2 releases

0.5.1 Feb 13, 2022
0.5.0 Feb 13, 2022

#193 in Concurrency

Download history 1467/week @ 2024-01-14 1454/week @ 2024-01-21 1251/week @ 2024-01-28 498/week @ 2024-02-04 786/week @ 2024-02-11 754/week @ 2024-02-18 523/week @ 2024-02-25 1205/week @ 2024-03-03 1613/week @ 2024-03-10 277/week @ 2024-03-17 398/week @ 2024-03-24 325/week @ 2024-03-31 463/week @ 2024-04-07 403/week @ 2024-04-14 463/week @ 2024-04-21 505/week @ 2024-04-28

1,895 downloads per month
Used in 11 crates (5 directly)

MPL-2.0 OR MIT OR Apache-2.0

4MB
1K SLoC

Parallel iterator processing library for Rust

See IteratorExt (latest IteratorExt on docs.rs) for supported operations.

Notable features

  • drop-in replacement for standard iterators(*)
    • preserves order
    • lazy, somewhat like single-threaded iterators
    • panic propagation
  • support for iterating over borrowed values using scoped threads
  • backpressure
  • profiling methods (useful for analyzing pipelined processing bottlenecks)

When to use and alternatives

This library is a good general purpose solution to adding multi-threaded processing to an existing iterator-based code. When you have a chain of iterator steps, and would like to process one or some of them in parallel to speed things up, this library goes a long way to make it as close to a drop-in replacement as possible in all aspects.

The implementation is based on spawning thread-pools of worker threads and sending them work using channels, then receiving and sorting the results to turn them into a normal iterator again.

Sending iterator items through channels is fast, but not free. Make sure to parallelize operations that are heavy enough to justify overhead of sending data through channels. E.g. operations involving IO or some CPU-heavy computation.

You can use cargo bench or view the /docs/bench-report/report/index.html locally for criterion.rs benchmark report, but as a rule of thumb, each call to function being parallized should take more than 200ns for the parallelization to outweight the overheads.

When you have a lot items already stored in a collection, that you want to "roll over and perform some simple computation" you probably want to use rayon instead. It's a library optimized for parallelizing processing of whole chunks of larger set of data, which minimizes any per-item overheads. A downside of that is that converting rayon's iterators back to ordered sequencial iterator is non-trivial.

Usage

Adding new ones based on the existing code should be relatively easy, so PRs are welcome.

In short, if you have:

# fn step_a(x: usize) -> usize {
#   x * 7
# }
# 
# fn filter_b(x: &usize) -> bool {
#   x % 2 == 0
# }
# 
# fn step_c(x: usize) -> usize {
#   x + 1
# }
assert_eq!(
  (0..10)
    .map(step_a)
    .filter(filter_b)
    .map(step_c).collect::<Vec<_>>(),
    vec![1, 15, 29, 43, 57]
);

You can change it to:

use pariter::IteratorExt as _;
# fn step_a(x: usize) -> usize {
#   x * 7
# }
# 
# fn filter_b(x: &usize) -> bool {
#   x % 2 == 0
# }
# 
# fn step_c(x: usize) -> usize {
#   x + 1
# }
assert_eq!(
  (0..10)
    .map(step_a)
    .filter(filter_b)
    .parallel_map(step_c).collect::<Vec<_>>(),
    vec![1, 15, 29, 43, 57]
);

and it will run faster (conditions apply), because step_c will run in parallel on multiple-threads.

Iterating over borrowed values

Hitting a borrowed value does not live long enough error? Looks like you are iterating over values containing borrowed references. Sending them over to different threads for processing could lead to memory unsafety issues. But no problem, we got you covered.

First, if the values you are iterating over can be cheaply cloned, just try adding .cloned() and turning them into owned values.

If you can't, you can use scoped-threads API from crossbeam crate:

use pariter::{IteratorExt as _, scope};
# fn step_a(x: &usize) -> usize {
#   *x * 7
# }
#
# fn filter_b(x: &usize) -> bool {
#   x % 2 == 0
# }
#
# fn step_c(x: usize) -> usize {
#   x + 1
# }
let v : Vec<_> = (0..10).collect();

scope(|scope| {
  assert_eq!(
    v
      .iter() // iterating over `&usize` now, `parallel_map` will not work
      .parallel_map_scoped(scope, step_a)
      .filter(filter_b)
      .map(step_c).collect::<Vec<_>>(),
      vec![1, 15, 29, 43, 57]
  );
});

// or:

assert_eq!(
  scope(|scope| {
  v
    .iter()
    .parallel_map_scoped(scope, step_a)
    .filter(filter_b)
    .map(step_c).collect::<Vec<_>>()}).expect("handle errors properly in production code"),
    vec![1, 15, 29, 43, 57]
);

The additional scope argument comes from crossbeam::thread::scope and is there to enforce memory-safety. Just wrap your iterator chain in a scope wrapper that does not outlive the borrowed value, and everything will work smoothly.

Customizing settings

If you need to change settings like buffer sizes and number of threads:

# use pariter::IteratorExt as _;
assert_eq!(
  (0..10)
    .map(|x| x + 1)
    .parallel_filter_custom(|o| o.threads(16), |x| *x == 5)
    .map(|x| x /2).collect::<Vec<_>>(),
    vec![2]
);

Status & plans

I keep needing this exact functionality, so I've cleaned up my ad-hoc code, put it in a proper library. I'm usually very busy, so if you want something added, please submit a PR.

I'm open to share/transfer ownership & maintenance into reputable hands.

Dependencies