Python and HDF5

Chapter 4. How Chunking and Compression Can Help You

So far we have avoided talking about exactly how the data you write is stored on disk. Some of the most interesting features in HDF5, including per-dataset compression, are tied up in the details of how data is arranged on disk.

Before we get down to the nuts and bolts, there’s a more fundamental issue we have to discuss: how multidimensional arrays are actually handled in Python and HDF5.

Contiguous Storage

Let’s suppose we have a four-element NumPy array of strings:

>>> a = np.array([ ["A","B"], ["C","D"] ])
>>> print a
[['A' 'B']
 ['C' 'D']]

Mathematically, this is a two-dimensional object. It has two axes, and can be indexed using a pair of numbers in the range 0 to 1:

>>> a[1,1]
'D'

However, there’s no such thing as “two-dimensional” computer memory, at least not in common use. The elements are actually stored in a one-dimensional buffer:

'A' 'B' 'C' 'D'

This is called contiguous storage, because all the elements of the array, whether it’s stored on disk or in memory, are stored one after another. NumPy uses a simple set of rules to turn an indexing expression into the appropriate offset into this one-dimensional buffer. In this case, indexing along the first axis advances us into the buffer in steps (strides, in NumPy lingo) of 2, while indexing along the second axis advances us in steps of 1.

For example, the indexing expression a[0,1] is handled as follows:

offset = 2*0 + 1*1 -> 1
buffer[offset] -> value "B"

Note

You might notice that ...