Add regridding benchmark

[hendrikmakait]

Closes #1556

[mrocklin]

Do we have enough context here to add this to the benchmark post? If you give me bullet points and an image (if one makes sense) I'm happy to write up words.

[aulemahal]

Hi! I'm one of the de facto maintainer of xESMF and I think this benchmark is a good start!

To complexify the test and test areas where xESMF needs more improvement, I would suggest using a very large grid either as input or output and have some chunking across the spatial dimensions.

IIUC, these benchmarks are more geared towards dask ? Another bottleneck of xESMF is the generation of the weights (the Regridder initialization) with very large grids and more complex methods ("conservative"). But that part is neither parallelized nor lazy, so benchmarking this might be out of scope here.
We do have some code to make the weights generation in parallel but I would say it is still experimental and of limited scope.

[hendrikmakait]

@aulemahal: Thanks for the input. I'll add a follow-up issue to look into some of the suggestions for increasing the complexity of this workload.

IIUC, these benchmarks are more geared towards dask ?

It is, but it's also aimed at reflecting real workloads. Would more complex methods also result in more complex computations or just in more complex weight generation?

We do have some code to make the weights generation in parallel but I would say it is still experimental and of limited scope.

Please let us know if we can help with anything from a Dask perspective.

@mrocklin:

Do we have enough context here to add this to the benchmark post? If you give me bullet points and an image (if one makes sense) I'm happy to write up words.

I'll whip something up. At first glance, this benchmark seems to do alright; it's mostly an embarrassingly parallel computation. Performance and # of tasks could probably look better but that's already a lot better than some of the other benchmarks.

[aulemahal]

Would more complex methods also result in more complex computations or just in more complex weight generation?

Mostly more complex weight generation, which is totally on the ESMF side, so partly in C/Fortran I think. Maybe two very different grids (curvilinear ones for example) and more complex methods would make for weights with more connected nodes, but I don't think this would affect the computation so much.

[slevang]

xarray-regrid is a much less established tool than xesmf, but could alternatively (or in addition) be used as a more "pure" test of the dask workload. It assumes rectilinear grids and therefore separates the operations along each axis, which makes weight generation near instantaneous.

The resulting dask workload is very similar to xesmf, just an einsum(data, weights), at least for the conservative method. I wrote up a little notebook comparing the two tools here. Both libraries now use sparse weights.

The case in which we have chunking along the dimensions to regrid would also be interesting to add to your benchmark, but I don't know of any publicly available equivalents to the GCP ERA5 ARCO stores with that sort of chunking.

[hendrikmakait]

@slevang, thanks for the additional input! Would you be interested in contributing a benchmark implemented with xarray-regrid?

[phofl]

small comment, otherwise lgtm

[phofl]

thx