One more general thing that I’d like to share with whomever it might concern: I’ve made the mistake in the past to think that something would be sped up by doing it in Rust where it turned out that the actual work that was performed in Rust was not worth it. There is a measurable overhead in context switching:
- Reading out Elixir datatypes on the Rust side often involves some parsing with many pointer jumps, since that is how the Elixir/Erlang datastructures are represented under the hood.
- If you want your results to be usable in Elixir, you often need to convert your Rust datatypes back to e.g. an immutable Erlang datastructure.
- I believe the runtime performs some work to make sure the BEAMworld matches what the NIFside would expect. (Note that this is limited; it is very easy to break everything when writing a NIF by e.g. subtly corrupting memory, which is one of the reasons why writing NIFs in Rust is nicer than C).
In these situations, either write a more clever algorithm in Erlang/Elixir, or restructure your algorithm such that more work is done per call to the NIF.
And as always: Measure the difference in performance, to be sure it is actually worth it 
.
