Crate simple_parallel

Straight-forward functions and types for basic data parallel operations.

This library provides a few building blocks for operating on data in parallel, particularly iterators. At the moment, it is not designed to be robust or eke out every last drop of performance, but rather explore some ways in which Rust's type system allows for some fairly fancy things to be written with a guarantee of safety, all without a garbage collector.

The only dependency is std and the basic functionality has no unsafe code at all (the thread pool does use unsafe). Other than the pool, parallelism is built directly from the functionality provided by std::thread and std::sync and leverages their correctness to automatically ensure the correctness of this library (at the memory safety level).

The core design is to simply allow for operations that could occur on a single thread to execute on many, it is not intending to serve as a hard boundary between threads; in particular, if something (a panic!) would take down the main thread when run sequentially, it will also take down the main thread (eventually) when run using the functions in this library.

On the point of performance and robustness, the top level functions do no thread pooling and so everything essentially spawns a new thread for each element, which is definitely suboptimal for many reasons. Fortunately, not all is lost, the functionality is designed to be as generic as possible, so the iterator functions work with many many iterators, e.g. instead of executing a thread on every element of a vector individually, a user can divide that vector into disjoint sections and spread those across much fewer threads (e.g. the chunks method).

Further, the thread pooling that does exist has a lot of synchronisation overhead, and so is actually rarely a performance improvement (although it is a robustness improvement over the top-level functions, since it limits the number of threads that will be spawned).

Either way, this is not recommended for general use.

simple_parallel and stability

This library is built around the same idea as std::thread::scoped, and some functions suffer from its problems. In particular, there are several functions that either use scoped directly, or expose a similar API.

The version of this library usable with Rust 1.0 only exposes Pool, which manually implement the same functionality as std::thread::scoped, but do it in a way that avoids feature gates. However, the map and unordered_map methods have the same unsafety problem and so are marked unsafe (at least until the standard library has a solution).

The APIs that need the scoped functionality are still available only nightly, with the simple_parallel/unstable feature.

Usage

This is available on crates.io. Add this to your Cargo.toml:

[dependencies]
simple_parallel = "0.2"

The latest development version can be obtained on GitHub.

Examples

Initialise an array, in parallel.

let mut data = [0; 10];
// fill the array, with one thread for each element:
simple_parallel::for_(data.iter_mut().enumerate(), |(i, elem)| {
    *elem = i as i32;
});

// now adjust that data, with a threadpool:
let mut pool = simple_parallel::Pool::new(4);
pool.for_(data.iter_mut(), |elem| *elem *= 2);

Transform each element of an ordered map in a fancy way, in parallel, with map (map ensures the output order matches the input order, unlike unordered_map),

#[cfg(feature = "unstable")] fn foo() {
use std::collections::BTreeMap;

let mut map = BTreeMap::new();
map.insert('a', 1);
map.insert('x', 55);

let f = |(&c, &elem): (&char, _)| {
    let mut x = elem  * c as i32;
    // ... something complicated and expensive ...
    return x as f64
};

// (`IntoIterator` is used, so "direct" iteration like this is fine.)
let par_iter = simple_parallel::map(&map, &f);

// the computation is executing on several threads in the
// background, so that elements are hopefully ready as soon as
// possible.

for value in par_iter {
    println!("I computed {}", value);
}

Sum an arbitrarily long slice, in parallel, by summing subsections and adding everything to a shared mutex, stored on the stack of the main thread. (A parallel fold is currently missing, hence the mutex.)

use std::sync::Mutex;

// limit the spew of thread spawning to something sensible
const NUM_CHUNKS: usize = 8;

fn sum(x: &[f64]) -> f64 {
    // (round up)
    let elements_per_chunk = (x.len() + NUM_CHUNKS - 1) / NUM_CHUNKS;

    let total = Mutex::new(0.0);
    simple_parallel::for_(x.chunks(elements_per_chunk), |chunk| {
        // sum up this little subsection
        let subsum = chunk.iter().fold(0.0, |a, b| a + *b);
        *total.lock().unwrap() += subsum;
    });

    let answer = *total.lock().unwrap();
    answer
}

Alternatively, one could use a thread pool, and assign an absolute number of elements to each subsection and let the pool manage distributing the work among threads, instead of being forced to computing the length of the subsections to limit the number of threads spawned.

use std::sync::Mutex;

// limit the spew of thread spawning to something sensible
const ELEMS_PER_JOB: usize = 1_000;

fn pooled_sum(pool: &mut simple_parallel::Pool, x: &[f64]) -> f64 {
    let total = Mutex::new(0.0);
    pool.for_(x.chunks(ELEMS_PER_JOB), |chunk| {
        // sum up this little subsection
        let subsum = chunk.iter().fold(0.0, |a, b| a + *b);
        *total.lock().unwrap() += subsum;
    });

    let answer = *total.lock().unwrap();
    answer
}

A sketch of a very simple recursive parallel merge-sort, using both to handle the recursion. (A working implementation may really need some temporary buffers to mangle the data, but the key point is both naturally running things in parallel.)

/// Merges the two sorted runs `left` and `right`.
/// That is, after `merge(left, right)`,
///
///    left[0] <= left[1] <= ... <= left[last] <= right[0] <= ...
fn merge<T: Ord>(left: &mut [T], right: &mut [T]) {
    // magic (but non-parallel, so boring)
}

fn parallel_merge_sort<T: Ord + Send>(x: &mut [T]) {
   // base case
   if x.len() <= 1 { return }

   // get two disjoint halves of the `x`,
   let half = x.len() / 2;
   let (left, right) = x.split_at_mut(half);
   // and sort them recursively, in parallel
   simple_parallel::both(&mut *left, &mut *right, |v| parallel_merge_sort(v));

   // now combine the two sorted halves
   merge(left, right)
}

The examples folder contains more intricate example(s), such as a parallel fast Fourier transform implementation (it really works, and the parallelism does buy something... when tuned).

Reexports

pub use pool::Pool;

Modules

pool