Bug#961108: openmpi: providing 64-bit MPI
Source: openmpi
Severity: normal
Bug#953116 requests that petsc be built with 64 bit support for node
indices (for addressing arrays and datasets, not simply 64-bit double).
That's fair enough, it would enable massive systems to be modelled
with our software. Coronavirus protein-membrane docking, say. Ocean
currents and weather forecasting. Formation of galaxies. The big
projects. FEniCS will be able to use it.
If I understand correctly, it's dangerous to simply enable 64-bit in
PETSc alone. It needs to be done all along the computational library
stack.
blas already provides 64-bit builds.
MUMPS can be built for 64-bit handling (64-bit fortran integers), but
its documentation says 64-bit handling needs to be enabled in the
subsystems it uses:
METIS, SCOTCH (and BLAS, LAPACK, already 64-bit enabled)
and MPI.
Of the lower libraries, 64-bit build in MPI (i.e. 64-bit MPI_INTEGER,
counts) will be the most awkward for us, given the complexity of the
mpi packaging (with openmpi, mpich alternatives)
So I'm starting the discussion here with Debian Science and openmpi,
are we ready to take on 64-bit enablement of our entire
computational library stack?
Drew
Reply to: