also aware of the following questions on a similar topic, but it seems that my question is not answered clearly

thanks

You can run multiple MPI ranks on a single GPU without MPS. In that case, all the rank (GPU) code will be sorted. The GPU code given will begin executing only after the GPU code associated with the previous rank completely expires and is out of the GPU.

If you want to get an opportunity to rank GPU code from a rank with a GPU code, then MPS will be required. If the GPU code associated with a rank makes full use of GPU, then you are unlikely to see a lot of profit from the MPS, the rank GPU code can execute concurrently with the GPU code of the second-rank.