Comparisons with xarray-regrid

[aulemahal]

The xarray-regrid package seems to have better performance in some areas where I don't see why xESMF should be different.

Two issues :

• Improve performance of conservative routine xarray-contrib/xarray-regrid#42
• Comparison with xESMF xarray-contrib/xarray-regrid#48

Probably that our chunking management is not the best.

Also, I think it would be good to mention xarray-regrid in our doc and to point out the differences. xarray-regrid only handles rectilinear grids, but it is much more lightweight and handles out-of-memory grids, so we could direct users there if it fits their needs.

[BSchilperoort]

Hi, thanks for mentioning xarray-regrid here!

I don't see why xESMF should be different

Well, xarray-regrid only solves a subset (rectilinear -> rectilinear girds) of what ESMF can do. This makes it so we can solve the regridding per coordinate. For example, to convert a 10x10 rectilinear grid to a 20x20 rectilinear grid, you need 2 vectors with 200 elements each for xarray-regrid. But to convert a 10x10 rectilinear grid to a 20x20 curvilinear grid, you need a matrix with 40,000 elements for ESMF.

Of course these matrices are very sparse (but this applies to both cases).

Lastly, in my experience (from trying out things over the last years) it's a lot easier to make the coordinate-by-coordinate dot product computations of xarray-regrid fast, compared to the big matrix multiplications of xESMF. Not saying it's not possible of course.