Priority queue traits and efficient d-ary heap implementations.
This crate aims to provide algorithms with the abstraction over priority queues. In order to achieve this, two traits are defined: PriorityQueue<N, K>
and PriorityQueueDecKey<N, K>
. The prior is a simple queue while the latter extends it by providing additional methods to change priorities of the items that already exist in the queue.
The separation is important since additional operations often requires the implementors to allocate internal memory for bookkeeping. Therefore, we would prefer PriorityQueueDecKey<N, K>
only when we need to change the priorities.
Three categories of d-ary heap implementations are provided.
DaryHeap<N, K, const D: usize>
is the basic queue implementing PriorityQueue<N, K>
. As mentioned before, it is to be preferred unless priority updates, decrease-key operations, are required.
The following two types additionally implement PriorityQueueDecKey<N, K>
which serve different purposes:
DaryHeapOfIndices<N, K, const D>
is a d-ary heap paired up with a positions array. It requires the nodes to implementHasIndex
trait which is nothing butfn index(&self) -> usize
. Further, it requires the maximum index that is expected to enter the queue (candidates coming from a closed set). Once these conditions are satisfied, it performs significantly faster than the alternative decrease key queues.- Although the closed set requirement might sound strong, it is often satisfied in mathematical algorithms. For instance, for most network traversal algorithms, the candidates set is the nodes of the graph.
DaryHeapWithMap<N, K, const D: usize>
is a d-ary heap paired up with a positionsHashMap
(BTreeMap
with no-std). This provides the open-set flexibility and fits better to more general cases, rather than mathematical algorithms.
All three variants of the d-ary heap implementations take complete benefit of const generics to speed up traversal on the heap when d is a power of two.
In addition the the following implementations are provided in this crate:
std::collections::BinaryHeap<(N, K)>
implementsPriorityQueue<N, K>
,priority_queue:PriorityQueue<N, K>
impelements bothPriorityQueue<N, K>
andPriorityQueueDecKey<N, K>
.
This allows to use all the queue implementations interchangeably and measure performance. See B. Benchmarks section to see the experiments.
Decrease key operations are critical in certain cases.
Assume a scenario where the same candidate node might be evaluated multiple times with different priorities throughout the life of the queue. Just as in Dijkstra's shortest path algorithm.
We can handle this with a basic PriorityQueue
as follows:
- every time a node is observed as a candidate with a given priority, we push it again to the priority queue;
- this means that we will pop the same node multiple times with priorities in worsening order;
- we additionally keep track of each node that is popped from the queue;
- when we pop the same node the second or third time, we simply ignore it.
This approach might work well in many cases. However, it is not the most memory efficient solution. If we assume that the algorithm using the queue is the Dijkstra's shortest path, and if our graph has N nodes, each node can enter the queue N times in the worst case. This would lead to a space complexity of O(N^2).
On the other hand, with a PriorityQueueDecKey
, each time we observe a candidate with a priority:
- we can push the (node, priority) if it doesn't exist in the queue,
- we can decrease its priority on the queue if it exists in the queue with a worse priority,
- or lastly, we can ignore the candidate if the node already exists in the queue with a better priority.
This approach would guarantee that each node enters the queue at most once. It would reduce the space complexity of the Dijkstra's shortest path to O(N). Another benefit of this approach is to push some part of the book keeping to the queue allowing the algorithm to be more concise. For instance, try_decrease_key_or_push
method removes almost all book keeping from Dijkstra's shortest path algorithm.
However, there is not a clear winner. Most performant queue often depends on the input data and can be decided empirically. As mentioned, allowing these experiments and benchmarks is one of the key motivations behind the priority queue traits. You may see such an exercise in the repository https://github.com/orxfun/orx-bench-shortest-path.
In summary, DaryHeap
performs comparable to (slightly faster than) std::collections::BinaryHeap
and significantly faster than decrease key variants.
For the cases where we need decrease key operations, DaryHeapOfIndices
performs significantly faster than other tested priority queues.
This benchmark compares basic push & pop operations on different PriorityQueue
implementations. See "benches/basic_queue.rs" for details. The computation times in µs for different test data sizes are presented in the following table.
DaryHeap
implementations with different d-values andstd::collections::BinaryHeap
perform at a similar scale.- Among these,
DaryHeap<_, _, 4>
performs consistently faster.
- Among these,
priority_queue::PriorityQueue
is around five times slower.
In the next benchmark, different PriorityQueueDecKey
implementations are tested on a benchmark where decrease key calls are made in addition to push & pop operations. See "benches/deckey_queue.rs" for details. The computation times in µs for different test data sizes are presented in the following table. Since std::collections::BinaryHeap
does not provide the required methods it is excluded from the analysis.
- There are three clusters with respect to performance:
DaryHeapOfIndices
is significantly faster than others (~5 times faster thanDaryHeapWithMap
and ~3.5 times faster thanpriority_queue::PriorityQueue
),priority_queue::PriorityQueue
is around 1.5 times faster thanDaryHeapWithMap
, andDaryHeapWithMap
is the slowest.
- Among all variants,
DaryHeapOfIndices<_, _, 4>
consistently outperforms the others.
use orx_priority_queue::*;
// generic over simple priority queues
fn test_priority_queue<P>(mut pq: P)
where
P: PriorityQueue<usize, f64>,
{
pq.clear();
pq.push(0, 42.0);
assert_eq!(Some(&0), pq.peek().map(|x| x.node()));
assert_eq!(Some(&42.0), pq.peek().map(|x| x.key()));
let popped = pq.pop();
assert_eq!(Some((0, 42.0)), popped);
assert!(pq.is_empty());
pq.push(0, 42.0);
pq.push(1, 7.0);
pq.push(2, 24.0);
pq.push(10, 3.0);
while let Some(popped) = pq.pop() {
println!("pop {:?}", popped);
}
}
// generic over decrease-key priority queues
fn test_priority_queue_deckey<P>(mut pq: P)
where
P: PriorityQueueDecKey<usize, f64>,
{
pq.clear();
pq.push(0, 42.0);
assert_eq!(Some(&0), pq.peek().map(|x| x.node()));
assert_eq!(Some(&42.0), pq.peek().map(|x| x.key()));
let popped = pq.pop();
assert_eq!(Some((0, 42.0)), popped);
assert!(pq.is_empty());
pq.push(0, 42.0);
assert!(pq.contains(&0));
pq.decrease_key(&0, 7.0);
assert_eq!(Some(&0), pq.peek().map(|x| x.node()));
assert_eq!(Some(&7.0), pq.peek().map(|x| x.key()));
let deckey_result = pq.try_decrease_key(&0, 10.0);
assert!(matches!(ResTryDecreaseKey::Unchanged, deckey_result));
assert_eq!(Some(&0), pq.peek().map(|x| x.node()));
assert_eq!(Some(&7.0), pq.peek().map(|x| x.key()));
while let Some(popped) = pq.pop() {
println!("pop {:?}", popped);
}
}
// d-ary heap generic over const d
const D: usize = 4;
test_priority_queue(DaryHeap::<usize, f64, D>::default());
test_priority_queue(DaryHeapWithMap::<usize, f64, D>::default());
test_priority_queue(DaryHeapOfIndices::<usize, f64, D>::with_index_bound(100));
test_priority_queue_deckey(DaryHeapWithMap::<usize, f64, D>::default());
test_priority_queue_deckey(DaryHeapOfIndices::<usize, f64, D>::with_index_bound(100));
// or type aliases for common heaps to simplify signature
// Binary or Quarternary to fix d of d-ary
test_priority_queue(BinaryHeap::default());
test_priority_queue(BinaryHeapWithMap::default());
test_priority_queue(BinaryHeapOfIndices::with_index_bound(100));
test_priority_queue_deckey(QuarternaryHeapOfIndices::with_index_bound(100));
Below are some additional traits we need to set up the example.
type Weight = u32;
pub trait OutEdgeData {
fn head(&self) -> usize;
fn weight(&self) -> Weight;
}
pub trait OutEdges {
type Edge: OutEdgeData;
fn next_edge(&mut self) -> Option<Self::Edge>;
fn count_edges(&mut self) -> usize;
}
pub trait Graph {
type OutEdges<'a>: OutEdges where where Self:'a;
fn num_nodes(&self) -> usize;
fn out_edges(&self, node: usize) -> Self::OutEdges<'_>;
}
You may see below two implementations one using a PriorityQueue
and the other with a PriorityQueueDecKey
.
use orx_priority_queue::*;
fn run_with_basic_pq<G: Graph, Q: PriorityQueue<usize, Weight>>(
graph: &G,
queue: &mut Q,
source: usize,
sink: usize,
) -> Option<Weight> {
// reset
queue.clear();
let mut dist = vec![Weight::MAX; graph.num_nodes()];
// init
dist[source] = 0;
queue.push(source, 0);
// iterate
while let Some((node, cost)) = queue.pop() {
if node == sink {
return Some(cost);
} else if cost > dist[node] {
continue;
}
let mut out_edges = graph.out_edges(node);
while let Some(edge) = out_edges.next_edge() {
let (head, weight) = (edge.head(), edge.weight());
let next_cost = cost + weight;
if next_cost < dist[head] {
queue.push(head, next_cost);
dist[head] = next_cost;
}
}
}
None
}
fn run_with_pq_deckey<G: Graph, Q: PriorityQueueDecKey<usize, Weight>>(
graph: &G,
queue: &mut Q,
source: usize,
sink: usize,
) -> Option<Weight> {
// reset
queue.clear();
let mut visited = vec![false; graph.num_nodes()];
// init
visited[source] = true;
queue.push(source, 0);
// iterate
while let Some((node, cost)) = queue.pop() {
if node == sink {
return Some(cost);
}
let mut out_edges = graph.out_edges(node);
while let Some(edge) = out_edges.next_edge() {
let (head, weight) = (edge.head(), edge.weight());
if !visited[head] {
// this call will:
// * push the `head` if not present in the queue, or
// * decrease its priority to `cost + weight` if it was greater, or
// * leave the queue untouched otherwise.
queue.try_decrease_key_or_push(&head, cost + weight);
}
}
visited[node] = true;
}
None
}
This library is licensed under MIT license. See LICENSE for details.