Help needed: Fixing performance problems w/js_sys::Math::random

[NateLevin1]

Hey everybody, I'm looking for some help fixing a performance problem in some Rust WASM code I am writing.

I'm building a simulation program that needs to calculate a ton of random numbers — I'm talking in the range of 16 to 144 million or more. I've benchmarked my code with the Chrome profiler, and it says that calls to js_sys::Math::random are taking up ~1.3s of blocking CPU time. I'm pretty certain I know why this is happening, the WASM is constantly having to call the JS glue code to run Math.random which slows everything down.

Is there a way I can re-architect my code to have better performance? Is there an alternate random number generator that will be good for my use case?

Any help would be greatly appreciated 😄

[Liamolucko]

You could give the rand crate a go, which includes random number generators written in Rust that could potentially be faster than using Math.random. To seed them from the browser's RNG, you'll need to enable the js feature of the getrandom crate, like so:

# Cargo.toml
[dependencies]
rand = "0.8.5"
getrandom = { version = "0.2.8", features = ["js"] }

Math.random wouldn't be slow because of the JS glue code, though, because it doesn't have any - all it does is return a single f64, which wasm can accept directly without any glue.

I haven't actually checked whether rand is faster. If you don't need a cryptographic RNG, rand::rngs::SmallRng should also be faster than its default RNG.

[NateLevin1]

Thanks for the reply — this seems super useful, though I’m having one slightly unrelated problem. I need a random f64 between 0-1, but SmallRng seems to only be able to give random u32/u64 values. Is there an idiomatic way to convert between these, or should I use a different generator?

Thanks again for the help!

[Liamolucko]

Yes, there's an idiomatic way to convert - you need to pick a Distribution to use, which defines how to make random values from raw random bytes. In your case, you need the Standard distribution, which samples numbers from 0-1 the same as Math.random. To actually use it, you write rng.sample(Standard), where rng is the RNG you've already constructed. You could also use rng.gen(), which is shorthand for the same thing.