Dask dask arrays

[sehHeiden]

After some years of using rioxarray, I used the first times with the chunks options.
Because the raster had a matrix size of about 70k * 60 k elements. For the first times something, that did not fit into the ram when computing on the raster.
I did have high hopes, because using multiple cores for one matrix and using less ram would be nice.

I tried loading with chunks set to:

• 'auto'
• (4, 4)
• {'x': 5, 'y': 5}, or higher, like 10, or 25 in each direction
  With Client set to four or five workers.

In general the dask functions I used never finished:

1. gradient
2. multiplication --> arctan, or arctan2
3. rad2deg

Therefor, I tried to add a re-projection from the resolution of 1 meter to 5 meter.
This never finished, too.

I encountered the following problems:

• opening with chunks=(4, 4), opened with chunks=(4, y_length). I did not encounter that with the dict.
• often got a gc warning , 40+ %, or 70+ %
• sometimes got a warning, that the size of the graph is to large.
• Instead of loading less data the huge matrix, was often loaded multiple times

Hence, I opened, without chunks, resized and than computed. Which worked in seconds instead of hours.

[snowman2]

Related: #119

[sehHeiden]

Although I ran in timing problems with and without re-projection.