search

radio-astro-tools / spectral-cube

Library for reading and analyzing astrophysical spectral data cubes
http://spectral-cube.rtfd.org
BSD 3-Clause "New" or "Revised" License
98 stars 65 forks source link

Idea for performance improvements with dask spectral cubes #754

Open astrofrog opened 3 years ago

astrofrog commented 3 years ago

Currently, for FITS cubes, we convert the numpy array to a dask array with default chunking, which splits up all axes into chunks. This is the right behavior for arbitrary operations on it, but for operations that use the whole cube, such as statistics(), and for image plane calculations, this is not optimal.

One idea would be to internally have e.g three dask arrays all backed by the same data but optionally with different chunking:

self._data
self._data_global  # optimized for global operations
self._data_image  # optimized for image plane operations

For CASA cubes, these would all be the same since the chunking we have can't really be improved much. For FITS cubes, self._data_global and self._data_image would use whole image plane chunks.

Using a dask array with whole image plane chunks improves the performance of statistics() by a factor of 5x or more, and undoubtedly would also improve the performance of image-plane convolution since it is a more 'native' chunking.

keflavich commented 3 years ago

how general is this? Would there need to be different accommodations for tall, skinny cubes and short, fat ones?

astrofrog commented 3 years ago

What I'm thinking is that for FITS cubes where the Numpy array is C-contiguous (i.e. no reordering of axes to do weird WCS) for global operations and image plane operations we definitely should be using whole image plane chunks. We could set a maximum size for the image plane chunks if we thought e.g. cubes with 100,000x100,000 pixel image planes could exist though.

astrofrog commented 3 years ago

So no special casing needed aside from maybe thinking what a sensible maximum image plane chunk size would be.