Tokio Async: Concurrent vs Parallel

main.rs

use futures::StreamExt;
use std::error::Error;
use tokio;
use tokio::macros::support::Pin;
use tokio::prelude::*;
use tokio::time::{Duration, Instant};

pub fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut multi_threaded_runtime = tokio::runtime::Builder::new()
        .threaded_scheduler()
        .enable_all()
        .core_threads(10)
        .max_threads(10)
        .thread_name("multi-threaded")
        .build()?;

    // multi_threaded_runtime.block_on(concurrent());
    multi_threaded_runtime.block_on(concurrentAndParallel());
    Ok(())
}

// As can be seen by the output below, despite specifying parallelism of 3, we are still bound by
// the fact that all of the future generated by this function execute within a single task.
async fn concurrent() {
    let before = Instant::now();

    let paths = (0..6).rev();

    let fetches = futures::stream::iter(paths.into_iter().map(|path| make_request(path)))
        .buffer_unordered(3)
        .map(|r| println!("finished request: {}", r))
        .collect::<Vec<_>>();
    fetches.await;

    println!("elapsed time: {:.2?}", before.elapsed());
}
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 5
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 4
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 3
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 2
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 1
// started request
// finished request: current thread ThreadId(1) | thread name main | request_duration 0
// elapsed time: 15.01s

// Here, we wrap every future within its own task using tokio::spawn. This allows the "requests" to
// execute in parallel (depending on how many threads the runtime is configured with; 10 in this
// case) using the multiplexing that Tokio does between different tasks and threads. You can see
// from the output how 3 threads with ids, 9, 10, and 11, are consistently used to execute all of
// the 6 "requests".
async fn concurrentAndParallel() {
    let before = Instant::now();

    let paths = (0..6).rev();

    let fetches = futures::stream::iter(
        paths
            .into_iter()
            .map(|path| tokio::spawn(make_request(path))),
    )
    .buffer_unordered(3)
    .map(|r| {
        println!(
            "finished request: {}",
            match r {
                Ok(rr) => rr,
                Err(_) => String::from("Bad"),
            }
        );
    })
    .collect::<Vec<_>>();
    fetches.await;

    println!("elapsed time: {:.2?}", before.elapsed());
}
// started request
// started request
// started request
// finished request: current thread ThreadId(9) | thread name multi-threaded | request_duration 3
// started request
// finished request: current thread ThreadId(10) | thread name multi-threaded | request_duration 4
// started request
// finished request: current thread ThreadId(11) | thread name multi-threaded | request_duration 5
// started request
// finished request: current thread ThreadId(11) | thread name multi-threaded | request_duration 0
// finished request: current thread ThreadId(10) | thread name multi-threaded | request_duration 1
// finished request: current thread ThreadId(9) | thread name multi-threaded | request_duration 2
// elapsed time: 5.01s

async fn make_request(sleep: u64) -> String {
    println!("started request");

    std::thread::sleep(Duration::from_secs(sleep));

    format!(
        "current thread {:?} | thread name {} | request_duration {:?}",
        std::thread::current().id(),
        std::thread::current()
            .name()
            .get_or_insert("default_thread_name"),
        sleep
    )
}

The documentation for std::thread::sleep says: This function is blocking, and should not be used in async functions. I assume that is the reason you need to run it parallel and that the concurrent version is slower.

Right. I just wanted to use something that blocked to show how you can effectively parallelize work by using multiple tasks (unlike with the concurrent example which uses a single task for everything)