Add benchmark for CellList

[padix-key]

No description provided.

[codspeed-hq]

CodSpeed Performance Report

Merging #734 will not alter performance

_{Comparing padix-key:celllist-benchmark (242a5cc) with main (226f2b4)}

Summary

✅ 59 untouched benchmarks

[Croydon-Brixton]

Hey @padix-key , I've been wondering this for a while but never had the time to investigate and thought you might know this: How does biotite's internal CellList performance compare to scipy's KDTrees (https://docs.scipy.org/doc/scipy-1.15.0/reference/generated/scipy.spatial.KDTree.html)?

The functionality they provide seem very similar (querying nearest neighbors, querying all points within a distance X, looping around in case of periodic boundary conditions, etc.) but I suspect there might have been a deeper reason to implement CellList instead of deferring to scipy's KDTrees?

[padix-key]

This is a good question. There are multiple reasons:

• Years ago, when I implemented the CellList I did not know about scipy.spatial.KDTree 😉.
• The implementation is inspired by MD programs. Specifically it is able to handle periodic boundary conditions, while I do not know if K-d trees can handle them as well. (It could if, simply adjacent periodic copies of the molecule are added to the K-d tree.)
• Neighbour search in a cell list has O(1) time complexity while for a K-d tree it is O(log n) (if I am not wrong).

That being said, I would still be curious how they compare with each other in terms of performance, especially over a range of structure sizes!

[Croydon-Brixton]

Thank you for the quick response - that's very helpful and makes a lot of sense!

I agree that it'd be interesting to compare the performance. I've used it on AtomArray's before when processing the entire PDB and it was very fast, but I never did a proper comparison with CellList.

I'll put it on my backlog of things to try out -- if it were to indeed work as fast and provide the equivalent functionality, we could consider switching in KD-trees to reduce the code volume we'd need to actively maintain.

[padix-key]

I generally agree. However one additional aspect to consider is that it would add scipy as mandatory dependency (which would probably be no big deal).