Permalink
402 lines (327 sloc) 16.5 KB

Rayon

Join the chat at https://gitter.im/rayon-rs/Lobby

Build Status

Rayon is a data-parallelism library for Rust. It is extremely lightweight and makes it easy to convert a sequential computation into a parallel one. It also guarantees data-race freedom. (You may also enjoy this blog post about Rayon, which gives more background and details about how it works, or this video, from the Rust Belt Rust conference.) Rayon is available on crates.io, and API Documentation is available on docs.rs.

You can use Rayon in two ways. Which way you will want will depend on what you are doing:

  • Parallel iterators: convert iterator chains to execute in parallel.
  • The join method: convert recursive, divide-and-conquer style problems to execute in parallel.

No matter which way you choose, you don't have to worry about data races: Rayon statically guarantees data-race freedom. For the most part, adding calls to Rayon should not change how your programs works at all, in fact. However, if you operate on mutexes or atomic integers, please see the notes on atomicity.

Rayon currently requires rustc 1.12.0 or greater.

Contribution

Rayon is an open source project! If you'd like to contribute to Rayon, check out the list of "help wanted" issues. These are all (or should be) issues that are suitable for getting started, and they generally include a detailed set of instructions for what to do. Please ask questions if anything is unclear! Also, check out the Guide to Development page on the wiki. Note that all code submitted in PRs to Rayon is assumed to be licensed under Rayon's dual MIT/Apache2 licensing.

Quick demo

To see Rayon in action, check out the rayon-demo directory, which includes a number of demos of code using Rayon. For example, run this command to get a visualization of an nbody simulation. To see the effect of using Rayon, press s to run sequentially and p to run in parallel.

> cd rayon-demo
> cargo +nightly run --release -- nbody visualize

For more information on demos, try:

> cd rayon-demo
> cargo +nightly run --release -- --help

Note: While Rayon is usable as a library with the stable compiler, running demos or executing tests requires nightly Rust.

Parallel Iterators

Rayon supports an experimental API called "parallel iterators". These let you write iterator-like chains that execute in parallel. For example, to compute the sum of the squares of a sequence of integers, one might write:

use rayon::prelude::*;
fn sum_of_squares(input: &[i32]) -> i32 {
    input.par_iter()
         .map(|&i| i * i)
         .sum()
}

Or, to increment all the integers in a slice, you could write:

use rayon::prelude::*;
fn increment_all(input: &mut [i32]) {
    input.par_iter_mut()
         .for_each(|p| *p += 1);
}

To use parallel iterators, first import the traits by adding something like use rayon::prelude::* to your module. You can then call par_iter and par_iter_mut to get a parallel iterator. Like a regular iterator, parallel iterators work by first constructing a computation and then executing it. See the ParallelIterator trait for the list of available methods and more details. (Sorry, proper documentation is still somewhat lacking.)

Using join for recursive, divide-and-conquer problems

Parallel iterators are actually implemented in terms of a more primitive method called join. join simply takes two closures and potentially runs them in parallel. For example, we could rewrite the increment_all function we saw for parallel iterators as follows (this function increments all the integers in a slice):

/// Increment all values in slice.
fn increment_all(slice: &mut [i32]) {
    if slice.len() < 1000 {
        for p in slice { *p += 1; }
    } else {
        let mid_point = slice.len() / 2;
        let (left, right) = slice.split_at_mut(mid_point);
        rayon::join(|| increment_all(left), || increment_all(right));
    }
}

Perhaps a more interesting example is this parallel quicksort:

fn quick_sort<T:PartialOrd+Send>(v: &mut [T]) {
    if v.len() <= 1 {
        return;
    }

    let mid = partition(v);
    let (lo, hi) = v.split_at_mut(mid);
    rayon::join(|| quick_sort(lo), || quick_sort(hi));
}

Note though that calling join is very different from just spawning two threads in terms of performance. This is because join does not guarantee that the two closures will run in parallel. If all of your CPUs are already busy with other work, Rayon will instead opt to run them sequentially. The call to join is designed to have very low overhead in that case, so that you can safely call it even with very small workloads (as in the example above).

However, in practice, the overhead is still noticeable. Therefore, for maximal performance, you want to have some kind of sequential fallback once your problem gets small enough. The parallel iterator APIs try to handle this for you. When using join, you have to code it yourself. For an example, see the quicksort demo, which includes sequential fallback after a certain size.

Safety

You've probably heard that parallel programming can be the source of bugs that are really hard to diagnose. That is certainly true! However, thanks to Rust's type system, you basically don't have to worry about that when using Rayon. The Rayon APIs are guaranteed to be data-race free. The Rayon APIs themselves also cannot cause deadlocks (though if your closures or callbacks use locks or ports, those locks might trigger deadlocks).

For example, if you write code that tries to process the same mutable state from both closures, you will find that fails to compile:

/// Increment all values in slice.
fn increment_all(slice: &mut [i32]) {
    rayon::join(|| process(slice), || process(slice));
}

However, this safety does have some implications. You will not be able to use types which are not thread-safe (i.e., do not implement Send) from inside a join closure. Note that almost all types are in fact thread-safe in Rust; the only exception is those types that employ "inherent mutability" without some form of synchronization, such as RefCell or Rc. Here is a list of the most common types in the standard library that are not Send, along with an alternative that you can use instead which is Send (but which also has higher overhead, because it must work across threads):

  • Cell -- replacement: AtomicUsize, AtomicBool, etc (but see warning below)
  • RefCell -- replacement: RwLock, or perhaps Mutex (but see warning below)
  • Rc -- replacement: Arc

However, if you are converting uses of Cell or RefCell, you must be prepared for other threads to interject changes. For more information, read the section on atomicity below.

How it works: Work stealing

Behind the scenes, Rayon uses a technique called work stealing to try and dynamically ascertain how much parallelism is available and exploit it. The idea is very simple: we always have a pool of worker threads available, waiting for some work to do. When you call join the first time, we shift over into that pool of threads. But if you call join(a, b) from a worker thread W, then W will place b into a central queue, advertising that this is work that other worker threads might help out with. W will then start executing a. While W is busy with a, other threads might come along and take b from the queue. That is called stealing b. Once a is done, W checks whether b was stolen by another thread and, if not, executes b itself. If b was stolen, then W can just wait for the other thread to finish. (In fact, it can do even better: it can go try to find other work to steal in the meantime.)

This technique is not new. It was first introduced by the Cilk project, done at MIT in the late nineties. The name Rayon is an homage to that work.

Warning: Be wary of atomicity

Converting a Cell (or, to a lesser extent, a RefCell) to work in parallel merits special mention for a number of reasons. Cell and RefCell are handy types that permit you to modify data even when that data is shared (aliased). They work somewhat differently, but serve a common purpose:

  1. A Cell offers a mutable slot with just two methods, get and set. Cells can only be used for Copy types that are safe to memcpy around, such as i32, f32, or even something bigger like a tuple of (usize, usize, f32).
  2. A RefCell is kind of like a "single-threaded read-write lock"; it can be used with any sort of type T. To gain access to the data inside, you call borrow or borrow_mut. Dynamic checks are done to ensure that you have either readers or writers but not both.

While there are threadsafe types that offer similar APIs, caution is warranted because, in a threadsafe setting, other threads may "interject" modifications in ways that are not possible in sequential code. While this will never lead to a data race --- that is, you need not fear undefined behavior --- you can certainly still have bugs.

Let me give you a concrete example using Cell. A common use of Cell is to implement a shared counter. In that case, you would have something like counter: Rc<Cell<usize>>. Now I can increment the counter by calling get and set as follows:

let value = counter.get();
counter.set(value + 1);

If I convert this to be a thread-safe counter, I would use the corresponding types tscounter: Arc<AtomicUsize>. If I then were to convert the Cell API calls directly, I would do something like this:

let value = tscounter.load(Ordering::SeqCst);
tscounter.store(value + 1, Ordering::SeqCst);

You can already see that the AtomicUsize API is a bit more complex, as it requires you to specify an ordering. (I won't go into the details on ordering here, but suffice to say that if you don't know what an ordering is, and probably even if you do, you should use Ordering::SeqCst.) The danger in this parallel version of the counter is that other threads might be running at the same time and they could cause our counter to get out of sync. For example, if we have two threads, then they might both execute the "load" before either has a chance to execute the "store":

Thread 1                                          Thread 2
let value = tscounter.load(Ordering::SeqCst);
// value = X                                      let value = tscounter.load(Ordering::SeqCst);
                                                  // value = X
tscounter.store(value+1);                         tscounter.store(value+1);
// tscounter = X+1                                // tscounter = X+1

Now even though we've had two increments, we'll only increase the counter by one! Even though we've got no data race, this is still probably not the result we wanted. The problem here is that the Cell API doesn't make clear the scope of a "transaction" -- that is, the set of reads/writes that should occur atomically. In this case, we probably wanted the get/set to occur together.

In fact, when using the Atomic types, you very rarely want a plain load or plain store. You probably want the more complex operations. A counter, for example, would use fetch_add to atomically load and increment the value in one step. Compare-and-swap is another popular building block.

A similar problem can arise when converting RefCell to RwLock, but it is somewhat less likely, because the RefCell API does in fact have a notion of a transaction: the scope of the handle returned by borrow or borrow_mut. So if you convert each call to borrow to read (and borrow_mut to write), things will mostly work fine in a parallel setting, but there can still be changes in behavior. Consider using a handle: RefCell<Vec<i32>> like :

let len = handle.borrow().len();
for i in 0 .. len {
    let data = handle.borrow()[i];
    println!("{}", data);
}

In sequential code, we know that this loop is safe. But if we convert this to parallel code with an RwLock, we do not: this is because another thread could come along and do handle.write().unwrap().pop(), and thus change the length of the vector. In fact, even in sequential code, using very small borrow sections like this is an anti-pattern: you ought to be enclosing the entire transaction together, like so:

let vec = handle.borrow();
let len = vec.len();
for i in 0 .. len {
    let data = vec[i];
    println!("{}", data);
}

Or, even better, using an iterator instead of indexing:

let vec = handle.borrow();
for data in vec {
    println!("{}", data);
}

There are several reasons to prefer one borrow over many. The most obvious is that it is more efficient, since each borrow has to perform some safety checks. But it's also more reliable: suppose we modified the loop above to not just print things out, but also call into a helper function:

let vec = handle.borrow();
for data in vec {
    helper(...);
}

And now suppose, independently, this helper fn evolved and had to pop something off of the vector:

fn helper(...) {
    handle.borrow_mut().pop();
}

Under the old model, where we did lots of small borrows, this would yield precisely the same error that we saw in parallel land using an RwLock: the length would be out of sync and our indexing would fail (note that in neither case would there be an actual data race and hence there would never be undefined behavior). But now that we use a single borrow, we'll see a borrow error instead, which is much easier to diagnose, since it occurs at the point of the borrow_mut, rather than downstream. Similarly, if we move to an RwLock, we'll find that the code either deadlocks (if the write is on the same thread as the read) or, if the write is on another thread, works just fine. Both of these are preferable to random failures in my experience.

But wait, isn't Rust supposed to free me from this kind of thinking?

You might think that Rust is supposed to mean that you don't have to think about atomicity at all. In fact, if you avoid inherent mutability (Cell and RefCell in a sequential setting, or AtomicUsize, RwLock, Mutex, et al. in parallel code), then this is true: the type system will basically guarantee that you don't have to think about atomicity at all. But often there are times when you WANT threads to interleave in the ways I showed above.

Consider for example when you are conducting a search in parallel, say to find the shortest route. To avoid fruitless search, you might want to keep a cell with the shortest route you've found thus far. This way, when you are searching down some path that's already longer than this shortest route, you can just stop and avoid wasted effort. In sequential land, you might model this "best result" as a shared value like Rc<Cell<usize>> (here the usize represents the length of best path found so far); in parallel land, you'd use a Arc<AtomicUsize>. Now we can make our search function look like:

fn search(path: &Path, cost_so_far: usize, best_cost: &Arc<AtomicUsize>) {
    if cost_so_far >= best_cost.load(Ordering::SeqCst) {
        return;
    }
    ...
    best_cost.store(...);
}

Now in this case, we really WANT to see results from other threads interjected into our execution!

License

Rayon is distributed under the terms of both the MIT license and the Apache License (Version 2.0). See LICENSE-APACHE for LICENSE-MIT for details. Opening a pull requests is assumed to signal agreement with these licensing terms.