Should we remove haskey(ENV, "PMI_RANK") || haskey(ENV, "OMPI_COMM_WORLD_RANK") check in context_type?

[charleskawczynski]

This seems to be causing issues for people when they don't explicitly request MPI (and it automatically gets scooped up). I think it might be less surprising if we change

function context_type()
    name = get(ENV, "CLIMACOMMS_CONTEXT", nothing)
    if !isnothing(name)
        if name == "MPI"
            return :MPICommsContext
        elseif name == "SINGLETON"
            return :SingletonCommsContext
        else
            error("Invalid context: $name")
        end
    end
    # detect common environment variables used by MPI launchers
    #   PMI_RANK appears to be used by MPICH and srun
    #   OMPI_COMM_WORLD_RANK appears to be used by OpenMPI
    if haskey(ENV, "PMI_RANK") || haskey(ENV, "OMPI_COMM_WORLD_RANK")
        return :MPICommsContext
    else
        return :SingletonCommsContext
    end
end

to

function context_type()
    name = get(ENV, "CLIMACOMMS_CONTEXT", nothing)
    if !isnothing(name)
        if name == "MPI"
            return :MPICommsContext
        elseif name == "SINGLETON"
            return :SingletonCommsContext
        else
            error("Invalid context: $name")
        end
    else
      return :SingletonCommsContext
    end
end

[haakon-e]

I was just hit by this, and am for the time being working around it by doing a few additional checks, e.g.:

auto_ctx = ClimaComms.context_type() == :MPICommsContext
two_or_more_gpus = length(split(cuda_devices, ',')) ≥ 2
two_or_more_slurm_tasks = parse(Int, get(ENV, "SLURM_NTASKS", 0)) ≥ 2
is_mpi_ctx = auto_ctx && (two_or_more_gpus || two_or_more_slurm_tasks)

main_context = if is_mpi_ctx
    import MPI
    ClimaComms.MPICommsContext(device)
else
    ClimaComms.SingletonCommsContext(device)
end

[nefrathenrici]

I support this change, using both the CLIMACOMMS_CONTEXT and the MPI variables leads to unexpected behavior, we should improve this check or just use the CLIMACOMMS_CONTEXT to set the context.

[nefrathenrici]

The main issue I have run into is that srun sets PMI_RANK, so we should change this variable check to something like MPICH_VERSION, or cast a wider net of variable checks to be safer.

[charleskawczynski]

Maybe we could change the context to contain a string / symbol indicating why that context was returned. Then, users could print:

context = ClimaComms.context()
@show ClimaComms.why(context) # prints, e.g., :explicitly_requested

This could be really helpful on the user-side.