Add in and out degree for states #45
Conversation
There was a problem hiding this comment.
Thanks @jeffa5. Some changes will be needed before this can be merged as I'm seeing a significant performance difference. For instance, the 2pc check 10 benchmark increases from ~70 seconds on my current laptop to ~290 seconds.
I didn't profile this to understand where the time is being taken, but I did leave a couple of thoughts in the comments.
Beyond that, my high level assessment is that this might be best implemented as a visitor so that the cost is only paid when someone opts in. e.g. they might opt out for unit tests but then opt in when debugging/etc.
Please let me know what you think and if there's anything I can do to help. If you want to quickly iterate on ideas then we can do that over Discord chat, for instance.
| @@ -255,6 +262,11 @@ where M: Model + Send + Sync + 'static, | |||
| // Otherwise enqueue newly generated states (with related metadata). | |||
| let mut is_terminal = true; | |||
| model.actions(&state, &mut actions); | |||
| let generated_fingerprint = fingerprint(&state); | |||
There was a problem hiding this comment.
Could you reuse state_fp to avoid an extra fingerprint calculation here?
| struct NodeData { | ||
| in_degree: usize, | ||
| out_degree: usize, | ||
| parent_fingerprint: Option<Fingerprint>, | ||
| } |
There was a problem hiding this comment.
One thing impacting performance is the change from storing Option<Fingerprint> in the map (which encodes to a single u64 as Fingerprint is a NonZeroU64 enabling a memory layout optimization) to this new structure that cannot encode as efficiently.
I don't know whether it would help or hurt performance, but perhaps adopting different maps along the lines of https://en.wikipedia.org/wiki/Data-oriented_design is worth exploring.
|
You might want to use something like https://github.com/TimelyDataflow/differential-dataflow or https://github.com/vmware/database-stream-processor to make the calculations incremental. An example of using the former is in https://github.com/dtolnay/cargo-tally and the latter has some good examples. There is an example for in and out degrees: https://github.com/vmware/database-stream-processor/blob/main/crates/dbsp/examples/degrees.rs |
Adds tracking of the in-degree and out-degree of states in the graph.
This is reported by collecting them all into a
Vec. This isn't the most efficient but allows full analysis by the (potentially custom) reporter, including calculating percentiles.