Why HDF5

Scientific codes need one file that holds many named arrays, the shape of each array, and a few labels such as units. HDF5 is that container. A group is a directory. A dataset is an array with a type and a shape. An attribute is a small value attached to a group or a dataset. The bytes on disk are a superblock, object headers, and the raw numbers.

This library writes superblock version 0. That is the original HDF5 superblock: an 8-byte signature, sizes of offsets and lengths, and the root group's symbol-table entry. Version 0 is what current h5dump still reads, and it avoids the checksummed version 2 superblock.

NetCDF classic is an older, smaller version of the same idea: a list of dimensions, variables, and text attributes in one header. CDF-1 stores that header in big-endian and the numbers after it. NetCDF-4 keeps the NetCDF model but stores it as HDF5, using dimension-scale attributes so a dataset can name the dimension it uses. This library writes both. Classic files are the ones ncdump is required to print. The NetCDF-4 files follow the dimension scale and _netcdf_dim_info conventions and are read back here.

ADIOS2 BP5 is a directory, not a single header. The array payload can be a raw block. The metadata that bpls understands is FFS, a separate schema encoding. This library writes the directory and the raw payload, and a metadata record it can read itself. It does not claim that bpls accepts that record.

The implementation is Mojo. It does not link the C libraries. Those tools are only used, when they are installed, to check a file after it has been written.