search

pysal / momepy

Urban Morphology Measuring Toolkit
https://docs.momepy.org
BSD 3-Clause "New" or "Revised" License
496 stars 59 forks source link

test efficiency of graphblas in straightness #562

Closed martinfleis closed 5 months ago

martinfleis commented 8 months ago

This is more of an exploration than anything. I am curious if graphblas_algorithms engine for networkx can help us with our stuff, so I reimplemented straigtness_centrality on top of it. For now only the core part. Using the network from our docs, I am able to get it from 1min 50s down to 17s, which means it is 6.5x faster. There's nothing to shave from our overhead as 98% of time is now spent on all_pairs_bellman_ford_path_length. I was expecting a bit more from this, given the speedups in benchmarks in their Readme are in hundreds, but better than nothing :).

The code for networkx and graphblas is similar but not the same. Networkx-compatible code is possible but much slower than this. So having graphblas as optional dependency (like numba elsewhere in PySAL) would mean having two implementations, which I am not a fan of. But as @jGaboardi pointed out, graphblas is not installable on Apple Silicon under Python 3.12 at the moment. If there'll be delays of this sort, we I am not sure I like depending on it. (Though it is installable from PyPI...).

martinfleis commented 8 months ago

On the installation, python-graphblas switched from noarch to compiled code some time ago but conda-forge is not set to build arm64. Fixing that in https://github.com/conda-forge/conda-forge-pinning-feedstock/pull/5676.

Once the recipe gets rebuilt, the installation should not be an issue any longer.

martinfleis commented 5 months ago

I'll close it as it was a test. It seems that while the performance can be improved, graphblas_algorithms might need to mature a bit before relying on it.