James McMurray's Blog

Rust, Linux and other curiosities

A practical introduction to async programming in Rust


In this post we will explore a brief example of asynchronous programming in Rust with the Tokio runtime, demonstrating different execution scenarios. This post is aimed at beginners to asynchronous programming.

The source code for this example is available on Github. A branch using the async-std runtime is also available (contributed by @BartMassey).

What is asynchronous programming?

Asynchronous programming allows you to continue carrying out computations whilst waiting for results from IO operations (often network requests or responses), even on a single OS thread.

This is achieved via the use of an async runtime which handles the assignment of async tasks (i.e. green threads) to actual OS threads.

Unlike OS threads, green threads are not expensive to create and so we needn't worry about hitting a hard limit. Whereas OS threads need to hold their own stack, leading to high memory usage when dealing with many threads. On Linux you can check your thread limit per process with: cat /proc/sys/kernel/threads-max, mine is 127,162.

This is a significant issue if we need a separate OS thread to handle each request on a web server for example, and is the origin of the C10k problem - how to handle 10,000 concurrent connections to a web server.

Early web servers did indeed use a separate OS thread per request, in order to handle requests in parallel. The key insight is that those threads spend most of their time waiting for network responses, rather than doing any computation.

Async and await

Rust has adopted the async/await syntax for defining asynchronous code blocks and functions.

The async keyword defines an async code block or function. Specifying that it will return a Future, a value that will need to be .awaited elsewhere in order to trigger the execution of the task (note the lazy execution) and wait for the return value to be available.

The .await keyword (which must be inside an async block/function) is used to wait asynchronously for an async task to finish, and get the return value. Note that while the task itself cannot progress until the Future is ready, the actual OS thread can have other tasks assigned to it by the runtime, and so continue to do work.

Effectively the task is informing the runtime that at this point it may yield the execution to another task (eventually, that other task will also await something, and if the Future in this task is ready, execution of this task may continue) - this is an implementation of co-operative multitasking.

This syntax is very elegant, and allows us to write asynchronous programs that maintain a structure similar to simple, synchronous programs.

When should I use async?

Asynchronous programming is useful when the thread would otherwise just be waiting on IO operations - for example when making network requests or responses, reading or writing to disk, or waiting for user input.

It is not useful if you are always doing computations and there is no waiting on IO operations, even if those computations could run in parallel - for example in a ray tracer. In this case it would be best to parallelise the computations over OS threads directly (taking advantage of the multiple cores in your CPU), for example using parallel iterators in the rayon crate (or if you want thread-level control, then with the crossbeam and threadpool crates). However, remember Amdahl's Law and consider that algorithmic improvements might yield a better return than focusing on parallelisation in this cases.

It is also not useful if there are no other tasks to do whilst waiting for the IO operations. For example in the previous blog post, Rusoto actually returns a RusotoFuture object when we request the DB credentials from AWS Secrets Manager, however in this case one invocation of our Lambda function corresponds to one request - there is no work to be done whilst waiting for the DB credentials to arrive. So we can just program synchronously (fortunately RusotoFuture provides the .sync() method to do exactly that).

Example

In this example we will simulate three very slow network requests, consisting of three stages:

  • Connection - an asynchronous delay of 2 seconds
  • Waiting for a response - an asynchronous delay of 8 seconds (the delay is on the server side)
  • Computation - a synchronous delay of 4 seconds (i.e. it has to block the current OS thread to do the computation).

We will use Tokio as our async runtime for this example, as it is currently the most popular. The other main alternative is the async_std runtime - code for using this runtime is available in the async-std branch of the Github repo (this code was contributed by @BartMassey).

Note that both use the common futures crate so you can swap the async runtime whilst keeping mostly the same API.

Server

The server used in this example is adapted from the Tokio tutorial. It echoes back the bytes received, after a delay of 8 seconds.

The full adapted code is as follows (and available in the source code for this post):

use futures::stream::StreamExt;
use tokio::net::TcpListener;

#[tokio::main]
async fn main() {
    let addr = "127.0.0.1:6142";
    let mut listener = TcpListener::bind(addr).await.unwrap();

    let server = {
        async move {
            let mut incoming = listener.incoming();
            while let Some(conn) = incoming.next().await {
                match conn {
                    Err(e) => eprintln!("accept failed = {:?}", e),
                    Ok(mut sock) => {
                        tokio::spawn(async move {
                            let (mut reader, mut writer) = sock.split();
                            tokio::time::delay_for(tokio::time::Duration::from_secs(8)).await;
                            match tokio::io::copy(&mut reader, &mut writer).await {
                                Ok(amt) => {
                                    println!("wrote {} bytes", amt);
                                }
                                Err(err) => {
                                    eprintln!("IO error {:?}", err);
                                }
                            }
                        });
                    }
                }
            }
        }
    };
    println!("Server running on localhost:6142");
    server.await;
}

Synchronous requests

In the synchronous case, we simply run each request in series, one after another. Therefore we would expect a total execution time of 3*(2+8+4) = 42 seconds to finish all 3 tasks.

We can visualise this case with a diagram:

Synchronous execution

We can implement this using only the standard library:

use std::io::prelude::*;
use std::net::TcpStream;
use std::thread::sleep;
use std::time::Instant;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let now = Instant::now();

    task("task1", now.clone())?;
    task("task2", now.clone())?;
    task("task3", now.clone())?;
    Ok(())
}

fn task(label: &str, now: std::time::Instant) -> Result<(), Box<dyn std::error::Error>> {
    // Simulate network delay using thread sleep for 2 seconds
    println!(
        "OS Thread {:?} - {} started: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed(),
    );
    sleep(std::time::Duration::from_secs(2));

    // Write to server - server will echo this back to us with 8 second delay
    let mut stream = TcpStream::connect("127.0.0.1:6142")?;
    stream.write_all(label.as_bytes())?;
    println!(
        "OS Thread {:?} - {} written: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed()
    );

    // Read 5 chars we expect (to avoid dealing with EOF, etc.)
    let mut buffer = [0; 5];
    stream.read_exact(&mut buffer)?;
    stream.shutdown(std::net::Shutdown::Both)?;
    println!(
        "OS Thread {:?} - {} read: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed()
    );

    // Simulate computation work by sleeping actual thread for 4 seconds
    sleep(std::time::Duration::from_secs(4));
    println!(
        "OS Thread {:?} - {} finished: {:?}",
        std::thread::current().id(),
        std::str::from_utf8(&buffer)?,
        now.elapsed()
    );
    Ok(())
}

Running this (see the repo for this workspace):

$ cargo run --release --bin server
$ cargo run --release --bin client_synchronous
OS Thread ThreadId(1) - task1 started: 578ns
OS Thread ThreadId(1) - task1 written: 2.000346788s
OS Thread ThreadId(1) - task1 read: 10.002177173s
OS Thread ThreadId(1) - task1 finished: 14.002328699s
OS Thread ThreadId(1) - task2 started: 14.002387112s
OS Thread ThreadId(1) - task2 written: 16.002673602s
OS Thread ThreadId(1) - task2 read: 24.006071003s
OS Thread ThreadId(1) - task2 finished: 28.006204147s
OS Thread ThreadId(1) - task3 started: 28.006263855s
OS Thread ThreadId(1) - task3 written: 30.00652763s
OS Thread ThreadId(1) - task3 read: 38.008234993s
OS Thread ThreadId(1) - task3 finished: 42.008389223s

Gives exactly the 42 seconds total execution time that we calculated above.

Synchronous requests (Tokio)

Note that it is possible to get synchronous behaviour from async functions when using Tokio (sometimes unintentionally). Implementing the above with Tokio:

use futures::stream::StreamExt;
use std::error::Error;
use std::thread::sleep;
use std::time::Instant;
use tokio::join;
use tokio::net::TcpStream;
use tokio::prelude::*;

#[tokio::main]
async fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
    let now = Instant::now();

    // Synchronous
    task("task1", now.clone()).await?;
    task("task2", now.clone()).await?;
    task("task3", now.clone()).await?;
    Ok(())
}

async fn task(label: &str, now: std::time::Instant) -> Result<(), Box<dyn Error + Send + Sync>> {
    // Simulate network delay using Tokio async delay for 2 seconds
    println!(
        "OS Thread {:?} - {} started: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed(),
    );
    tokio::time::delay_for(tokio::time::Duration::from_secs(2)).await;

    // Write to server - server will echo this back to us with 8 second delay
    let mut stream = TcpStream::connect("127.0.0.1:6142").await?;
    stream.write_all(label.as_bytes()).await?;
    println!(
        "OS Thread {:?} - {} written: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed()
    );

    // Read 5 chars we expect (to avoid dealing with EOF, etc.)
    let mut buffer = [0; 5];
    stream.read_exact(&mut buffer).await?;
    stream.shutdown(std::net::Shutdown::Both)?;
    println!(
        "OS Thread {:?} - {} read: {:?}",
        std::thread::current().id(),
        label,
        now.elapsed()
    );

    // Simulate computation work by sleeping actual thread for 4 seconds
    sleep(std::time::Duration::from_secs(4));
    println!(
        "OS Thread {:?} - {} finished: {:?}",
        std::thread::current().id(),
        std::str::from_utf8(&buffer)?,
        now.elapsed()
    );
    Ok(())
}

Running this produces the same output as before:

$ cargo run --release --bin client_async
OS Thread ThreadId(1) - task1 started: 333ns
OS Thread ThreadId(1) - task1 written: 2.001476012s
OS Thread ThreadId(1) - task1 read: 10.003284491s
OS Thread ThreadId(1) - task1 finished: 14.003404307s
OS Thread ThreadId(1) - task2 started: 14.003476979s
OS Thread ThreadId(1) - task2 written: 16.005013941s
OS Thread ThreadId(1) - task2 read: 24.005471439s
OS Thread ThreadId(1) - task2 finished: 28.005575307s
OS Thread ThreadId(1) - task3 started: 28.005615372s
OS Thread ThreadId(1) - task3 written: 30.007082377s
OS Thread ThreadId(1) - task3 read: 38.009223127s
OS Thread ThreadId(1) - task3 finished: 42.009349576s

It is the .awaiting of the tasks in series which causes this to be synchronous. The main function is async, but the use of .await causes it to wait for the result of that Future before continuing. There is nothing special about the main function compared to other async functions in this regard. No other tasks exist at that time to yield execution to, causing the execution to be synchronous in practice.

Note the Send + Sync bounds are not required for the above implementation (since it runs on only one OS thread), but we will need them in the last example. This is also why we clone now instead of borrowing it in task() (alternatively we could wrap it in an Arc).

We will use the same async fn task() definition for the following examples, where it will be omitted.

Asynchronous requests (one OS thread)

In the asynchronous, single OS thread case we start the waiting steps (connection and getting the server response) concurrently. However, the final computation step will still need to be done in series for each task. Therefore we expect a total execution time of 8+2+(3*4)=22 seconds.

In a diagram:

Asynchronous execution on a single OS thread

Using the same definition of async fn task() as before:

use futures::stream::futures_unordered::FuturesUnordered;
use futures::stream::StreamExt;
use std::error::Error;
use std::thread::sleep;
use std::time::Instant;
use tokio::net::TcpStream;
use tokio::prelude::*;

#[tokio::main]
async fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
    let now = Instant::now();

    // Asynchronous single-thread
    let mut futs = FuturesUnordered::new();

    futs.push(task("task1", now.clone()));
    futs.push(task("task2", now.clone()));
    futs.push(task("task3", now.clone()));

    while let Some(_handled) = futs.next().await {}
    Ok(())
}

Running this gives our expected total execution time of 22 seconds:

OS Thread ThreadId(1) - task1 started: 3.994µs
OS Thread ThreadId(1) - task2 started: 21.174µs
OS Thread ThreadId(1) - task3 started: 25.511µs
OS Thread ThreadId(1) - task3 written: 2.002221984s
OS Thread ThreadId(1) - task2 written: 2.002406898s
OS Thread ThreadId(1) - task1 written: 2.002483563s
OS Thread ThreadId(1) - task3 read: 10.003326999s
OS Thread ThreadId(1) - task3 finished: 14.003478669s
OS Thread ThreadId(1) - task2 read: 14.00365763s
OS Thread ThreadId(1) - task2 finished: 18.00379238s
OS Thread ThreadId(1) - task1 read: 18.003951713s
OS Thread ThreadId(1) - task1 finished: 22.004094444s

In this case we use a FuturesUnordered collection to allow us to repeatedly await the different Futures. However, we never use tokio::spawn() so it is only running on the single OS thread (as we never allow the creation of more).

Note that we could use the join! macro here instead of allocating a FuturesUnordered, we will see an example of this later. However, that is only practical for a small number of Futures.

We can also force Tokio to only use one thread by putting arguments in the attribute for our main function:

#[tokio::main(core_threads = 1, max_threads = 1)]
async fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
...
}

Asynchronous requests (multiple OS threads)

In the case of asynchronous requests across multiple OS threads, we can do every step concurrently (and the OS threads could complete other tasks when the tasks await). This means that we can do the final computation step in parallel on different OS threads.

Asynchronous execution on multiple OS threads

Therefore we would expect a total execution time of 2+8+4=14 seconds for all 3 requests. This is the best we can achieve - the same time as completing a single request.

Note that this requires the types we send across threads to be thread-safe, i.e. implementing Send or Sync - just as we would if we were using OS threads directly.

The implementation is very similar to the previous example, but instead of awaiting on the futures directly, we tokio::spawn the tasks and await their handles. This allows Tokio to execute them on different OS threads.

#[tokio::main]
async fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
    let now = Instant::now();

    let mut futs = FuturesUnordered::new();
    futs.push(tokio::spawn(task("task1", now.clone())));
    futs.push(tokio::spawn(task("task2", now.clone())));
    futs.push(tokio::spawn(task("task3", now.clone())));
    while let Some(_handled) = futs.next().await {}
    Ok(())
}

And we observe the 14 second execution time (note we don't care about the order of execution):

OS Thread ThreadId(2) - task1 started: 17.055µs
OS Thread ThreadId(3) - task2 started: 30.227µs
OS Thread ThreadId(2) - task3 started: 32.513µs
OS Thread ThreadId(2) - task3 written: 2.001499145s
OS Thread ThreadId(3) - task1 written: 2.00153689s
OS Thread ThreadId(5) - task2 written: 2.001721878s
OS Thread ThreadId(3) - task3 read: 10.003403756s
OS Thread ThreadId(2) - task1 read: 10.003501s
OS Thread ThreadId(5) - task2 read: 10.003417328s
OS Thread ThreadId(3) - task3 finished: 14.003584085s
OS Thread ThreadId(2) - task1 finished: 14.003664981s
OS Thread ThreadId(5) - task2 finished: 14.003698375s

The different OS Thread IDs demonstrate that the tasks are being executed on separate OS threads.

For completeness, here is the same implementation using the join! macro instead of allocating a FuturesUnordered:

#[tokio::main]
async fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
    let now = Instant::now();
    // Asynchronous multi-threaded

    match join!(
        tokio::spawn(task("task1", now.clone())),
        tokio::spawn(task("task2", now.clone())),
        tokio::spawn(task("task3", now.clone()))
    ) {
        (x, y, z) => {
            (x.ok(), y.ok(), z.ok())
        }
    };
    Ok(())
}

This saves the allocation of the FuturesUnordered but it can be awkward to work with the tuple of Results that is returned (especially for many Futures).

How does this differ from OS thread parallelisation?

You could implement something similar for the synchronous case using OS threads directly, for example with the rayon crate mentioned previously.

However, in this case each request would need its own OS thread, and if we had to carry out 10,000 concurrent requests, we might hit the thread limit.

That is, the execution diagram would look like this:

Synchronous execution on multiple OS threads

Note that the OS threads would spend a lot of time just waiting on IO operations, unable to start other requests.

We can modify our first synchronous example above to do this with rayon:

use rayon::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    let now = Instant::now();

    ["task1", "task2", "task3"]
        .par_iter()
        .map(|x| task(x, now.clone()))
        .collect::<Result<Vec<_>, _>>()?;
    Ok(())
}

And it terminates in 14 seconds as expected (note that each task is allocated to one OS thread):

OS Thread ThreadId(3) - task1 started: 280.871µs
OS Thread ThreadId(6) - task2 started: 281.03µs
OS Thread ThreadId(7) - task3 started: 283.838µs
OS Thread ThreadId(6) - task2 written: 2.000605562s
OS Thread ThreadId(7) - task3 written: 2.000619598s
OS Thread ThreadId(3) - task1 written: 2.000679853s
OS Thread ThreadId(3) - task1 read: 10.002321036s
OS Thread ThreadId(6) - task2 read: 10.00233185s
OS Thread ThreadId(7) - task3 read: 10.002384653s
OS Thread ThreadId(3) - task1 finished: 14.002447762s
OS Thread ThreadId(6) - task2 finished: 14.002540969s
OS Thread ThreadId(7) - task3 finished: 14.002589621s

However, this would not be resource-efficient when handling a large number of tasks since one task corresponds to one OS thread.

Whereas in the async case we only need the additional threads during the computation step (which is synchronous). This means that we could use a fixed OS threadpool size and still benefit from some parallelisation in the computation step whilst having guarantees about our resource usage (i.e. we can limit the maximum number of OS threads and still be able to start new requests).

Conclusion

I hope this blog post has helped you to better understand when and how to apply asynchronous programming in Rust.

The elegant async/await syntax allows for clear and concise asynchronous programming. However, it may take some getting used to if you do not have prior experience with asynchronous programming.

These examples also demonstrate the difference between concurrency and parallelism. In the asynchronous single OS thread case we are handling tasks concurrently, but nothing is running in parallel (i.e. we use only one OS thread).

Current Limitations

Note that currently you cannot use async methods in traits, and nor can you create async destructors. This is a problem if you want a struct to make a network request when it is dropped, but you don't want to block the OS thread whilst it does that - you cannot .await because drop() (from Drop) is not async.

I hit the latter issue in my recent s3rename crate, and am still attempting to work around it. See this issue for details.