Skip to content

A priority queue for Rust with efficient change function.

Notifications You must be signed in to change notification settings

garro95/priority-queue

Repository files navigation

PriorityQueue

crate Build Test

This crate implements a Priority Queue with a function to change the priority of an object. Priority and items are stored in an IndexMap and the queue is implemented as a Heap of indexes.

Please read the API documentation here

Usage

To use this crate, simply add the following string to your Cargo.toml:

priority-queue = "2.0.0"

or use the command cargo add priority-queue

Version numbers follow the semver convention.

Then use the data structure inside your Rust source code as in the following Example.

Remember that, if you need serde support, you should compile using --features serde.

Examples

use priority_queue::PriorityQueue;

fn main() {
    let mut pq = PriorityQueue::new();

    assert!(pq.is_empty());
    pq.push("Apples", 5);
    pq.push("Bananas", 8);
    pq.push("Strawberries", 23);

    assert_eq!(pq.peek(), Some((&"Strawberries", &23)));

    for (item, _) in pq.into_sorted_iter() {
        println!("{}", item);
    }
}

By default, the highest priority element will be extracted first. The order can be easily reversed using the standard wrapper Reverse<T>.

use priority_queue::PriorityQueue;
use std::cmp::Reverse;

fn main() {
    let mut pq = PriorityQueue::new();

    assert!(pq.is_empty());
    pq.push("Apples", Reverse(5));
    pq.push("Bananas", Reverse(8));
    pq.push("Strawberries", Reverse(23));

    assert_eq!(pq.peek(), Some((&"Apples", &Reverse(5))));

    for (item, _) in pq.into_sorted_iter() {
        println!("{}", item);
    }
}

Speeding up

You can use custom BuildHasher for the underlying IndexMap and therefore achieve better performance. For example you can create the queue with the speedy FxHash hasher:

use hashbrown::hash_map::DefaultHashBuilder;

let mut pq = PriorityQueue::<_, _, DefaultHashBuilder>::with_default_hasher();

Attention: FxHash does not offer any protection for dos attacks. This means that some pathological inputs can make the operations on the hashmap O(n^2). Use the standard hasher if you cannot control the inputs.

Benchmarks

Some benchmarks have been run to compare the performances of this priority queue to the standard BinaryHeap, also using the FxHash hasher. On a Ryzen 9 3900X, the benchmarks produced the following results:

test benchmarks::priority_change_on_large_double_queue     ... bench:          25 ns/iter (+/- 1)
test benchmarks::priority_change_on_large_double_queue_fx  ... bench:          21 ns/iter (+/- 1)
test benchmarks::priority_change_on_large_queue            ... bench:          15 ns/iter (+/- 0)
test benchmarks::priority_change_on_large_queue_fx         ... bench:          11 ns/iter (+/- 0)
test benchmarks::priority_change_on_large_queue_std        ... bench:     190,345 ns/iter (+/- 4,976)
test benchmarks::priority_change_on_small_double_queue     ... bench:          26 ns/iter (+/- 0)
test benchmarks::priority_change_on_small_double_queue_fx  ... bench:          20 ns/iter (+/- 0)
test benchmarks::priority_change_on_small_queue            ... bench:          15 ns/iter (+/- 0)
test benchmarks::priority_change_on_small_queue_fx         ... bench:          10 ns/iter (+/- 0)
test benchmarks::priority_change_on_small_queue_std        ... bench:       1,694 ns/iter (+/- 21)
test benchmarks::push_and_pop                              ... bench:          31 ns/iter (+/- 0)
test benchmarks::push_and_pop_double                       ... bench:          31 ns/iter (+/- 0)
test benchmarks::push_and_pop_double_fx                    ... bench:          24 ns/iter (+/- 1)
test benchmarks::push_and_pop_fx                           ... bench:          26 ns/iter (+/- 0)
test benchmarks::push_and_pop_min_on_large_double_queue    ... bench:         101 ns/iter (+/- 2)
test benchmarks::push_and_pop_min_on_large_double_queue_fx ... bench:          98 ns/iter (+/- 0)
test benchmarks::push_and_pop_on_large_double_queue        ... bench:         107 ns/iter (+/- 2)
test benchmarks::push_and_pop_on_large_double_queue_fx     ... bench:         106 ns/iter (+/- 2)
test benchmarks::push_and_pop_on_large_queue               ... bench:          84 ns/iter (+/- 1)
test benchmarks::push_and_pop_on_large_queue_fx            ... bench:          78 ns/iter (+/- 2)
test benchmarks::push_and_pop_on_large_queue_std           ... bench:          71 ns/iter (+/- 1)
test benchmarks::push_and_pop_std                          ... bench:           4 ns/iter (+/- 0)

The priority change on the standard queue was obtained with the following:

pq = pq.drain().map(|Entry(i, p)| {
    if i == 50_000 {
        Entry(i, p/2)
    } else {
        Entry(i, p)
    }
}).collect()

The interpretation of the benchmarks is that the data structures provided by this crate is generally slightly slower than the standard Binary Heap.

On small queues (<10000 elements), the change_priority function, obtained on the standard Binary Heap with the code above, is way slower than the one provided by PriorityQueue and DoublePriorityQueue. With the queue becoming bigger, the operation takes almost the same amount of time on PriorityQueue and DoublePriorityQueue, while it takes more and more time for the standard queue.

It also emerges that the ability to arbitrarily pop the minimum or maximum element comes with a cost, that is visible in all the operations on DoublePriorityQueue, that are slower then the corresponding operations executed on the PriorityQueue.

Contributing

Feel free to contribute to this project with pull requests and/or issues.

All contribution shall be under a license compatible with the GNU LGPL version 3 or any later version and with the MPL version 2.0.

Changes

  • 2.1.1 Bug fix: #56
  • 2.1.0 Implement drain and reserve variations
  • 2.0.3 Some licensing-related housekeeping
  • 2.0.2 Fix docs.rs build
  • 2.0.1 Documentation improvements
  • 2.0.0 This release contains breaking changes
    • Some methods now require the trait bound H: BuildHasher. This change will likely have a small impact or none.
    • The standard library support is no longer auto-detected. The feature "std" is included in the default feature set, or else can be enabled like any other Cargo feature. Users that need to support no_std targets will have to disable default features.
  • 1.4.0 Improve shrink_to_fit to also shrink the internal IndexMap (#50)
  • 1.3.2 Bug fix in the log2_fast internal function
  • 1.3.1 Bug fix: #42
  • 1.3.0 Return bool from change_priority_by (Merged #41)
  • 1.2.3 Further performance optimizations (mainly on DoublePriorityQueue)
  • 1.2.2 Performance optimizations
  • 1.2.1 Bug fix: #34
  • 1.2.0 Implement DoublePriorityQueue data structure
  • 1.1.1 Convert documentation to Markdown
  • 1.1.0 Smooth Q: Sized requirement on some methods (fix #32)
  • 1.0.5 Bug fix: #28
  • 1.0.4 Bug fix: #28
  • 1.0.3 Bug fix: #26
  • 1.0.2 Added documentation link to Cargo.toml so the link is shown in the results page of crates.io
  • 1.0.1 Documentation
  • 1.0.0 This release contains breaking changes!
    • From and FromIterator now accept custom hashers -- Breaking: every usage of from and from_iter must specify some type to help the type inference. To use the default hasher (RandomState), often it will be enough to add something like

        let pq: PriorityQueue<_, _> = PriorityQueue::from...

      or you can add a type definition like

        type Pq<I, P> = PriorityQueue<I, P>

      and then use Pq::from() or Pq::from_iter()

    • Support no-std architectures

    • Add a method to remove elements at arbitrary positions

    • Remove take_mut dependency -- Breaking: change_priority_by signature has changed. Now it takes a priority_setter F: FnOnce(&mut P). If you want you can use the unsafe take_mut yourself or also use std::mem::replace

  • 0.7.0 Implement the push_increase and push_decrease convenience methods.
  • 0.6.0 Allow the usage of custom hasher
  • 0.5.4 Prevent panic on extending an empty queue
  • 0.5.3 New implementation of the Default trait avoids the requirement that P: Default
  • 0.5.2 Fix documentation formatting
  • 0.5.1 Add some documentation for iter_mut()
  • 0.5.0 Fix #7 implementing the iter_mut features
  • 0.4.5 Fix #6 for change_priority and change_priority_by
  • 0.4.4 Fix #6
  • 0.4.3 Fix #4 changing the way PriorityQueue serializes. Note that old serialized PriorityQueues may be incompatible with the new version. The API should not be changed instead.
  • 0.4.2 Improved performance using some unsafe code in the implementation.
  • 0.4.1 Support for serde when compiled with --features serde. serde marked as optional and serde-test as dev-dipendency. Now compiling the crate won't download and compile also serde-test, neither serde if not needed.
  • 0.4.0 Support for serde when compiled with cfg(serde)
  • 0.3.1 Fix #3
  • 0.3.0 Implement PartialEq and Eq traits