Re: netcdf parallel
On 14/01/2016 13:00, Bas Couwenberg wrote:
> On 2016-01-14 13:16, Alastair McKinstry wrote:
>> As Nico points out, the two versions (serial, parallel) provide
>> different features: notably only the serial version supports compression
>> (and threads), only the parallel version does parallel IO.
>> However any sufficiently complex application will eventually be linked
>> to _both_. Examples are Paraview / VisIt, which have plugins for reading
>> "all scientific" data, including met data (using netcdf4/hdf5,
>> compression), and the exodus/adios plugins for parallel reads.
>>
>> So, we need to copy the HDF5 method and install libnetcdf*-serial.so* ,
>> etc. libraries that are co-installable.
>> These should change SONAMES, symbol suffixes so they can be both linked
>> into the same binary.
>>
>> We probably need libnetcdf-serial-dev and libnetcdf-mpi-dev packages to
>> be co-installable too.
>
> I'd prefer to keep the serial flavor unchanged wrt to upstream, and
> only customize the mpi flavors (mpich & openmpi).
That would be good, at least in terms of least surprise for the user.
NetCDF does not do parallelism unless explicitly asked by the caller,
so just dropping a parallel implementation
in place of a serial one will not provide any benefits to the user; the
cost of having two libraries co-installed
is negligible so I don't think there is any benefit in having the
'alternatives' approach we use in HDF5 packages.
In terms of -dev, we _should_ be able to get away with just
libnetcdf_mpi_dev, linking against
libhdf5-mpi-dev. I think we should just build one MPI flavour, -mpi in
the package name
and _MPI in the symbols, rather than libnetcdf-openmpi-* libnetcdf-mpich-*
We need to avoid combinatorial explosion.
> So if you'd like to see a parallel netcdf library materialize, active
> contributors are required to make it happen. My requirement is mostly
> that these changes need to be coordinated with upstream, if netcdf
> upstream is not interested in those changes I'm not willing to
> maintain it either.
>
Yes. I've done some work with it, largely because parallel netcdf is
needed for my packages exodusII and ADIOS
(IO libraries used in HPC, and dependencies of VisIT which I'm ITP'ing),
so I'll try to get it done for Stretch.
> Kind Regards,
>
> Bas
>
Regards
Alastair
--
Alastair McKinstry, <alastair@sceal.ie>, <mckinstry@debian.org>, https://diaspora.sceal.ie/u/amckinstry
Misentropy: doubting that the Universe is becoming more disordered.
Reply to: