Keap is a heap data structure written in Kotlin similar to binary heap. It maintains separately the queue of elements and the tournament tree atop the queue.
Keap is stable, that is, it keeps initial order of equal elements.
It's faster than binary heap in terms of number of comparisons. For any kind of input (random or ordered) of size n
,
the heapify procedure requires exactly n - 1
comparisons. Binary heap reaches this number of comparisons
only for ordered input. For random input, keap compared to binary heap does approximately 90% less comparisons
to heapify, 20% more comparisons to offer[todo], and more than 3 times less
comparisons to poll. Though, array-backed keap consumes 2-3 times more memory than array-backed binary heap.
Performance summary of keap and binary heap (both array-backed) is as follows:
Keap average | Keap worst case | Binary heap average | Binary heap worst case | |
---|---|---|---|---|
Heapify | exactly n - 1 | exactly n - 1 | Θ(n) | Θ(n) |
Peek | Θ(1) | Θ(1) | Θ(1) | Θ(1) |
Poll | Θ(log n) | Θ(log n) | Θ(log n) | Θ(log n) |
Offer | Θ(1) | Θ(log n) | Θ(1) | Θ(log n) |
Memory used | Θ(n) | Θ(n) | exactly n | exactly n |
Here are two applications: keap-based PriorityQueue
as a replacement of java.util.PriorityQueue
and Keapsort
sorting algorithm. Both might be useful in two cases:
- stability is a must;
- comparing elements is rather heavyweight operation (e.g., it requires a database access).
PriorityQueue is a keap-based stable replacement of java.util.PriorityQueue
.
Keapsort is a Heapsort with keap used instead of binary heap. Keapsort is a comparison-based sorting algorithm with a worst-case O(n log n) runtime. As keap is a stable priority queue, Keapsort is stable. Unlike Heapsort, Keapsort doesn't have an in-place version.
Like Heapsort, Keapsort produces first output after Θ(n) comparisons.
For Keapsort, this Θ(n) estimate is just equal to n - 1
. To sort random input completely, Keapsort does almost
two times less comparisons than Heapsort, nearly 1% less comparisons than
Mergesort and 1-2% more comparisons than
Timsort.
Like Heapsort, Keapsort is not a modern CPU-friendly algorithm since it has poor data cache performance.
Like Heapsort, Keapsort doesn't seem to parallelize well.
JMH Gradle Plugin is used to build and run benchmarks.
Benchmark results are obtained on a PC running under Windows 7 with Intel(R) Core(TM) i7-3770 3.4 GHz CPU
and 64-bit JRE build 1.8.0_112-b15 with the following java parameters: -Xms1g -Xmx1g
. To get results in your
environment, run:
./gradlew clean jar jmh
For both java.util.PriorityQueue
and keap-based PriorityQueue, there are two types of benchmarks:
- Benchmarks examining operations with random elements: heapify (building the queue from a collection), offer, peek, and poll. The queue elements are random strings of length ~30 characters with constant 10-characters prefix. Queue size is 10000.
- Benchmarks examining offering ordered elements: offerIncreasing (successively increasing elements) and offerDecreasing (successively decreasing elements). The queue elements are strings of length ~30 characters with constant 10-characters prefix. Queue size is 10000.
Current results grouped for easy review are as follows. Building the queue from collections of random elements:
Benchmark Mode Cnt Score Error Units
JavaQueueRandomBenchmark.heapify thrpt 20 1.812 ± 0.041 ops/ms
KeapQueueRandomBenchmark.heapify thrpt 20 3.315 ± 0.019 ops/ms
Basic queue operations with random elements:
Benchmark Mode Cnt Score Error Units
JavaQueueRandomBenchmark.offer thrpt 20 3.652 ± 0.075 ops/us
JavaQueueRandomBenchmark.peek thrpt 20 81.031 ± 0.488 ops/us
JavaQueueRandomBenchmark.poll thrpt 20 4.267 ± 0.074 ops/us
KeapQueueRandomBenchmark.offer thrpt 20 3.349 ± 0.077 ops/us
KeapQueueRandomBenchmark.peek thrpt 20 82.927 ± 0.484 ops/us
KeapQueueRandomBenchmark.poll thrpt 20 11.522 ± 0.264 ops/us
Offering ordered elements:
Benchmark Mode Cnt Score Error Units
JavaQueueOrderedBenchmark.offerDecreasing thrpt 20 5.506 ± 0.071 ops/us
JavaQueueOrderedBenchmark.offerIncreasing thrpt 20 31.228 ± 0.363 ops/us
KeapQueueOrderedBenchmark.offerDecreasing thrpt 20 8.318 ± 0.048 ops/us
KeapQueueOrderedBenchmark.offerIncreasing thrpt 20 7.539 ± 0.099 ops/us
The scores above are numbers of operations per microsecond, for heapify - per millisecond. So the greater the score, the better performance.
<!-- in Maven project -->
<dependency>
<groupId>com.github.penemue</groupId>
<artifactId>keap</artifactId>
<version>0.1.0</version>
</dependency>
// in Gradle project
dependencies {
compile 'com.github.penemue:keap:0.1.0'
}
Gradle is used to build, test, and run benchmarks. JDK 1.8 and Kotlin 1.1.1 are required. To build the project, run:
./gradlew
-
On increasing capacity, get rid of heapify, thus reducing number of comparisons on offer.
-
Compare Keapsort to Quicksort.
-
Looks like a tree-backed version of keap could be exposed as an immutable/persistent/lock-free heap data structure. In addition, it could support heap merge operation in Θ(1) time.