Add parameters to limit number of settled nodes during preparation, much faster preparation for certain graphs

[easbar]

Fixes #34.

Long story short: When there are large-weight edges the preparation time can become much slower because witness searches explore too many nodes. Preventing this and still not adding too many shortcuts is non-trivial, so this PR introduces some parameters that can be used to tune performance depending for different types of graphs. It also provides some default values that should work reasonably well in most cases.

todo:

• update benchmark values in README.md
• add benchmark values to mini-performance tests in lib.rs and maybe further tune parameters

[dabreegster]

Awesome, thanks!

[dabreegster]

These explanations are great, thank you! I kind of want to ask for something similar for hierarchy_depth_factor and edge_quotient_factor -- would users ever want to deviate from defaults?

[easbar]

Good point, I added some documentation for these as well. I also ran some measurements for varying values of hierarchy_depth_factor, see here: https://gist.github.com/easbar/a7980e834206b6065e848ba29e228fdd#hierarchy-depth-factor

For the abstreet maps you sent reducing hierarchy_depth_factor from 0.1 to 0.01 might actually reduce preparation times further (but also might yield somewhat slower queries, ~10% according to these experiments, but that could also just be noise).

[easbar]

That sounds unexpected for bike/car/walk. Most nodes are directed road segments, and from a quick glance at a reachability debugger, there are only a few roads not reachable:

Hm, strange. I might also check if the graphs are strongly connected. I agree that looking at your reachability debugger 25% of failed paths are unexpected.

The new default settings are indeed a huge performance boost.

I'm glad to hear this. Note that for your maps I think Params::new(0.01, 100, 10, 100) and ParamsWithOrder::new(100) work better than the defaults. I'll send a PR once this is merged.

For now, there's no recorded set of realistic user edits, but I'll keep an eye on the re-preparation times as I test things.

Yes, a set of realistic edits would be helpful to analyze this further for sure.