pm4py
draft: Perf: optimize petri net replay
This was worked on during the Celonis Impact Day on the 29th of September 2023
This PR suggests a few changes to the Petri net semantics class to optimize the Petri net replay. The changes go from less to more radical, so it's up for discussion about which changes should be accepted. The optimizations is guided by the benchmarks. Given a Petri net (the 5 added), it measures the total runtime to compute the reachable markings of the Petri net. The results can be found below (runtime in seconds):
| Model | # States | Initial Time | Decorators | Slots | Faster Hashing | Cache The ID | Iterate and Unpack Items | Native Copy | Marking as Dictionary | Compiled Semantics | Use Place IDs |
|---|---|---|---|---|---|---|---|---|---|---|---|
| M1 | 144 | 0.0076 | 0.0077 | 0.0072 | 0.0070 | 0.0065 | 0.0061 | 0.0054 | 0.0042 | 0.0019 | 0.0015 |
| M2 | 230 | 0.013 | 0.013 | 0.012 | 0.012 | 0.011 | 0.010 | 0.0090 | 0.0067 | 0.0038 | 0.0028 |
| M7 | 36740 | 4.69 | 4.83 | 4.56 | 4.24 | 3.83 | 3.48 | 2.99 | 2.18 | 1.30 | 0.94 |
| ML2 | 82939 | 16.63 | 16.75 | 15.91 | 14.87 | 13.46 | 12.79 | 12.06 | 10.33 | 2.45 | 1.83 |
| ML4 | 1154 | 0.076 | 0.076 | 0.074 | 0.069 | 0.063 | 0.059 | 0.051 | 0.036 | 0.023 | 0.018 |
If you ask me, only slots are not worth the trouble, because it will potentially break existing code (if the users have been misusing the class). The change from counter to dict would also require some rework before deploying. In the end, we would end up reimplementing the Counter Python class, but that is alright because the counter class is just very slow. The usage of place IDs instead of places for the markings could be hidden behind some nice abstractions (example: a "CompiledMarking" class)
Attention Do not merge the PR with the Petri nets (and the benchmark script)
@Javert899 Is there anything missing on this PR? I took a look at the Dijkstra less memory method and it should be possible to rewrite the compiled semantics class to use some of the techniques there. This should make the alignment code cleaner as well as exposing this efficient representation to the end user
@EduardoGoulart1 we will integrate more when our contribution level agreement will be approved by the Fraunhofer.
Until then, we will try to identify specific contributions with high impact and integrate giving due credits
@fit-alessandro-berti Since there has been no action on this PR, do you think we should close it?
@fit-alessandro-berti closing this PR after a long period of inactivity. Feel free to incorporate the ideas on your own, if you find them useful