suggestions to accelerate rust phone_encoder

[omac777]

I understand the original implementations did not consider any parallelism mechanisms at that time since they weren't easy to use.

A few suggestions to accelerate the rust implementation of phone_encoder:

• Tokio and rayon crates could definitely speed things up with your rust implementation. Also hardware choices would impact the performance of the parallel implementation as well. Rather than using traditional mechanical hard-drives, use nvme drives.
  • on nvme drives you may be able to parallelize the reading and the writing of the data. The data layout could be done differently to maximize that.
  • reading the lines of words.txt and numbers.txt could be done in parallel
  • did the other implementations have checking if input was alphanumeric for numbers.txt? That's actually wasted time. It should treat the numbers only and panic/fail/error otherwise. That will save time to take out that check
  • what's the purpose of collecting the numbers if you already have a vector of different input on each line? Just parse it and pass the data to the next step via a channel. Don't collect. collect is literally a funnel which ultimately is a bottleneck.
  • the load_dict could be done in parallel too. You may read the lines of words in parallel, word_to_number, then pass the number to the next step via a channel to create an entry and push it.
  • the print_translations is doing two things in sequence: translating each bigint and printing the result. It should be split up into two different functions. the translate could be done in parallel for each bigint and sent to a channel for printing when completed.

Thank you for listening.

[renatoathaydes]

I would be happy to see what code that does this would look like, and how performant it would be!

But I won't have time to do that myself. Feel free to submit a PR!

If you haven't seen it yet, I've just written a followup post in which I analyse the Rust code in more depth thanks to the guys at #1 : https://renato.athaydes.com/posts/how-to-write-slow-rust-code.html

[ElhamAryanpur]

I'd also like to add while reading the Cargo.toml that, no release profile optimizations were added...

A release profile such as this could improve by quite a bit:

[profile.release]
opt-level = 3
lto = true
codegen-units = 1
debug = false

although at the cost of compile time

[renatoathaydes]

@ElhamAryanpur that was done in my series of blog posts for Rust. See Profile Guided Optimisation in https://renato.athaydes.com/posts/how-to-write-fast-rust-code

Nice! Just by re-compiling with PGO, without having to make any code changes, we get yet another 5% improvement in performance.

[ElhamAryanpur]

@renatoathaydes oh my bad then, I didn't see it. Loved the article truly!

[renatoathaydes]

No problem, glad that you liked it! I should add links to the previous articles, I realized I haven't done that properly yet.