A Benchmark comparative analysis tool

#In Go, you can easily perform performance testing on a certain function point by writing a Benchmark function. For important functions, we can add corresponding testing processes in CI/CD to be able to detect changes in function performance in a timely manner. Then the question comes, how to detect the performance changes of the function?

In other words, you write a certain function but find that it runs very slowly and you need to optimize the function. When you search on Google and find a better The implementation method is to use the Benchmark function to find that it is indeed faster. But you can't tell how much it has become faster. You want to know the performance comparison before and after function optimization, how many percentage points it has improved, and is it highly reliable?

For the above demand scenarios, there is a tool that can help you, it is benchstat.

Benchmark Example

Let’s review the benchmark first. To facilitate understanding, here we take the classic calculation of Fibonacci sequence values ​​as an example.

func FibSolution(n int) int {
 if n < 2 {
  return n
 }

 return FibSolution(n-1) + FibSolution(n-2)
}

The above code is a recursive implementation. Obviously, when n becomes larger and larger, the operation of this function will become very time-consuming. Taking n as 20 as an example, the Benchmark function is as follows

func BenchmarkFib20(b *testing.B) {
 for i := 0; i < b.N; i++ {
  FibSolution(20)
 }
}

Command line execution<span style="font-size: 15px;">go test -bench=BenchmarkFib20</span> Get performance results

BenchmarkFib20-8           39452             30229 ns/op

其中，-8 代表的是 8 cpu，函数运行次数为 39452，每次函数的平均花费时间为 30229ns。如果我们想得到多次样本数据，可以指定 go test 的 <span style="font-size: 15px;">-count=N</span> 参数。例如想得到 5 次样本数据，则执行<span style="font-size: 15px;">go test -bench=BenchmarkFib20 -count=5</span>

BenchmarkFib20-8           39325             30297 ns/op
BenchmarkFib20-8           39216             30349 ns/op
BenchmarkFib20-8           39901             30251 ns/op
BenchmarkFib20-8           39336             30455 ns/op
BenchmarkFib20-8           39423             30894 ns/op

计算斐波那契数列值的迭代式实现如下

func FibSolution(n int) int {
 if n < 2 {
  return n
 }
 p, q, r := 0, 0, 1
 for i := 2; i <= n; i++ {
  p = q
  q = r
  r = p + q
 }
 return r
}

对比这两种函数的性能差异，最朴素的方式就是分别对这两个函数进行基准测试，然后通过手工分析这些基准测试结果，但是这并不直观。

benchstat

benchstat 是 Go 官方推荐的一款命令行工具，它用于计算和比较基准测试的相关统计数据。

我们可以通过以下命令进行安装

go install golang.org/x/perf/cmd/benchstat@latest

执行 -h 参数可以看到该工具的使用描述

~ $ benchstat -h
usage: benchstat [options] old.txt [new.txt] [more.txt ...]
options:
  -alpha α
     consider change significant if p < α (default 0.05)
  -csv
     print results in CSV form
  -delta-test test
     significance test to apply to delta: utest, ttest, or none (default "utest")
  -geomean
     print the geometric mean of each file
  -html
     print results as an HTML table
  -norange
     suppress range columns (CSV only)
  -sort order
     sort by order: [-]delta, [-]name, none (default "none")
  -split labels
     split benchmarks by labels (default "pkg,goos,goarch")

我们想比较 FibSolution(n) 从 15 到 20，两种实现方式的性能基准测试。

$ go test -bench=. -count=5 | tee old.txt
$ go test -bench=. -count=5 | tee new.txt

注意，这两条命令执行时，分别对应 FibSolution 函数采用递归式和迭代式实现逻辑。

此时，我们可以对这两个函数实现逻辑进行性能对比

 $ benchstat old.txt new.txt 
name     old time/op  new time/op  delta
Fib15-8  2.67µs ± 2%  0.01µs ± 5%  -99.81%  (p=0.008 n=5+5)
Fib16-8  4.20µs ± 1%  0.01µs ± 2%  -99.87%  (p=0.008 n=5+5)
Fib17-8  6.81µs ± 0%  0.01µs ± 2%  -99.92%  (p=0.008 n=5+5)
Fib18-8  11.1µs ± 1%   0.0µs ± 1%  -99.95%  (p=0.008 n=5+5)
Fib19-8  18.0µs ± 2%   0.0µs ± 4%  -99.97%  (p=0.008 n=5+5)
Fib20-8  29.2µs ± 1%   0.0µs ± 3%  -99.98%  (p=0.008 n=5+5)

可以看到，递归式实现的函数，他的执行时间随着 n 值变大增加非常明显。迭代式实现方式，相较于递归式，它的平均时间开销降低了 99 % 以上，优化效果非常明显。

另外，p=0.008 表示结果的可信程度，p 值越大表明可信度越低。一般以 0.05 作为临界值，超过该值，则结果不可信。n=5+5 表示分别使用的有效样本数量。

Summary

benchstat is a benchmark statistical tool that can be used to reduce the cost of manual analysis of data when we do some optimization work.

If your project deploys automated tests in the CI/CD process, you may wish to add this tool. When changes are made to functions that increase performance loss, it may help you detect problems in advance.

##

The above is the detailed content of A Benchmark comparative analysis tool. For more information, please follow other related articles on the PHP Chinese website!