Writing on external temporary files with MPI.jl

General Usage
1

Hi all. I’m learning how to use MPI with Julia, specifically the MPI.jl package.
I’m trying to understand how to print on external temporary files, and I can’t make it work.

This is my current minimal non-working example script, called mpi_test.jl:

using MPI
MPI.Init()

let
    comm = MPI.COMM_WORLD
    prank = MPI.Comm_rank(comm)
    io_file, _ = mktemp()
    io_file = MPI.bcast(io_file, comm)
    f = MPI.File.open(comm, io_file; write=true)
    MPI.File.write(f, "$prank: hello!")
    
    io_files = MPI.gather(io_file, comm)
    prank == 0 && for p in 0:MPI.Comm_size(comm)-1
        println(p, ": ", io_files[p+1])
    end
end

which I run with mpiexecjl -np 4 julia --project mpi_test.jl.
The output is

0: /tmp/jl_FokE6l
1: /tmp/jl_FokE6l
2: /tmp/jl_FokE6l
3: /tmp/jl_FokE6l

but the files are not actually there:

cat: /tmp/jl_FokE6l: No such file or directory

If I write to an ordinary file, say io_file = "test.txt" instead of using mktemp(), the file is correctly created in the working directory and contains the given string.
With mktemp() instead, no file is created at all in the /tmp folder.

Can someone explain what I’m missing here?

2
  • Did you mean to call tempname() instead of mktemp(), since you want MPI.File.open to be the function that actually creates the file?
  • Note that there is no point in calling tempname() on anything except for prank == 0, since that is where your are broadcasting (bcast) from.
  • Did you forget to call close(f)?

(I assume you are running on a single node, so that all the MPI processes have access to a single /tmp filesystem?)

3
  • Did you mean to call tempname() instead of mktemp(), since you want MPI.File.open to be the function that actually creates the file?

Huh, I never realised mktemp() already opened the file. I guess I always used it wrong.

  • Note that there is no point in calling tempname() on anything except for prank == 0, since that is where your are broadcasting (bcast) from.

Yeah I know, it doesn’t matter for this simple example, I just found it simpler to write.

  • Did you forget to call close(f)?

I deliberately left it out, I assumed it didn’t change the outcome.

  • (I assume you are running on a single node, so that all the MPI processes have access to a single /tmp filesystem?)

Yes, I’m running it on a single node, on my laptop.

Anyway I switched from mktemp to tempname and closed the file at the end, but the behaviour didn’t change.

4

Well, I’m just dumb, I found out why this happens, and it has nothing to do with MPI: the temporary files were deleted by Julia when the script ended because I didn’t supply the cleanup=false argument to mktemp().