CUDA MPS

Hi

I am trying out CUDA MPS with my applications which are ROS based.
Is it mandatory to have MPI appication to work with CUDA MPS?
What is the typical context switching time expected when we are running application without MPS?
and any detailed explanation on how GPU context swithching is done with and without MPS?

Thanks,
njs

it’s not mandatory to use MPI

the typical context switching time will vary by exact workload, GPU, and platform. Probably best for you to just take some examples in your case and measure it.

more detailed explanations of MPS are available on this forum, in questions that have been asked recently, if you care to search for them, such as this one:

[url]Question about GPU sharing of Multi-process service - CUDA Programming and Performance - NVIDIA Developer Forums

Hi txbob,

thank you for the information.

BR,
njs