Dispatch/user distribution calculation using Kullback-Leibler divergence. Allow float weights.

[tdadela]

• simplified code
• added/fixed support for float weights
• yielding next value now has O(log n) complexity, instead of Θ(n) for roundrobin.smooth
• one dependency less

[cyberw]

Btw, did you run the dispatch performance tests?

[tdadela]

I reduced test matrix to:

    worker_count_cases = [10, 100, 1000, 5_000]
    user_count_cases = [100, 1000, 10_000, 20_000]
    number_of_user_classes_cases = [1, 10, 50, 100]
    spawn_rate_cases = [100, 1000, 5000, 20_000]

and removed IO operations (print, file.write).
I executed benchmark using: time python3.12 benchmarks/dispatch.py
Results:

• old implementation:
  12.38s user 3.18s system 99% cpu 15.594 total
  12.40s user 2.99s system 99% cpu 15.420 total
  12.58s user 3.27s system 99% cpu 15.886 total
  12.11s user 3.11s system 99% cpu 15.253 total
  12.04s user 3.01s system 99% cpu 15.086 total
• new implementation:
  11.07s user 2.61s system 99% cpu 13.713 total
  11.35s user 3.04s system 99% cpu 14.414 total
  11.17s user 3.05s system 99% cpu 14.261 total
  10.94s user 3.07s system 99% cpu 14.043 total
  11.29s user 2.64s system 99% cpu 13.964 total

[cyberw]

Nice. Not a huge improvement, but an improvement nonetheless.

[tdadela]

I rerun the same benchmark for this dummy implementation:

cycle_fixed_gen = itertools.cycle([u.__name__ for u in fixed_users.values()])
cycle_weighted_gen = itertools.cycle([u.__name__ for u in self._user_classes if not u.fixed_count])

All results were above 11s. [So most time is consumed by other parts of code]
Mind you that for certain inputs, old implementation may be faster. Mostly due to "caching" in the line 412:

return itertools.cycle(gen() for _ in range(generation_length_to_get_proper_distribution))

I ran the following test a few times:

if __name__ == "__main__":

    input_data = [(u, u.weight) for u in USER_CLASSES]
    ts = time.perf_counter()
    gen = _kl_generator(input_data)
    for _ in range(1_000_000):
        next(gen)
    instantiate_duration = time.perf_counter() - ts
    print(instantiate_duration * 1000)

(USER_CLASSES is exactly the same users list as in benchmark/dispatch.py)
On my hardware, it consistently takes less than 600 ms.
Equivalent version for roundrobin.smooth takes more than 9000 ms.
So, I think we have strong confidence that this change will not cause performance regression.
Furthermore, I believe that more important than performance gains are: code simplification, removal of weird normalization which led to issues (ex. #2662) and float weight support.

[cyberw]

Great stuff! I’ll use it myself a couple times next week, and if I dont find any issues I’ll merge.

[cyberw]

I noticed your branch is based on a pretty old commit. Can you rebase on latest master?

[cyberw]

Thanks!