[Paraview-developers] Problems with mpi4py and SGI's mpt 2.14 version of MPI

[Ben Boeckel]

On Mon, Apr 03, 2017 at 14:15:49 +0000, Hennessey, Joseph G CTR USARMY RDECOM ARL (US) wrote:
> I have been running into problems with
> mpi4py and SGI's mpt 2.14 version of MPI
> while compiling ParaView 5.3.0 on HPC systems.
> 
> mpi4py cannot find MPI_CONVERSION_FN_NULL
> when compiling mpi4py.MPI.c
> 
> MPI_CONVERSION_FN_NULL was not used by the
> version of mpi4py that came with ParaView 5.2.0
> and when I swap the ParaView 5.2.0 version of mpi4py
> in place of the ParaView 5.3.0 version of mpi4py then
> ParaView 5.3.0 (with the 5.2.0 version of mpi4py) compiles just fine.
> 
> Is this a known problem? Is there a minimum version
> of mpt that is required for ParaView 5.3.0? What 
> will break if use the ParaView 5.2.0 version of mpi4py
> instead with ParaView 5.3.0?

We updated from mpi4py from 1.3.0 to 2.0.0. You can see
upstream-documented differences here:

    https://gitlab.kitware.com/vtk/vtk/blob/master/ThirdParty/mpi4py/vtkmpi4py/CHANGES.rst

It doesn't appear that VTK cares about the backwards-incompat changes
listed there (we use it to allow MPI Controllers to be shared between
VTK's C++ code and other Python code).

It seems that the MPI_CONVERSION_FN_NULL is used because the
upstream-generated Cython code says it exists. Maybe the module could be
updated to optionally rerun the Cython code generator so that it works
with an arbitrary MPI?

--Ben