orx-concurrent-iter

A thread-safe, ergonomic and lightweight concurrent iterator trait and efficient implementations.

• ergonomic: An iterator implementing ConcurrentIter can safely be shared among threads. It may be iterated over concurrently by multiple threads with for or while let. It further provides higher level methods such as for_each and fold which allow for safe, simple and efficient parallelism.
• efficient and lightweight: All concurrent iterator implementations provided in this crate extend atomic iterators which are lock-free and depend only on atomic primitives.

Examples

Basic Usage

A ConcurrentIter can be safely shared among threads and iterated over concurrently. As expected, it will yield each element only once and in order. The yielded elements will be shared among the threads which concurrently iterates based on first come first serve. In other words, threads concurrently pull remaining elements from the iterator.

use orx_concurrent_iter::*;
use std::fmt::Debug;

fn fake_work<T: Debug>(_x: T) {
    std::thread::sleep(std::time::Duration::from_nanos(10));
}

/// `process` elements of `iter` concurrently using `num_threads` threads
fn process_concurrently<T, I, F>(process: &F, num_threads: usize, iter: I)
where
    T: Send + Sync,
    F: Fn(T) + Send + Sync,
    I: ConcurrentIter<Item = T>,
{
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                // concurrently iterate over values in a `for` loop
                for value in iter.values() {
                    process(value);
                }
            });
        }
    });
}

/// executes `fake_work` concurrently on all elements of the `concurrent_iter`
fn run<T: Send + Sync + Debug>(concurrent_iter: impl ConcurrentIter<Item = T>) {
    process_concurrently(&fake_work, 8, concurrent_iter)
}

// non-consuming iteration over references
let names: [String; 3] = [
    String::from("foo"),
    String::from("bar"),
    String::from("baz"),
];
run::<&String>(names.con_iter());

let values: Vec<i32> = (0..8).map(|x| 3 * x + 1).collect();
run::<&i32>(values.con_iter());

let slice: &[i32] = values.as_slice();
run::<&i32>(slice.con_iter());
run::<i32>(slice.con_iter().cloned());

// consuming iteration over values
run::<String>(names.into_con_iter());
run::<i32>(values.into_con_iter());

// any Iterator into ConcurrentIter
let values: Vec<i32> = (0..1024).collect();

let evens = values.iter().filter(|x| *x % 2 == 0);
run::<&i32>(evens.into_con_iter());

let evens = values.iter().filter(|x| *x % 2 == 0);
run::<i32>(evens.into_con_iter().cloned());

let iter_val = values
    .iter()
    .filter(|x| *x % 2 == 0)
    .map(|x| (7 * x + 3) as usize)
    .skip(2)
    .take(5);
run::<usize>(iter_val.into_con_iter());

Ways to Loop

ConcurrentIters implement the next method, which is the concurrent counterpart of Iterator::next. Therefore, the iterator can be used almost the same as a regular Iterator safely across multiple threads. Slight difference of different ways to iterate over a ConcurrentIter are demonstrated and explained in the following example.

use orx_concurrent_iter::*;
use std::fmt::Debug;

fn process_one_by_one<T, I, F>(process: &F, num_threads: usize, iter: &I)
where
    T: Send + Sync + Debug,
    F: Fn(T) + Send + Sync,
    I: ConcurrentIter<Item = T>,
{
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                // pull values 1 by 1
                for value in iter.values() {
                    process(value);
                }

                while let Some(value) = iter.next() {
                    process(value);
                }

                // pull values and corresponding index 1 by 1
                for (idx, value) in iter.ids_and_values() {
                    dbg!(idx);
                    process(value);
                }

                while let Some(x) = iter.next_id_and_value() {
                    dbg!(x.idx);
                    process(x.value);
                }
            });
        }
    });
}

fn process_in_chunks<T, I, F>(
    process: &F,
    num_threads: usize,
    iter: &I,
    chunk_size: usize,
) where
    T: Send + Sync + Debug,
    F: Fn(T) + Send + Sync,
    I: ConcurrentIter<Item = T>,
{
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                // pull values in chunks of `chunk_size`
                let mut chunk_iter = iter.buffered_iter(16);
                while let Some(chunk) = chunk_iter.next() {
                    assert!(chunk.values.len() <= chunk_size);

                    for (i, value) in chunk.values.enumerate() {
                        let idx = chunk.begin_idx + i;

                        dbg!(idx);
                        process(value);
                    }
                }
            });
        }
    });
}

let process = |x| {
    dbg!(x);
};

process_one_by_one(&process, 8, &(0..1024).con_iter());
process_in_chunks(&process, 8, &(0..1024).con_iter(), 64);

• for and while let loops of process_one_by_one demonstrate the most basic usage where threads will pull the next element of the iterator whenever they complete processing the prior element.
• Note that each thread will pull different elements at different positions of the iterator depending on how fast they finish the execution of the task inside the loop. Therefore, an enumerate call inside the thread, or counting the pulled elements by that particular thread, does not provide the index of the element in the original data source. ConcurrentIter additionally provides the original index with ids_and_values or next_id_and_value methods.
• Whenever the work to be done inside the loop is too small (like just the dbg call in the above example), taking elements 1-by-1 might be suboptimal. In such cases, a better idea is to pull elements in chunks. In process_in_chunks, we create a buffered chunk iterator which pulls chunk_size (or less, if not enough left) consecutive elements at each next call. Note that chunk returned by chunk_iter.next() is an ExactSizeIterator with a known len.
• While iterating in chunks, we can still access the original idx of the elements. chunk.begin_idx represents the original index of the first element of the returned chunk.values iterator. Note that chunk.values is always non-empty; i.e., always has at least one element, otherwise, next returns None. Further, the chunk iterator contains consecutive elements. Hence, we can get the original indices of all elements by combining chunk.begin_idx with the local indices of the current chunk obtained by the chunk.values.enumerate; i.e., let idx = chunk.begin_idx + i.

Parallel Fold

Considering the elements of the iteration as inputs of a process, ConcurrentIter conveniently allows distribution of tasks to multiple threads. See below a parallel fold implementation using the concurrent iterator.

use orx_concurrent_iter::*;

fn compute(input: u64) -> u64 {
    input * 2
}

fn fold(aggregated: u64, value: u64) -> u64 {
    aggregated + value
}

fn par_fold(num_threads: usize, inputs: impl ConcurrentIter<Item = u64>) -> u64 {
    std::thread::scope(|s| {
        (0..num_threads)
            .map(|_| s.spawn(|| inputs.values().map(compute).fold(0u64, fold)))
            .collect::<Vec<_>>()
            .into_iter()
            .map(|x| x.join().expect("-_-"))
            .fold(0u64, fold)
    })
}

// validate
for num_threads in [1, 2, 4, 8] {
    let values = (0..1024).map(|x| 2 * x);
    let par_result = par_fold(num_threads, values.into_con_iter());
    assert_eq!(par_result, 2 * 1023 * 1024);
}

Notes on the implementation:

• Concurrent iterator allows for a simple 7-line parallel fold implementation.
• Parallel fold operation is defined as two fold operations.
  • The first .map(_).fold(_) defines the parallel fold operation executed by num_threads threads. Each thread returns its own aggregated result.
  • The second map(_).fold(_) defines the final sequential fold operation executed to fold over the num_threads results obtained by each thread.

Parallel Map

Parallel map can also be implemented by merging returned transformed collections, such as vectors. Especially for larger data types, a more efficient approach could be to pair ConcurrentIter with a concurrent collection such as orx_concurrent_bag::ConcurrentBag which allows to efficiently collect results concurrently without copies.

use orx_concurrent_iter::*;
use orx_concurrent_bag::*;

fn map(input: u64) -> String {
    input.to_string()
}

fn parallel_map(num_threads: usize, iter: impl ConcurrentIter<Item = u64>) -> SplitVec<String> {
    let outputs = ConcurrentBag::new();
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                for output in iter.values().map(map) {
                    outputs.push(output);
                }
            });
        }
    });
    outputs.into_inner()
}

// test
for num_threads in [1, 2, 4, 8] {
    let inputs = (0..1024).map(|x| 2 * x);
    let outputs = parallel_map(num_threads, inputs.into_con_iter());
    assert_eq!(1024, outputs.len());
}

Note that due to parallelization, outputs is not guaranteed to be in the same order as inputs. In order to preserve the order of the input in the output, iteration with indices can be used to sort the result accordingly. Alternative to post-sorting, ConcurrentBag can be replaced with orx_concurrent_bag::ConcurrentOrderedBag to already collect in order.

Parallel Find, A Little Communication Among Threads

As illustrated above, efficient parallel implementations of different methods are conveniently possible with ConcurrentIter. There is only one bit of information implicitly shared among threads by the concurrent iterator: the elements left. In scenarios where we do not need to iterate over all elements, we can use this information to share a message among threads. We might call such cases as early-return scenarios. A common example is the find method, where we are looking for a matching element and we want to terminate the search as soon as we find one.

You may see a parallel implementation of the find method below.

use orx_concurrent_iter::*;

fn par_find<I, P>(iter: I, predicate: P, n_threads: usize) -> Option<(usize, I::Item)>
where
    I: ConcurrentIter,
    P: Fn(&I::Item) -> bool + Send + Sync,
{
    std::thread::scope(|s| {
        (0..n_threads)
            .map(|_| {
                s.spawn(|| {
                    for (i, x) in iter.ids_and_values() {
                        if predicate(&x) {
                            iter.skip_to_end();
                            return Some((i, x));
                        }
                    }
                    None
                })
            })
            .collect::<Vec<_>>()
            .into_iter()
            .flat_map(|x| x.join().expect("(-)"))
            .min_by_key(|x| x.0)
    })
}

let mut names: Vec<_> = (0..8785).map(|x| x.to_string()).collect();
names[42] = "foo".to_string();

let result = par_find(names.con_iter(), |x| x.starts_with('x'), 4);
assert_eq!(result, None);

let result = par_find(names.con_iter(), |x| x.starts_with('f'), 4);
assert_eq!(result, Some((42, &"foo".to_string())));

names[43] = "foo_second_match".to_string();
let result = par_find(names.con_iter(), |x| x.starts_with('f'), 4);
assert_eq!(result, Some((42, &"foo".to_string())));

Notice that the parallel find implementation is in two folds:

• (parallel search) Inside each thread, we loop through the elements of the concurrent iterator and return the first value satisfying the predicate together with its index.
• (sequential wrap up) Since this is a parallel execution, we might end up receiving multiple matches from multiple threads. In the second part, we investigate the thread results and return the one with the minimum position index (min_by_key(|x| x.0)) since that is the element which appears first in the original iterator.

So far, this is straightforward and similar to the parallel fold implementation. The difference; however, is the additional iter.skip_to_end() call. This call will immediately consume all remaining elements of the iterator. Therefore, whenever, another thread tries to advance the iterator in the for (i, x) in iter.ids_and_values(), it will not receive any further elements. Hence, they will as well return as soon as they complete processing their last pulled element. This establishes a very trivial communication among threads, which is critical in achieving efficiency in early return scenarios, as the find method. To demonstrate, assume the case we didn't use iter.skip_to_end() in the above implementation.

• In the second example, the iterator has 8785 elements where there exists only one element satisfying the predicate, "foo" at position 42.
• One of the 4 threads used, say A, will find this element and return immediately.
• The other 3 threads will never see this element, since it is pulled by A. They will iterate over all remaining elements and will eventually return None.
• The final result will be correct. However, this implementation will evaluate all elements of the iterator regardless of where the first matching element is. Although parallelized, this would be a very inefficient implementation.

Traits and Implementors

The trait defining types that can be safely be iterated concurrently by multiple threads is ConcurrentIter.

Further, there are two traits which define types that can provide a ConcurrentIter.

• A ConcurrentIterable type implements the con_iter(&self) method which returns a concurrent iterator without consuming the type itself.
• On the other hand, types implementing IntoConcurrentIter trait has the into_con_iter(self) method which consumes and converts the type into a concurrent iterator. Additionally there exists IterIntoConcurrentIter trait which is functionally identical to IntoConcurrentIter and only implemented by regular iterators, separated only to allow for special implementations for vectors and arrays.

The following table summarizes the implementations of the standard types in this crate.

Finally, concurrent iterators having an element type which is a reference to a Clone or Copy type, have the cloned() or copied() methods, allowing to iterate over cloned values.

Contributing

Contributions are welcome! If you notice an error, have a question or think something could be improved, please open an issue or create a PR.

License

This library is licensed under MIT license. See LICENSE for details.