这是我参与「第三届青训营 -后端场」笔记创作活动的的第11篇笔记。PC端阅读效果更佳,点击文末:阅读原文即可。
「数据结构与算法」 第三届字节跳动青训营 - 后端专场
课程背景
- 这节课是介绍生产环境使用的算法和数据结构,然后重点从排序算法这个在课本上耳熟能详的算法分类开始,一步步打造出一个在工程实践领域性能一流的排序算法,介绍目前工业界最新的实践成果。
课程大纲
01 为什么要学习数据结构与算法
举个例子
规则:某个时间段内,直播间礼物数TOP10房间获得奖励,需要在每个房间展示排行榜
解决方案
- 礼物数量存储在Redis-zset中,使用skiplist使得元素整体有序
- 使用Redis集群,避免单机压力过大,使用主从算法、分片算法
- 保证集群原信息的稳定,使用一致性算法
- 后端使用缓存算法(LRU)降低Redis压力,展示房间排行榜
什么是最快的排序算法?
具体看特殊场景。
Python —-timsort
C + + —-introsort
Rust—- pdqsort
Go的排序算法有没有提升空间?
Go(< := 1.18) — introsort
课程讲师是给go官方提供了重新实现go排序算法
重新实现了Go的排序算法,在某些常见场景中比之前算法快~ 10倍,成为Go 1.19的默认排序算法
02 经典排序算法
插入排序
将元素不断插入已经排序好的array中
- 起始只有一个元素5,其本身是一个有序序列
- 后续元素插入有序序列中,即不断交换,直到找到第一个比其小的元素
时间复杂度:
最好 | 平均 | 最坏 |
---|---|---|
$O(n)$ | $O(n^2)$ | $O(n^2)$ |
Quick Sort 快速排序
分治思想,不断分割序列直到序列整体有序
- 选定一个pivot (轴点)
- 使用pivot分割序列,分成元素比pivot大和元素比pivot小两个序列
时间复杂度:
最好 | 平均 | 最坏 |
---|---|---|
$O(n*logn)$ | $O(n*logn)$ | $O(n^2)$ |
Heap Sort 堆排序
利用堆的性质形成的排序算法
- 构造一个大顶堆
- 将根节点(最大元素)交换到最后一个位置,调整整个堆,如此反复
时间复杂度:
最好 | 平均 | 最坏 |
---|---|---|
$O(n*logn)$ | $O(n*logn)$ | $O(n*logn)$ |
结论
- 插入排序平均和最坏情况时间复杂度都是$O(n^2)$,性能不好
- 快速排序整体性能处于中间层次
- 堆排序性能稳定,“众生平等”
实际场景benchmark
根据序列元素排列情况划分
- 完全随机的情况(random)
- 有序/逆序的情况(sorted/reverse)
- 元素重复度较高的情况(mod8)
在此基础上,还需要根据序列长度的划分(16/128/1024)
03 从零开始打造pdqsort
pdqsort -简介
pdqsort (pattern- defeating-quicksort)
是一种不稳定的混合排序算法,它的不同版本被应用在C++,BOOST、Rust 以及Go 1.19中。它对常见的序列类型做了特殊的优化,使得在不同条件下都拥有不错的性能
pdqsort - version1
结合三种排序方法的优点
对于短序列(小于一定长度),我们使用插入排序。其他情况,使用快速排序来保证整体性能。
当快速排序表现不佳时,使用堆排序来保证最坏情况下时间复杂度仍然为 $O(n*logn)$
Q&A
短序列的具体长度是多少呢?
- 12 ~ 32,在不同语言和场景中会有不同,在泛型版本根据测试选定24
如何得知快速排序表现不佳,以及何时切换到堆排序?
- 当最终pivot的位置离序列两端很接近时(距离小于length/8)判定其表现不佳,当这种情况的次数达到limit (即bits.Len(length)) 时,切换到堆排序。
- 对于短序列(<=24) 我们使用插入排序。
- 其他情况,使用快速排序(选择首个元素作为pivot)来保证整体性能。
- 当快速排序表现不佳时(limit==0) ,使用堆排序来保证最坏情况下时间复杂度仍然为O(n*logn)。
如何让pdqsort速度更快?
- 尽量使得QuickSort的pivot为序列的中位数 -> 改进choose pivot
- Partition速度更快->改进partition,但是此优化在Go表现不好,略
pdqsort - version2
思考关于pivot的选择
- 使用首个元素作为pivot(最简单的方案),实现简单,但是往往效果不好,例如在sorted情况下性能很差
- 遍历数组,寻找真正的中位数遍历比对代价很高,性能不好
根据序列长度的不同,来决定选择策略
- 优化 - Pivot的选择
- 短序列(<=8),选择固定元素
- 中序列(<=50),采样三个元素
- 长序列(>50),采样九个元素
Pivot的采样方式使得我们有探知序列当前状态的能力!
- 采样的元素都是逆序排列—-序列可能已经逆序—-翻转整个序列
- 采样的元素都是顺序排列—-序列可能已经有序—-使用插入排序
注:插入排序实际使用partiallnsertionSort,即有限制次数的插入排序。
Version1升级到version2优化总结
- 升级pivot选择策略(近似中位数)
- 发现序列可能逆序,则翻转序列->应对reverse场景
- 发现序列可能有序,使用有限插入排序->应对sorted场景
还有什么场景我们没有优化?
短序列情况
- 使用插入排序(v1)
极端情况
- 使用堆排序保证算法的可行性(v1)
完全随机的情况(random)
- 更好的pivot选择策略(v2)
有序/逆序的情况(sorted/reverse)
- 根据序列状态 翻转或者插入排序(v2)
元素重复度较高的情况(mod8) -> ?
pdqsort - final version
如何优化重复元素很多的情况?
- 采样pivot的时候检测重复度?不是很好,因为采样数量有限,不一定能采样到相同元素
解决方案:
- 如果两次partition生成的pivot相同,即partition进行了无效分割,此时认为pivot的值为重复元素
(相比上一种方法有更高的采样率)
优化-重复元素较多的情况(partitionEqual)
- 当检测到此时的pivot和上次相同时(发 生在leftSubArray)
使用partitionEqual将重复元素排列在一起,减少重复元素对于pivot选
择的干扰
优化-当pivot选择策略表现不佳时,随机交换元素
- 避免一些极端情况使得QuickSort总是表现不佳,以及一些黑客攻击情况
高性能的排序算法是如何设计的?
- 根据不同情况选择不同策略,取长补短
生产环境中使用的的排序算法和课本上的排序算法有什么区别?
- 理论算法注重理论性能,例如时间、空间复杂度等。生产环境中的算法需要面对不同的实践场景,更加注重实践性能
Go语言(<= 1.18)的排序算法是快速排序么?
- 实际一直是混合排序算法,主体是快速排序。Go <= 1.18时的算法也是基于快速排序,和pdqsort的区别在于fallback时机、pivot 选择策略、是否有针对不同pattern优化等
最后
欣赏一波大佬贡献的代码:
https://github.com/golang/go/blob/master/src/sort/zsortinterface.go
// Code generated by gen_sort_variants.go; DO NOT EDIT.
// Copyright 2022 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package sort
// insertionSort sorts data[a:b] using insertion sort.
func insertionSort(data Interface, a, b int) {
for i := a + 1; i < b; i++ {
for j := i; j > a && data.Less(j, j-1); j-- {
data.Swap(j, j-1)
}
}
}
// siftDown implements the heap property on data[lo:hi].
// first is an offset into the array where the root of the heap lies.
func siftDown(data Interface, lo, hi, first int) {
root := lo
for {
child := 2*root + 1
if child >= hi {
break
}
if child+1 < hi && data.Less(first+child, first+child+1) {
child++
}
if !data.Less(first+root, first+child) {
return
}
data.Swap(first+root, first+child)
root = child
}
}
func heapSort(data Interface, a, b int) {
first := a
lo := 0
hi := b - a
// Build heap with greatest element at top.
for i := (hi - 1) / 2; i >= 0; i-- {
siftDown(data, i, hi, first)
}
// Pop elements, largest first, into end of data.
for i := hi - 1; i >= 0; i-- {
data.Swap(first, first+i)
siftDown(data, lo, i, first)
}
}
// pdqsort sorts data[a:b].
// The algorithm based on pattern-defeating quicksort(pdqsort), but without the optimizations from BlockQuicksort.
// pdqsort paper: https://arxiv.org/pdf/2106.05123.pdf
// C++ implementation: https://github.com/orlp/pdqsort
// Rust implementation: https://docs.rs/pdqsort/latest/pdqsort/
// limit is the number of allowed bad (very unbalanced) pivots before falling back to heapsort.
func pdqsort(data Interface, a, b, limit int) {
const maxInsertion = 12
var (
wasBalanced = true // whether the last partitioning was reasonably balanced
wasPartitioned = true // whether the slice was already partitioned
)
for {
length := b - a
if length <= maxInsertion {
insertionSort(data, a, b)
return
}
// Fall back to heapsort if too many bad choices were made.
if limit == 0 {
heapSort(data, a, b)
return
}
// If the last partitioning was imbalanced, we need to breaking patterns.
if !wasBalanced {
breakPatterns(data, a, b)
limit--
}
pivot, hint := choosePivot(data, a, b)
if hint == decreasingHint {
reverseRange(data, a, b)
// The chosen pivot was pivot-a elements after the start of the array.
// After reversing it is pivot-a elements before the end of the array.
// The idea came from Rust's implementation.
pivot = (b - 1) - (pivot - a)
hint = increasingHint
}
// The slice is likely already sorted.
if wasBalanced && wasPartitioned && hint == increasingHint {
if partialInsertionSort(data, a, b) {
return
}
}
// Probably the slice contains many duplicate elements, partition the slice into
// elements equal to and elements greater than the pivot.
if a > 0 && !data.Less(a-1, pivot) {
mid := partitionEqual(data, a, b, pivot)
a = mid
continue
}
mid, alreadyPartitioned := partition(data, a, b, pivot)
wasPartitioned = alreadyPartitioned
leftLen, rightLen := mid-a, b-mid
balanceThreshold := length / 8
if leftLen < rightLen {
wasBalanced = leftLen >= balanceThreshold
pdqsort(data, a, mid, limit)
a = mid + 1
} else {
wasBalanced = rightLen >= balanceThreshold
pdqsort(data, mid+1, b, limit)
b = mid
}
}
}
// partition does one quicksort partition.
// Let p = data[pivot]
// Moves elements in data[a:b] around, so that data[i]<p and data[j]>=p for i<newpivot and j>newpivot.
// On return, data[newpivot] = p
func partition(data Interface, a, b, pivot int) (newpivot int, alreadyPartitioned bool) {
data.Swap(a, pivot)
i, j := a+1, b-1 // i and j are inclusive of the elements remaining to be partitioned
for i <= j && data.Less(i, a) {
i++
}
for i <= j && !data.Less(j, a) {
j--
}
if i > j {
data.Swap(j, a)
return j, true
}
data.Swap(i, j)
i++
j--
for {
for i <= j && data.Less(i, a) {
i++
}
for i <= j && !data.Less(j, a) {
j--
}
if i > j {
break
}
data.Swap(i, j)
i++
j--
}
data.Swap(j, a)
return j, false
}
// partitionEqual partitions data[a:b] into elements equal to data[pivot] followed by elements greater than data[pivot].
// It assumed that data[a:b] does not contain elements smaller than the data[pivot].
func partitionEqual(data Interface, a, b, pivot int) (newpivot int) {
data.Swap(a, pivot)
i, j := a+1, b-1 // i and j are inclusive of the elements remaining to be partitioned
for {
for i <= j && !data.Less(a, i) {
i++
}
for i <= j && data.Less(a, j) {
j--
}
if i > j {
break
}
data.Swap(i, j)
i++
j--
}
return i
}
// partialInsertionSort partially sorts a slice, returns true if the slice is sorted at the end.
func partialInsertionSort(data Interface, a, b int) bool {
const (
maxSteps = 5 // maximum number of adjacent out-of-order pairs that will get shifted
shortestShifting = 50 // don't shift any elements on short arrays
)
i := a + 1
for j := 0; j < maxSteps; j++ {
for i < b && !data.Less(i, i-1) {
i++
}
if i == b {
return true
}
if b-a < shortestShifting {
return false
}
data.Swap(i, i-1)
// Shift the smaller one to the left.
if i-a >= 2 {
for j := i - 1; j >= 1; j-- {
if !data.Less(j, j-1) {
break
}
data.Swap(j, j-1)
}
}
// Shift the greater one to the right.
if b-i >= 2 {
for j := i + 1; j < b; j++ {
if !data.Less(j, j-1) {
break
}
data.Swap(j, j-1)
}
}
}
return false
}
// breakPatterns scatters some elements around in an attempt to break some patterns
// that might cause imbalanced partitions in quicksort.
func breakPatterns(data Interface, a, b int) {
length := b - a
if length >= 8 {
random := xorshift(length)
modulus := nextPowerOfTwo(length)
for idx := a + (length/4)*2 - 1; idx <= a+(length/4)*2+1; idx++ {
other := int(uint(random.Next()) & (modulus - 1))
if other >= length {
other -= length
}
data.Swap(idx, a+other)
}
}
}
// choosePivot chooses a pivot in data[a:b].
//
// [0,8): chooses a static pivot.
// [8,shortestNinther): uses the simple median-of-three method.
// [shortestNinther,∞): uses the Tukey ninther method.
func choosePivot(data Interface, a, b int) (pivot int, hint sortedHint) {
const (
shortestNinther = 50
maxSwaps = 4 * 3
)
l := b - a
var (
swaps int
i = a + l/4*1
j = a + l/4*2
k = a + l/4*3
)
if l >= 8 {
if l >= shortestNinther {
// Tukey ninther method, the idea came from Rust's implementation.
i = medianAdjacent(data, i, &swaps)
j = medianAdjacent(data, j, &swaps)
k = medianAdjacent(data, k, &swaps)
}
// Find the median among i, j, k and stores it into j.
j = median(data, i, j, k, &swaps)
}
switch swaps {
case 0:
return j, increasingHint
case maxSwaps:
return j, decreasingHint
default:
return j, unknownHint
}
}
// order2 returns x,y where data[x] <= data[y], where x,y=a,b or x,y=b,a.
func order2(data Interface, a, b int, swaps *int) (int, int) {
if data.Less(b, a) {
*swaps++
return b, a
}
return a, b
}
// median returns x where data[x] is the median of data[a],data[b],data[c], where x is a, b, or c.
func median(data Interface, a, b, c int, swaps *int) int {
a, b = order2(data, a, b, swaps)
b, c = order2(data, b, c, swaps)
a, b = order2(data, a, b, swaps)
return b
}
// medianAdjacent finds the median of data[a - 1], data[a], data[a + 1] and stores the index into a.
func medianAdjacent(data Interface, a int, swaps *int) int {
return median(data, a-1, a, a+1, swaps)
}
func reverseRange(data Interface, a, b int) {
i := a
j := b - 1
for i < j {
data.Swap(i, j)
i++
j--
}
}
func swapRange(data Interface, a, b, n int) {
for i := 0; i < n; i++ {
data.Swap(a+i, b+i)
}
}
func stable(data Interface, n int) {
blockSize := 20 // must be > 0
a, b := 0, blockSize
for b <= n {
insertionSort(data, a, b)
a = b
b += blockSize
}
insertionSort(data, a, n)
for blockSize < n {
a, b = 0, 2*blockSize
for b <= n {
symMerge(data, a, a+blockSize, b)
a = b
b += 2 * blockSize
}
if m := a + blockSize; m < n {
symMerge(data, a, m, n)
}
blockSize *= 2
}
}
// symMerge merges the two sorted subsequences data[a:m] and data[m:b] using
// the SymMerge algorithm from Pok-Son Kim and Arne Kutzner, "Stable Minimum
// Storage Merging by Symmetric Comparisons", in Susanne Albers and Tomasz
// Radzik, editors, Algorithms - ESA 2004, volume 3221 of Lecture Notes in
// Computer Science, pages 714-723. Springer, 2004.
//
// Let M = m-a and N = b-n. Wolog M < N.
// The recursion depth is bound by ceil(log(N+M)).
// The algorithm needs O(M*log(N/M + 1)) calls to data.Less.
// The algorithm needs O((M+N)*log(M)) calls to data.Swap.
//
// The paper gives O((M+N)*log(M)) as the number of assignments assuming a
// rotation algorithm which uses O(M+N+gcd(M+N)) assignments. The argumentation
// in the paper carries through for Swap operations, especially as the block
// swapping rotate uses only O(M+N) Swaps.
//
// symMerge assumes non-degenerate arguments: a < m && m < b.
// Having the caller check this condition eliminates many leaf recursion calls,
// which improves performance.
func symMerge(data Interface, a, m, b int) {
// Avoid unnecessary recursions of symMerge
// by direct insertion of data[a] into data[m:b]
// if data[a:m] only contains one element.
if m-a == 1 {
// Use binary search to find the lowest index i
// such that data[i] >= data[a] for m <= i < b.
// Exit the search loop with i == b in case no such index exists.
i := m
j := b
for i < j {
h := int(uint(i+j) >> 1)
if data.Less(h, a) {
i = h + 1
} else {
j = h
}
}
// Swap values until data[a] reaches the position before i.
for k := a; k < i-1; k++ {
data.Swap(k, k+1)
}
return
}
// Avoid unnecessary recursions of symMerge
// by direct insertion of data[m] into data[a:m]
// if data[m:b] only contains one element.
if b-m == 1 {
// Use binary search to find the lowest index i
// such that data[i] > data[m] for a <= i < m.
// Exit the search loop with i == m in case no such index exists.
i := a
j := m
for i < j {
h := int(uint(i+j) >> 1)
if !data.Less(m, h) {
i = h + 1
} else {
j = h
}
}
// Swap values until data[m] reaches the position i.
for k := m; k > i; k-- {
data.Swap(k, k-1)
}
return
}
mid := int(uint(a+b) >> 1)
n := mid + m
var start, r int
if m > mid {
start = n - b
r = mid
} else {
start = a
r = m
}
p := n - 1
for start < r {
c := int(uint(start+r) >> 1)
if !data.Less(p-c, c) {
start = c + 1
} else {
r = c
}
}
end := n - start
if start < m && m < end {
rotate(data, start, m, end)
}
if a < start && start < mid {
symMerge(data, a, start, mid)
}
if mid < end && end < b {
symMerge(data, mid, end, b)
}
}
// rotate rotates two consecutive blocks u = data[a:m] and v = data[m:b] in data:
// Data of the form 'x u v y' is changed to 'x v u y'.
// rotate performs at most b-a many calls to data.Swap,
// and it assumes non-degenerate arguments: a < m && m < b.
func rotate(data Interface, a, m, b int) {
i := m - a
j := b - m
for i != j {
if i > j {
swapRange(data, m-i, m, j)
i -= j
} else {
swapRange(data, m-i, m+j-i, i)
j -= i
}
}
// i == j
swapRange(data, m-i, m, i)
}
参考资料: