Warn users if mpi4py running under different MPI implementation

It does happen to unexperienced users that they run pacakges using mpi4py under an MPI implementation different from the one that mpi4y was built with.

MPI.jl, the analogue of mpi4py in Julia land, tries to warn users if this happens by detecting at MPI_INIT time if MPI_COMM_WORLD reports only 1 rank but the known environment variables (MPI_LOCALNRANKS, OMPI_COMM_WORLD_SIZE) suggest that there should be more than 1 rank. This happens exactly when mpi4py would be built with mpich but run under open-mpi.

If this condition is detected it prints a warning.

What do you think about doing this? Is it something that you would merge into mpi4py?

[dalcinl]

@leofang These envars are set by the mpiexec command in each process it executes, and then the MPI library takes over to assign a rank to MPI_COMM_WORLD. If mpiexec is not used (or one from a different implementation is used), then the expected environ vars are not set and the MPI library assumes a singleton MPI_Init() case, thus bootstrapping a sequential run (world size 1).

If there is a mismatch between the size of MPI_COMM_WORLD and these environment variables, you can reasonably assume that users are messing up: the mpiexec command does not match the MPI library. Did I make myself clear enough or should I provide a specific example using the command line?

[hzhou]

As I said, the reliable way to check PMI is through PMI. We are about to release mpich 4.1a1, which will have a separate libpmi package that third-party can directly use. A simple call to PMI_Init should reveal enough information on the PMI versions and compatibility. We haven’t changed the PMI protocols ever, so it should work all the way back to old releases.

PS: although making multiple PMI_Init for the purpose of version check is untested, but I think should work.