orx-concurrent-ordered-bag

An efficient, convenient and lightweight grow-only concurrent data structure allowing high performance and ordered concurrent collection.

• convenient: ConcurrentBag can safely be shared among threads simply as a shared reference. It is a PinnedConcurrentCol with a special concurrent state implementation. Underlying PinnedVec and concurrent bag can be converted back and forth to each other. Further, as you may see in the parallel map example, it enables efficient parallel methods with possibly the most convenient and simple implementation.
• efficient: ConcurrentBag is a lock free structure making use of a few atomic primitives, this leads to high performance concurrent growth while enabling to collect the results in the desired order.

Comparison to ConcurrentBag

Note that ConcurrentBag is a similar structure with the following differences.

Safety Requirements

As the comparison reveals, ConcurrentBag is much safer to use than ConcurrentOrderedBag, which could be the first choice if the order of collected elements does not matter. On the other hand, required safety guarantees fortunately are not too difficult to satisfy. ConcurrentOrderedBag can safely be used provided that the following two conditions are satisfied:

• Each position is written exactly once, so that there exists no race condition.
• At the point where into_inner is called (not necessarily always), the bag must not contain any gaps.
  • Let m be the maximum index of the position that we write an element to.
  • The bag assumes that the length of the vector is equal to m + 1.
  • Then, it expects that exactly m + 1 elements are written to the bag.
  • If the first condition was satisfied; then, this condition is sufficient to conclude that the bag can be converted to the underlying vector of m + 1 elements.

Examples

Safety guarantees to push to the bag with a shared reference makes it easy to share the bag among threads. However, the caller is required to make sure that the collection does not contain gaps and each position is written exactly once.

Manual Example

In the following example, we split computation among two threads: the first thread processes inputs with even indices, and the second with odd indices. This provides the required guarantee mentioned above.

use orx_concurrent_ordered_bag::*;

let n = 1024;

let evens_odds = ConcurrentOrderedBag::new();

// just take a reference and share among threads
let bag = &evens_odds;

std::thread::scope(|s| {
    s.spawn(move || {
        for i in (0..n).filter(|x| x % 2 == 0) {
            unsafe { bag.set_value(i, i as i32) };
        }
    });

    s.spawn(move || {
        for i in (0..n).filter(|x| x % 2 == 1) {
            unsafe { bag.set_value(i, -(i as i32)) };
        }
    });
});

let vec = unsafe { evens_odds.into_inner().unwrap_only_if_counts_match() };
assert_eq!(vec.len(), n);
for i in 0..n {
    if i % 2 == 0 {
        assert_eq!(vec[i], i as i32);
    } else {
        assert_eq!(vec[i], -(i as i32));
    }
}

Note that as long as no-gap and write-only-once guarantees are satisfied, ConcurrentOrderedBag is very flexible in the order of writes. They can simply happen in arbitrary order. Consider the following instance for instance. We spawn a thread just two write to the end of the collection, and then spawn a bunch of other threads to fill the beginning of the collection. This just works without any locks or waits.

use orx_concurrent_ordered_bag::*;

let n = 1024;
let num_additional_threads = 4;

let bag = ConcurrentOrderedBag::new();
let con_bag = &bag;

std::thread::scope(|s| {
    s.spawn(move || {
        // start writing to the end
        unsafe { con_bag.set_value(n - 1, 42) };
    });

    for thread in 0..num_additional_threads {
        s.spawn(move || {
            // then fill the rest concurrently from the beginning
            for i in (0..(n - 1)).filter(|i| i % num_additional_threads == thread) {
                unsafe { con_bag.set_value(i, i as i32) };
            }
        });
    }
});

let vec = unsafe { bag.into_inner().unwrap_only_if_counts_match() };
assert_eq!(vec.len(), n);
for i in 0..(n - 1) {
    assert_eq!(vec[i], i as i32);
}
assert_eq!(vec[n - 1], 42);

These examples represent cases where the work can be trivially split among threads while providing the safety requirements. However, it requires special care and correctness. This complexity can significantly be avoided by pairing the ConcurrentOrderedBag with a ConcurrentIter on the input side.

Parallel Map with ConcurrentIter

Parallel map operation is one of the cases where we would care about the order of the collected elements, and hence, ConcurrentBag would not suffice. On the other hand, an efficient implementation can be achieved with ConcurrentOrderedBag and ConcurrentIter. Further, it might possibly be the most convenient parallel map implementation that is almost identical to a single-threaded map implementation.

use orx_concurrent_ordered_bag::*;
use orx_concurrent_iter::*;

fn parallel_map<In, Out, Map, Inputs>(
    num_threads: usize,
    inputs: Inputs,
    map: &Map,
) -> ConcurrentOrderedBag<Out>
where
    Inputs: ConcurrentIter<Item = In>,
    Map: Fn(In) -> Out + Send + Sync,
    Out: Send + Sync,
{
    let outputs = ConcurrentOrderedBag::new();
    let inputs = &inputs;
    let out = &outputs;
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                while let Some(next) = inputs.next_id_and_value() {
                    unsafe { out.set_value(next.idx, map(next.value)) };
                }
            });
        }
    });
    outputs
}

let len = 2465;
let vec: Vec<_> = (0..len).map(|x| x.to_string()).collect();

let bag = parallel_map(4, vec.into_con_iter(), &|x| x.to_string().len());
let output = unsafe { bag.into_inner().unwrap_only_if_counts_match() };

assert_eq!(output.len(), len);
for (i, value) in output.iter().enumerate() {
    assert_eq!(value, &i.to_string().len());
}

As you may see, we are not required to share the work manually, we simply use a while let Some loop. The work is pulled by threads from the iterator. This both leads to an efficient implementation especially in cases of heterogeneous work loads of each task and automatically provides the safety requirements.

Parallel Map with ExactSizeConcurrentIter

A further performance improvement to the parallel map implementation above is to distribute the tasks among the threads in chunks. The aim of this approach is to avoid false sharing, you may see further details here. This can be achieved by pairing an ExactSizeConcurrentIter rather than a ConcurrentIter with the set_values method of the ConcurrentOrderedBag.

use orx_concurrent_ordered_bag::*;
use orx_concurrent_iter::*;

fn parallel_map<In, Out, Map, Inputs>(
    num_threads: usize,
    inputs: Inputs,
    map: &Map,
    chunk_size: usize,
) -> ConcurrentOrderedBag<Out>
where
    Inputs: ExactSizeConcurrentIter<Item = In>,
    Map: Fn(In) -> Out + Send + Sync,
    Out: Send + Sync,
{
    let outputs = ConcurrentOrderedBag::new();
    let inputs = &inputs;
    let out = &outputs;
    std::thread::scope(|s| {
        for _ in 0..num_threads {
            s.spawn(|| {
                while let Some(next) = inputs.next_exact_chunk(chunk_size) {
                    unsafe { out.set_values(next.begin_idx(), next.values().map(map)) };
                }
            });
        }
    });
    outputs
}

let len = 2465;
let vec: Vec<_> = (0..len).map(|x| x.to_string()).collect();
let bag = parallel_map(4, vec.into_con_iter(), &|x| x.to_string().len(), 64);
let output = unsafe { bag.into_inner().unwrap_only_if_counts_match() };
for (i, value) in output.iter().enumerate() {
    assert_eq!(value, &i.to_string().len());
}

Construction

ConcurrentOrderedBag can be constructed by wrapping any pinned vector; i.e., ConcurrentOrderedBag<T> implements From<P: PinnedVec<T>>. Likewise, a concurrent vector can be unwrapped without any allocation to the underlying pinned vector with into_inner method, provided that the safety requirements are satisfied.

Further, there exist with_ methods to directly construct the concurrent bag with common pinned vector implementations.

use orx_concurrent_ordered_bag::*;

// default pinned vector -> SplitVec<T, Doubling>
let bag: ConcurrentOrderedBag<char> = ConcurrentOrderedBag::new();
let bag: ConcurrentOrderedBag<char> = Default::default();
let bag: ConcurrentOrderedBag<char> = ConcurrentOrderedBag::with_doubling_growth();
let bag: ConcurrentOrderedBag<char, SplitVec<char, Doubling>> = ConcurrentOrderedBag::with_doubling_growth();

let bag: ConcurrentOrderedBag<char> = SplitVec::new().into();
let bag: ConcurrentOrderedBag<char, SplitVec<char, Doubling>> = SplitVec::new().into();

// SplitVec with [Linear](https://docs.rs/orx-split-vec/latest/orx_split_vec/struct.Linear.html) growth
// each fragment will have capacity 2^10 = 1024
// and the split vector can grow up to 32 fragments
let bag: ConcurrentOrderedBag<char, SplitVec<char, Linear>> = ConcurrentOrderedBag::with_linear_growth(10, 32);
let bag: ConcurrentOrderedBag<char, SplitVec<char, Linear>> = SplitVec::with_linear_growth_and_fragments_capacity(10, 32).into();

// [FixedVec](https://docs.rs/orx-fixed-vec/latest/orx_fixed_vec/) with fixed capacity.
// Fixed vector cannot grow; hence, pushing the 1025-th element to this bag will cause a panic!
let bag: ConcurrentOrderedBag<char, FixedVec<char>> = ConcurrentOrderedBag::with_fixed_capacity(1024);
let bag: ConcurrentOrderedBag<char, FixedVec<char>> = FixedVec::new(1024).into();

Of course, the pinned vector to be wrapped does not need to be empty.

use orx_concurrent_ordered_bag::*;

let split_vec: SplitVec<i32> = (0..1024).collect();
let bag: ConcurrentOrderedBag<_> = split_vec.into();

Concurrent State and Properties

The concurrent state is modeled simply by an atomic capacity. Combination of this state and PinnedConcurrentCol leads to the following properties:

• Writing to a position of the collection does not block other writes, multiple writes can happen concurrently.
• Caller is required to guarantee that each position is written exactly once.
• ⟹ caller is responsible to avoid write & write race conditions.
• Only one growth can happen at a given time.
• Reading is only possible after converting the bag into the underlying PinnedVec.
• ⟹ no read & write race condition exists.

License

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