Concurrency

Rayon is the first crate to reach for when you want to use all CPU cores. Changing iter() to par_iter() runs the loop across a work-stealing thread pool, and Rust's ownership rules ensure the closure cannot cause data races. It also provides join and scope for recursive divide-and-conquer work. Rayon is for CPU-bound tasks; for waiting on network or disk, use an async runtime instead.

Crossbeam provides the building blocks under many concurrent programs: fast multi-producer multi-consumer channels with select, epoch-based memory reclamation for lock-free data structures, and work-stealing deques. Scoped threads started in Crossbeam and are now part of the standard library as std::thread::scope. The standard library's mpsc channel has also been reimplemented using Crossbeam's channel design, so it is no longer the slow option it once was.

archery abstracts over Rc and Arc so data structures can be generic over thread safety, and orx-parallel is another parallel iterator implementation. coroutine-rs and coio-rs explore coroutines, and rate-limiters-rs provides rate-limiting algorithms. For most programs, std plus Rayon and Crossbeam is enough. Reach for the more specialised crates only when profiling shows contention or you need a structure these do not offer.