Simon Byrne

For what it's worth, I was unable to get `MPIEXEC_PREFIX_DEFAULT` to do anything with MPICH, but the `-prepend-rank` (which documented in the [help screen](https://github.com/pmodels/mpich/blob/cb944baeb07061759f7c22d85704cc9c673056f9/src/pm/hydra2/mpiexec/mpiexec_params.c#L95) but not the man page) does...

One option would be an interface such as `Cprintf` in Chapter 8 of the "Parallel Programming with MPI" book https://github.com/cyliustack/benchmark/blob/b91924d5dc842906ebf94d4b154d548d944a030f/mpi/ppmpi/chap08/cio.c We could define an interface like ```julia MPI.Cprint(comm, root) do...

I also suggested what I think is a better solution to the MPI forum: https://github.com/mpi-forum/mpi-issues/issues/296

I'm not quite sure yet how it would work. One option would be to modify `Base.stdout`, but I don't think that is a good idea as it won't help with...

cf https://github.com/pmodels/mpich/issues/4632

Not that I know of: unfortunately there are no APIs for controlling buffers (each MPI implementation handles the output combination differently).

> It would be good to have this be configurable, I have to think a bit more about this, but MPI implementation are free to not be threadsafe (or use...

Would that still work if we were calling MPI on multiple threads?

What I was thinking was something where each thread was creating its own CuArray, executing a kernel and then passing it to MPI: ```julia Threads.@threads for i in 1:Threads.nthreads() data...

Thanks. So I guess the question then is should we call this by default whenever a CuArray is passed as a buffer? Or leave it to the user?