让代码飞起来——高性能Julia学习笔记（三）

前面两篇让代码飞起来——高性能 Julia 学习笔记（一） 让代码飞起来——高性能 Julia 学习笔记（二）， 介绍了如何写出高性能的 Julia 代码， 这篇结合我最近的项目， 简单测试对比一下各种语言用 monte carlo 算法计算 pi 的效率。

首先声明一下， 本文不能算严格意义上的性能测试， 也不想挑起语言圣战， 个人能力有限， 实现的不同语言版本代码也未必是最高效的， 基本都是 naive 实现。

如果对 Monte Carlo 算法不熟悉， 可以参考下面两个资料， 我就不浪费时间重复了：

• https://zh.wikipedia.org/wiki...
• http://www.ruanyifeng.com/blo...

机器是 2015 年的 MacPro：

Processor: 2.5GHz Intel Core i7
Memory: 16GB 1600 MHZ DDR3
Os: macOS High Sierra Version 10.13.4

JS 版本

function pi(n) {
  let inCircle = 0;
  for (let i = 0; i <= n; i++) {
    x = Math.random();
    y = Math.random();
    if (x * x + y * y < 1.0) {
      inCircle += 1;
    }
  }
  return (4.0 * inCircle) / n;
}
const N = 100000000;
console.log(pi(N));

结果：

➜  me.magicly.performance git:(master) ✗ node --version
v10.11.0
➜  me.magicly.performance git:(master) ✗ time node mc.js
3.14174988
node mc.js  10.92s user 0.99s system 167% cpu 7.091 total

Go 版本

package main

import (
    "math/rand"
)

func PI(samples int) (result float64) {
    inCircle := 0
    r := rand.New(rand.NewSource(42))

    for i := 0; i < samples; i++ {
        x := r.Float64()
        y := r.Float64()
        if (x*x + y*y) < 1 {
            inCircle++
        }
    }

    return float64(inCircle) / float64(samples) * 4.0
}

func main() {
    samples := 100000000
    PI(samples)
}

结果：

➜  me.magicly.performance git:(master) ✗ go version
go version go1.11 darwin/amd64
➜  me.magicly.performance git:(master) ✗ time go run monte_carlo.go
go run monte_carlo.go  2.17s user 0.10s system 101% cpu 2.231 total

C 版本

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <string.h>
#define SEED 42

int main(int argc, char **argv)
{
  int niter = 100000000;
  double x, y;
  int i, count = 0;
  double z;
  double pi;

  srand(SEED);
  count = 0;
  for (i = 0; i < niter; i++)
  {
    x = (double)rand() / RAND_MAX;
    y = (double)rand() / RAND_MAX;
    z = x * x + y * y;
    if (z <= 1)
      count++;
  }
  pi = (double)count / niter * 4;
  printf("# of trials= %d , estimate of pi is %g \n", niter, pi);
}

结果：

➜  me.magicly.performance git:(master) ✗ gcc --version
Configured with: --prefix=/Applications/Xcode.app/Contents/Developer/usr --with-gxx-include-dir=/usr/include/c++/4.2.1
Apple LLVM version 9.1.0 (clang-902.0.39.2)
Target: x86_64-apple-darwin17.5.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin
➜  me.magicly.performance git:(master) ✗ gcc -O2 -o mc-pi-c mc-pi.c
➜  me.magicly.performance git:(master) ✗ time ./mc-pi-c
# of trials= 100000000 , estimate of pi is 3.14155
./mc-pi-c  1.22s user 0.00s system 99% cpu 1.226 total

C++ 版本

#include <iostream>
#include <cstdlib> //defines rand(), srand(), RAND_MAX
#include <cmath>   //defines math functions

using namespace std;

int main()
{
  const int SEED = 42;
  int interval, i;
  double x, y, z, pi;
  int inCircle = 0;

  srand(SEED);

  const int N = 100000000;
  for (i = 0; i < N; i++)
  {
    x = (double)rand() / RAND_MAX;
    y = (double)rand() / RAND_MAX;

    z = x * x + y * y;
    if (z < 1)
    {
      inCircle++;
    }
  }
  pi = double(4 * inCircle) / N;

  cout << "\nFinal Estimation of Pi = " << pi << endl;
  return 0;
}

结果：

➜  me.magicly.performance git:(master) ✗ c++ --version
Apple LLVM version 9.1.0 (clang-902.0.39.2)
Target: x86_64-apple-darwin17.5.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin
➜  me.magicly.performance git:(master) ✗ c++ -O2 -o mc-pi-cpp mc-pi.cpp
➜  me.magicly.performance git:(master) ✗ time ./mc-pi-cpp

Final Estimation of Pi = 3.14155
./mc-pi-cpp  1.23s user 0.01s system 99% cpu 1.239 total

Julia 版本

function pi(N::Int)
  inCircle = 0
  for i = 1:N
      x = rand() * 2 - 1
      y = rand() * 2 - 1

      r2 = x*x + y*y
      if r2 < 1.0
          inCircle += 1
      end
  end

  return inCircle / N * 4.0
end

N = 100_000_000
println(pi(N))

结果：

➜  me.magicly.performance git:(master) ✗ julia
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.0.1 (2018-09-29)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> versioninfo()
Julia Version 1.0.1
Commit 0d713926f8 (2018-09-29 19:05 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin14.5.0)
  CPU: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, haswell)

➜  me.magicly.performance git:(master) ✗ time julia mc.jl
3.14179496
julia mc.jl  0.85s user 0.17s system 144% cpu 0.705 total

另外 Rust 开发环境升级搞出了点问题， 没弄好， 不过根据之前的经验， 我估计跟 C++差不多。

github 上找到一份对比， 包含了更多的语言， 有兴趣的可以参考一下https://gist.github.com/jmoir... ， LuaJIT 居然跟 Rust 差不多一样快， 跟 Julia 官网的 benchmark 比较一致https://julialang.org/benchma... 。

另外实现了两个 Go 的并发版本：

package main

import (
    "fmt"
    "math/rand"
    "runtime"
    "time"
)

type Job struct {
    n int
}

var threads = runtime.NumCPU()
var rands = make([]*rand.Rand, 0, threads)

func init() {
    fmt.Printf("cpus: %d\n", threads)
    runtime.GOMAXPROCS(threads)

    for i := 0; i < threads; i++ {
        rands = append(rands, rand.New(rand.NewSource(time.Now().UnixNano())))
    }
}

func MultiPI2(samples int) float64 {
    t1 := time.Now()

    threadSamples := samples / threads

    jobs := make(chan Job, 100)
    results := make(chan int, 100)

    for w := 0; w < threads; w++ {
        go worker2(w, jobs, results, threadSamples)
    }

    go func() {
        for i := 0; i < threads; i++ {
            jobs <- Job{
                n: i,
            }
        }
        close(jobs)
    }()

    var total int
    for i := 0; i < threads; i++ {
        total += <-results
    }

    result := float64(total) / float64(samples) * 4
    fmt.Printf("MultiPI2: %d times, value: %f, cost: %s\n", samples, result, time.Since(t1))
    return result
}
func worker2(id int, jobs <-chan Job, results chan<- int, threadSamples int) {
    for range jobs {
        // fmt.Printf("worker id: %d, job: %v, remain jobs: %d\n", id, job, len(jobs))
        var inside int
        // r := rand.New(rand.NewSource(time.Now().UnixNano()))
        r := rands[id]
        for i := 0; i < threadSamples; i++ {
            x, y := r.Float64(), r.Float64()

            if x*x+y*y <= 1 {
                inside++
            }
        }
        results <- inside
    }
}

func MultiPI(samples int) float64 {
    t1 := time.Now()

    threadSamples := samples / threads
    results := make(chan int, threads)

    for j := 0; j < threads; j++ {
        go func() {
            var inside int
            r := rand.New(rand.NewSource(time.Now().UnixNano()))
            for i := 0; i < threadSamples; i++ {
                x, y := r.Float64(), r.Float64()

                if x*x+y*y <= 1 {
                    inside++
                }
            }
            results <- inside
        }()
    }

    var total int
    for i := 0; i < threads; i++ {
        total += <-results
    }

    result := float64(total) / float64(samples) * 4
    fmt.Printf("MultiPI: %d times, value: %f, cost: %s\n", samples, result, time.Since(t1))
    return result
}

func PI(samples int) (result float64) {
    t1 := time.Now()
    var inside int = 0
    r := rand.New(rand.NewSource(time.Now().UnixNano()))

    for i := 0; i < samples; i++ {
        x := r.Float64()
        y := r.Float64()
        if (x*x + y*y) < 1 {
            inside++
        }
    }

    ratio := float64(inside) / float64(samples)

    result = ratio * 4

    fmt.Printf("PI: %d times, value: %f, cost: %s\n", samples, result, time.Since(t1))

    return
}

func main() {
    samples := 100000000
    PI(samples)
    MultiPI(samples)
    MultiPI2(samples)
}

结果：

➜  me.magicly.performance git:(master) ✗ time go run monte_carlo.1.go
cpus: 8
PI: 100000000 times, value: 3.141778, cost: 2.098006252s
MultiPI: 100000000 times, value: 3.141721, cost: 513.008435ms
MultiPI2: 100000000 times, value: 3.141272, cost: 485.336029ms
go run monte_carlo.1.go  9.41s user 0.18s system 285% cpu 3.357 total

可以看出， 效率提升了 4 倍。 为什么明明有8 个 CPU， 只提升了 4 倍呢？ 其实我的 macpro 就是 4 核的， 8 是超线程出来的虚拟核，在 cpu 密集计算上并不能额外提升效率。 可以参考这篇文章： 物理 CPU、CPU 核数、逻辑 CPU、超线程 。

下一篇，我们就来看一下 Julia 中如何利用并行进一步提高效率。

欢迎加入知识星球一起分享讨论有趣的技术话题。