par_iter_sync: Parallel Iterator With Sequential Output

Crate like rayon do not offer synchronization mechanism. This crate provides easy mixture of parallelism and synchronization.

Consider the case where multiple threads share a cache which can be read only after prior tasks have written to it (e.g., reads of task 4 depends on writes of task 1-4).

Using IntoParallelIteratorSync trait

// in concurrency: task1 write | task2 write | task3 write | task4 write
//                      \_____________\_____________\_____________\
//             task4 read depends on task 1-4 write  \___________
//                                                               \
// in concurrency:              | task2 read  | task3 read  | task4 read

use par_iter_sync::IntoParallelIteratorSync;
use std::sync::{Arc, Mutex};
use std::collections::HashSet;

// there are 100 tasks
let tasks = 0..100;

// an in-memory cache for integers
let cache: Arc<Mutex<HashSet<i32>>> = Arc::new(Mutex::new(HashSet::new()));
let cache_clone = cache.clone();

// iterate through tasks
tasks.into_par_iter_sync(move |task_number| {

    // writes cache (write the integer in cache), in parallel
    cache.lock().unwrap().insert(task_number);
    // return the task number to the next iterator
    Ok(task_number)

}).into_par_iter_sync(move |task_number| { // <- synced to sequential order

    // reads
    assert!(cache_clone.lock().unwrap().contains(&task_number));
    Ok(())

// append a for each to actually run the whole chain
}).for_each(|_| ());

Usage Caveat

This crate is designed to clone all resources captured by the closure for each thread. To prevent unintended RAM usage, you may wrap large data structure using Arc (especially vectors of Clone objects).

Sequential Consistency

The output order is guaranteed to be the same as the upstream iterator, but the execution order is not sequential.

Overhead Benchmark

Platform: Macbook Air (2015 Late) 8 GB RAM, Intel Core i5, 1.6GHZ (2 Core).

Result

One million (1,000,000) empty iteration for each run.

test iter_async::test_par_iter_async::bench_into_par_iter_async
    ... bench: 110,277,577 ns/iter (+/- 28,510,054)

test test_par_iter::bench_into_par_iter_sync
    ... bench: 121,063,787 ns/iter (+/- 103,787,056)

Result:

• Async iterator overhead 110 ns (+/- 28 ns).
• Sync iterator overhead 121 ns (+/- 103 ns).

Bench Programs

iter_async

    #[bench]
    fn bench_into_par_iter_async(b: &mut Bencher) {
        b.iter(|| {
            (0..1_000_000).into_par_iter_async(|a| Ok(a)).for_each(|_|{})
        });
    }

iter_sync

    #[bench]
    fn bench_into_par_iter_sync(b: &mut Bencher) {
        b.iter(|| {
            (0..1_000_000).into_par_iter_sync(|a| Ok(a)).for_each(|_|{})
        });
    }

Examples

Mix Syncing and Parallelism By Chaining

use par_iter_sync::IntoParallelIteratorSync;

(0..100).into_par_iter_sync(|i| {
    Ok(i)                     // <~ async execution
}).into_par_iter_sync(|i| { // <- sync order
    Ok(i)                     // <~async execution
}).into_par_iter_sync(|i| { // <- sync order
    Ok(i)                     // <~async execution
});                           // <- sync order

Use std::iter::IntoIterator interface

use par_iter_sync::IntoParallelIteratorSync;

let mut count = 0;

// for loop
for i in (0..100).into_par_iter_sync(|i| Ok(i)) {
    assert_eq!(i, count);
    count += 1;
}

// sum
let sum: i32 = (1..=100).into_par_iter_sync(|i| Ok(i)).sum();

// take and collect
let results: Vec<i32> = (0..10).into_par_iter_sync(|i| Ok(i)).take(5).collect();

assert_eq!(sum, 5050);
assert_eq!(results, vec![0, 1, 2, 3, 4])

Closure Captures Variables

Variables captured are cloned to each thread automatically.

use par_iter_sync::IntoParallelIteratorSync;
use std::sync::Arc;

// use `Arc` to save RAM
let resource_captured = Arc::new(vec![3, 1, 4, 1, 5, 9, 2, 6, 5, 3]);
let len = resource_captured.len();

let result_iter = (0..len).into_par_iter_sync(move |i| {
    // `resource_captured` is moved into the closure
    // and cloned to worker threads.
    let read_from_resource = resource_captured.get(i).unwrap();
    Ok(*read_from_resource)
});

// the result is produced in sequential order
let collected: Vec<i32> = result_iter.collect();
assert_eq!(collected, vec![3, 1, 4, 1, 5, 9, 2, 6, 5, 3])

Fast Fail During Exception

The iterator stops once the inner function returns an Err.

use par_iter_sync::IntoParallelIteratorSync;
use std::sync::Arc;
use log::warn;

/// this function returns `Err` when it reads 1000
fn error_at_1000(n: i32) -> Result<i32, ()> {
    if n == 1000 {
        // you may log this error
        warn!("Some Error Occurs");
        Err(())
    } else {
        Ok(n)
    }
}

let results: Vec<i32> = (0..10000).into_par_iter_sync(move |a| {
    Ok(a)
}).into_par_iter_sync(move |a| {
    // error at 1000
    error_at_1000(a)
}).into_par_iter_sync(move |a| {
    Ok(a)
}).collect();

let expected: Vec<i32> = (0..1000).collect();
assert_eq!(results, expected)

You may choose to skip error

If you do not want to stop on Err, this is a workaround.

use par_iter_sync::IntoParallelIteratorSync;
use std::sync::Arc;

let results: Vec<Result<i32, ()>> = (0..5).into_par_iter_sync(move |n| {
    // error at 3, but skip
    if n == 3 {
        Ok(Err(()))
    } else {
        Ok(Ok(n))
    }
}).collect();

assert_eq!(results, vec![Ok(0), Ok(1), Ok(2), Err(()), Ok(4)])

Implementation Note

Output Buffering

• Each worker use a synced single-producer mpsc channel to buffer outputs. So, when a thread is waiting for its turn to get polled, it does not get blocked. The channel size is hard-coded to 100 for each thread.
• The number of threads equals to the number of logical cores.

Synchronization Mechanism

• When each thread fetch a task, it registers its thread ID (thread_number) and the task ID (task_number) into a mpsc channel.
• When next() is called, the consumer fetch from the task registry (task_order) the next thread ID and task ID.
• If next() detect that some thread has not produced result due to exception, it calls kill(), which stop threads from fetching new tasks, flush remaining tasks, and joining the worker threads.

Error handling and Dropping

• When any exception occurs, stop producers from fetching new task.
• Before dropping the structure, stop all producers from fetching tasks, flush all remaining tasks, and join all threads.