Handling large datasets

The SpectralCube class is designed to allow working with files larger than can be stored in memory. To take advantage of this and work effectively with large spectral cubes, you should keep the following three ideas in mind:

  • Work with small subsets of data at a time.
  • Minimize data copying.
  • Minimize the number of passes over the data.

Work with small subsets of data at a time

Numpy supports a memory-mapping mode which means that the data is stored on disk and the array elements are only loaded into memory when needed. spectral_cube takes advantage of this if possible, to avoid loading large files into memory.

Typically, working with NumPy involves writing code that operates on an entire array at once. For example:

x = <a numpy array>
y = np.sum(np.abs(x * 3 + 10), axis=0)

Unfortunately, this code creates several temporary arrays whose size is equal to x. This is infeasible if x is a large memory-mapped array, because an operation like (x * 3) will require more RAM than exists on your system. A better way to compute y is by working with a single slice of x at a time:

y = np.zeros_like(x[0])
for plane in x:
    y += np.abs(plane * 3 + 10)

Many methods in SpectralCube allow you to extract subsets of relevant data, to make writing code like this easier:

Many methods in SpectralCube iterate over smaller chunks of data, to avoid large memory allocations when working with big cubes. Some of these have a how keyword parameter, for fine-grained control over how much memory is accessed at once. how='cube' works with the entire array in memory, how='slice' works with one slice at a time, and how='ray' works with one ray at a time.

As a user, your best strategy for working with large datasets is to rely on builtin methods to SpectralCube, and to access data from filled_data() and unmasked_data() in smaller chunks if possible.

Warning

At the moment, argmax() and argmin(), are not optimized for handling large datasets.

Minimize Data Copying

Methods in SpectralCube() avoid copying as much as possible. For example, all of the following operations create new cubes or masks without copying any data:

>>> mask = cube > 3  
>>> slab = cube.spectral_slab(...)  
>>> subcube = cube[0::2, 10:, 0:30]  
>>> cube2 = cube.with_fill(np.nan)  
>>> cube2 = cube.apply_mask(mask)

Minimize the number of passes over the data

Accessing memory-mapped arrays is much slower than a normal array, due to the overhead of reading from disk. Because of this, it is more efficient to perform computations that iterate over the data as few times as possible.

An even subtler issue pertains to how the 3D or 4D spectral cube is arranged as a 1D sequence of bytes in a file. Data access is much faster when it corresponds to a single contiguous scan of bytes on disk. For more information on this topic, see this tutorial on Numpy strides.