CUDA MPS clients

I am reading the MPS documentation of NVIDIA and from the beginning it refers to MPS clients.
For me the term of clients is not that clear.
My question is:
MPS clients are:
a) the different processes of the user that has started the MPS daemon?
or
b) the different users that ask to start the MPS daemon?

MPS clients are the processes that submit CUDA work after the MPS server is started.

Refer to section 3.2:

https://docs.nvidia.com/deploy/pdf/CUDA_Multi_Process_Service_Overview.pdf

The MPI process A and process B are example MPS clients (in that example, where we have two MPI processes that are each submitting “work” i.e. kernels to be handled by the MPS server).