[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: netcdf dev-coinstallable (Was: NetCDF 4.4.1 and HDF5 1.10.0)

I'm not fixated on this namechange either way, what does everyone think?
If we were to go with the name change, this would be a good time to do it.

> I don't want to entangle this with the regular transition, adopting your
> changes needs a whole lot more thought, testing, and upstream adoption.
The dev-coinstallable change is not mature enough yet, and 'entanglement'
would be a bad idea. What i'm discussing here is just the idea of the
libnetcdf11 -> libnetcdf-11, making this at the time of this transition.
(Ideally dev-coinstallable, if agreed, would then just be a binNMU later).
> I'm very uncomfortable with adopting such a drastic change without
> buy-in from the upstream developers. Have you discussed this issue with
> the NetCDF developers?
Not yet.
I'm posting it here for examination and consideration, and by upstream.

The changes involved are:
(1) path changes, matching hdf5 design.
(2) symbol versioning, again matching hdf5. (HDF5_1.8.8 ->
HDF_MPI_1.8.8, etc)
There are no changes to existing (serial) netcdf (apart from potentially
the package name change).
The serial library is still libnetcdf.so, symbol names remain the same;
only pnetcdf and mpi
versions change.
In the event of upstream successfully merging the variants in the future
(ie. a single version,
with hdf5 in parallel and with compression, and threads working ... ),
then the changes
can go away with just a recompile, while upgrades for a transition would
just involve a temporary
thunking library for pre-existing codes (ie a version of
libnetcdf_mpi.so with symbols
nc_copy@NETCDF_MPI_XXX  etc. that links to libnetcdf.so  nc_copy @

This work then allows us to build codes against multiple versions of
netcdf simultaneously,
e.g. the ADIOS library, or Paraview/Visit, that currently must be linked
against the serial version
(in order to get compression, needed to read many datasets) but then
can't read in parallel
(needed for acceptable performance) on other sets.

Best regards
> Kind Regards,
> Bas

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: