It turns out that cargo run --release does the trick! What I was doing was cargo build --release and then target/kmeans. Is there a way to build with optimizations on, instead of relying on Cargo to run the program?

Turns out, this implementation is even slower than the previous one. On my laptop, this implementation runs in 1807ms. You can see the whole Rust implementation, as well as a few other languages, at github.com/andreaferretti/kmeans

Thank you very much! I was suspecting that the use of a TreeMap was the culprit, but I was unable to implement Hash for Point. I understand the issues of floating point arithmetic, but still being able to group points by their nearest neighbour seems a quite reasonable task for which a hashmap is definitely desired. In fact, your solution requires the use of unsafe code, which is unexpected for this task! I will try this implementation on the same laptop I have used for the other benchmarks and let you know the time

@Shepmaster For 100000 points, 10 clusters, 15 iterations, averaged over 100 repetitions, I get 1036 ms for Scala and 1373 ms for Rust. Since then, I also tried Pypy, which got an amazing 721 ms