I don't know why these packages are always such a pain to install. I have been using NetCDF/HDF5 for a long time now and it's always been a pure horror trip getting them to install or run properly, no matter if it's on Linux or OSX, no matter if C, C++ or now python. The simple dependency between netcdf4 and hdf5 is a source of great pain for many people and I really wish the developers of those packages would finally do something about it.
So, the latest concrete problem I am facing is this: I am trying to install netCDF4 for python. I get the following error:
Package hdf5 was not found in the pkg-config search path Perhaps you should add the directory containing `hdf5.pc'
I tried to install the hdf5 packages using apt-get, including:
- libhdf5-serial-dev
- libhdf5-serial
- libhdf5-7
- python-h5py
- libhdf5-dev
- hdf5-tools
- hdf5-helpers
- libhdf5-7-dbg
Using pip, I tried:
pip install h5py
which failed miserably to resolve a dependency to Cython, which I then installed manually. After that it installed (apparently) but I can not find the file hdf5.pc anywhere.
I am pulling my hairs out here. Anyone know how to work around this problem?
When you mix distribution packages and self-built packages, you are increasing your chance of problems (as you are finding out).
Also, do you want h5py or do you want netcdf-python? I don't think netcdf-python has a dependency on h5py. Rather, netcdf-python binds to the C netcdf library, which in turn depends on the C HDF5 library.
h5py likewise binds to C HDF5
There is a lot of software involved, it's true. Work your way through step by step and it will make more sense eventually (says the guy who has been doing this for 15 years... it gets easier!)
- If you are going to do any parallel programming, you'll need an MPI implementation
- HDF5 now provides the foundation for NetCDF4. If you want parallel programming, build HDF5 against your MPI implementation.
- Install the C library of NetCDF4
- now the python bindings can pick up what they need from NetCDF4, HDF5, and MPI
Yes it is a lot of software to configure and build. pkg-config can help a lot here! When you see Package hdf5 was not found in the pkg-config search path, that means you should adjust your PKG_CONFIG_DIR to point to the location of the package-config files. Unfortunately, hdf5 doesn't provide a .pc (package-config) file, so you'll have to just do that part by hand. Oh, and netcdf doesn't provide a pkg-config either: it provides a script nc-config that netcdf-python will use.
Let me provide a concrete example:
- MPICH-master installed in
/home/robl/soft/mpich-master - HDF5 installed in
/home/robl/soft/hdf5-1.8.16 - e.g configured like
../../hdf5-1.8.16/configure --prefix=/home/robl/work/soft/hdf5-1.8.16 CC=/home/robl/work/soft/mpich/bin/mpicc --enable-parallel
- NetCDF4 installed in
/home/robl/soft/netcdf-master - e.g. configured like
./configure CC=${HOME}/work/soft/mpich/bin/mpicc --prefix=${HOME}/work/soft/netcdf-master CPPFLAGS=-I${HOME}/work/soft/hdf5-1.8.16/include LDFLAGS=-L${HOME}/work/soft/hdf5-1.8.16/lib
now you have all the pre-requisietes for netcdf-python
by the way, http://unidata.github.io/netcdf4-python/ lays out the prerequisites and the necessary configure options
Don't get hung up on the carping about hdf5.pc. If you have nc-config in your path, it will provide the needed information.
If you are building for parallel programming, set CC to your MPI compiler. if not, you can skip the ``export CC=...'' step:
cd netcdf-python export CC=${HOME}/work/soft/mpich/bin/mpicc export PATH=${HOME}/work/soft/netcdf-master/bin:${PATH} python setup.py build
