JS和Python之间的基准性能差异答案

【问题标题】：Benchmarking performance difference between JS and PythonJS和Python之间的基准性能差异
【发布时间】：2019-02-28 19:12:33
【问题描述】：

到目前为止，我们有一个应用程序只保留在前端，但是在使用大型网格配置对其进行测试时，我们遇到了几个处理和内存限制。因此，我们正在探索将需要资源的任务推送到后端的选项。

因此，我目前正在运行一些基准测试，以了解我们可能会看到的性能差异。更具体地说，鉴于性能瓶颈的很大一部分是在嵌套循环上，我在 js 和 python 中编写了一个简单的 sn-p 来测试处理嵌套循环和构建不同大小的数组的性能。

令我惊讶的是，无论我测试过什么排列，js 似乎总是更快。

JS sn-p:

var timeTotal = 0;
var benchmarkTests = 100;
var testSizes = [50, 100, 500, 1000, 3000];

for (var a = 0; a < testSizes.length; a++) {
  var minX = 0;
  var maxX = testSizes[a];
  var minY = 0;
  var maxY = testSizes[a];
  var cellDimensionsX = 0.991; // cell width
  var cellDimensionsY = 1.652; // cell height
  var cellDimensionsXHalf = cellDimensionsX / 2; // half cell width
  var cellDimensionsYHalf = cellDimensionsY / 2; // half cell height

  var maxCellsXCount = Math.floor(maxX / cellDimensionsX);
  var maxCellsYCount = Math.floor(maxY / cellDimensionsY);

  console.log("X:", maxCellsXCount, "| Y", maxCellsYCount, "| Total", maxCellsXCount * maxCellsYCount);

  for (var k = 0; k < benchmarkTests; k++) {
    var cellCoordsTime = new Date().getTime();

    var cellCoords = {};
    var index = 0;

    for (var i = 0; i < maxCellsXCount; i++) {
      var xCoord = (i * cellDimensionsX) + cellDimensionsXHalf;

      for (var d = 0; d < maxCellsYCount; d++) {
        cellCoords[index] = {
          x: xCoord,
          y: (d * cellDimensionsY) + cellDimensionsYHalf
        };
        index++;
      }
    }

    var thisTime = new Date().getTime() - cellCoordsTime;
    timeTotal += thisTime;
    // console.log('cellCoords', thisTime, 'for grid with', index, 'cells');
  }
  console.log('Testing with a', testSizes[a], '*', testSizes[a], 'grid area. Total time', timeTotal, '. Avg', (timeTotal / benchmarkTests), 'for', benchmarkTests, 'tests');
}

Python sn-p:

#!/usr/bin/python
import time

timeTotal = 0
benchmarkTests = 100

testSizes = [50, 100, 500, 1000, 3000]

for a in range(len(testSizes)):
    minX = 0
    maxX = testSizes[a]
    minY = 0
    maxY = testSizes[a]
    cellDimensionsX = 0.991 # cell width
    cellDimensionsY = 1.652 # cell height
    cellDimensionsXHalf = cellDimensionsX / 2 # half cell width
    cellDimensionsYHalf = cellDimensionsY / 2 # half cell height

    maxCellsXCount = int(maxX / cellDimensionsX)
    maxCellsYCount = int(maxY / cellDimensionsY)

    print("X: %s | Y %s | Total %s" % (maxCellsXCount, maxCellsYCount, maxCellsXCount * maxCellsYCount))

    for k in range(benchmarkTests):
        start = time.time()
        cellCoords = {}
        index = 0

        for i in range(maxCellsXCount):
            xCoord = (i * cellDimensionsX) + cellDimensionsXHalf
            for d in range(maxCellsYCount):
                cellCoords[index] = {'x': xCoord, 'y': (d * cellDimensionsY) + cellDimensionsYHalf}
                index += 1

        thisTime = (time.time() - start) * 1000;
        timeTotal = timeTotal + thisTime;
        # print("Elapsed Time: %s for grid with %s cells" % (thisTime, index))

    print("Testing with a %s*%s grid area. Total time %s. Avg %s for %s tests" % (testSizes[a], testSizes[a], timeTotal, (timeTotal / benchmarkTests), benchmarkTests))

运行这些我得到：

JS：

Testing with a 50 * 50 grid area. Total time 26 . Avg 0.26 for 100 tests
Testing with a 100 * 100 grid area. Total time 85 . Avg 0.85 for 100 tests
Testing with a 500 * 500 grid area. Total time 4539 . Avg 45.39 for 100 tests
Testing with a 1000 * 1000 grid area. Total time 23160 . Avg 231.6 for 100 tests
Testing with a 3000 * 3000 grid area. Total time 243760 . Avg 2437.6 for 100 tests

Python：

Testing with a 50*50 grid area. Total time 50.8642196655. Avg 0.508642196655 for 100 tests
Testing with a 100*100 grid area. Total time 262.931108475. Avg 2.62931108475 for 100 tests
Testing with a 500*500 grid area. Total time 6338.83333206. Avg 63.3883333206 for 100 tests
Testing with a 1000*1000 grid area. Total time 30769.4478035. Avg 307.694478035 for 100 tests
Testing with a 3000*3000 grid area. Total time 304995.391846. Avg 3049.95391846 for 100 tests

所有时间都以毫秒为单位，所有测试都在同一个本地主机上运行。

我本来希望 python 比 js 快得多。我有什么遗漏吗？

更新 #1（将所有内容移至函数中 | 总体收益 ~17%）：

Testing with a 50*50 grid area. Total time 41.2473678589. Avg 0.412473678589 for 100 tests
Testing with a 100*100 grid area. Total time 174.555540085. Avg 1.74555540085 for 100 tests
Testing with a 500*500 grid area. Total time 5617.09475517. Avg 56.1709475517 for 100 tests
Testing with a 1000*1000 grid area. Total time 21199.390173. Avg 211.99390173 for 100 tests
Testing with a 3000*3000 grid area. Total time 255921.251535. Avg 2559.21251535 for 100 tests

更新 #2（xrange 的交换范围 | 更新 #1 后的总体额外增益 ~15% | 与初始代码相比的总增益 ~30%）：

Testing with a 50*50 grid area. Total time 38.7289524078. Avg 0.387289524078 for 100 tests
Testing with a 100*100 grid area. Total time 176.453590393. Avg 1.76453590393 for 100 tests
Testing with a 500*500 grid area. Total time 5346.49443626. Avg 53.4649443626 for 100 tests
Testing with a 1000*1000 grid area. Total time 21618.1008816. Avg 216.181008816 for 100 tests
Testing with a 3000*3000 grid area. Total time 213622.769356. Avg 2136.22769356 for 100 tests

更新 #3（交换列表的字典 | 更新 #2 后的总体额外增益 ~35% | 与初始代码相比的总增益 ~55%）：

Testing with a 50*50 grid area. Total time 20.7185745239. Avg 0.207185745239 for 100 tests
Testing with a 100*100 grid area. Total time 100.9953022. Avg 1.009953022 for 100 tests
Testing with a 500*500 grid area. Total time 3033.61153603. Avg 30.3361153603 for 100 tests
Testing with a 1000*1000 grid area. Total time 12399.708271. Avg 123.99708271 for 100 tests
Testing with a 3000*3000 grid area. Total time 140118.921518. Avg 1401.18921518 for 100 tests

将 JS 与 Python 的 UPDATE #3 匹配（数组的交换对象 | 与初始代码相比总体损失 ~165%）：

Testing with a 50 * 50 grid area. Total time 30 . Avg 0.3 for 100 tests
Testing with a 100 * 100 grid area. Total time 48 . Avg 0.48 for 100 tests
Testing with a 500 * 500 grid area. Total time 12694 . Avg 126.94 for 100 tests
Testing with a 1000 * 1000 grid area. Total time 81402 . Avg 814.02 for 100 tests
Testing with a 3000 * 3000 grid area. Total time 625615 . Avg 6256.15 for 100 tests

更新 #4（将 Cython 拉入战斗 | 更新 #3 后的总体额外增益 ~26% | 与初始代码相比总增益 ~66%）：

Testing with a 50*50 grid area. Total time 30. Avg 0.300 for 100 tests
Testing with a 100*100 grid area. Total time 68. Avg 0.680 for 100 tests
Testing with a 500*500 grid area. Total time 2475. Avg 24.750 for 100 tests
Testing with a 1000*1000 grid area. Total time 9924. Avg 99.240 for 100 tests
Testing with a 3000*3000 grid area. Total time 101697. Avg 1016.970 for 100 tests

更新 #5（键入变量 | 更新 #4 后的总体额外增益 ~23% | 与初始代码相比总增益 ~74%）：

Testing with a 50*50 grid area. Total time 5. Avg 0.048 for 100 tests
Testing with a 100*100 grid area. Total time 43. Avg 0.426 for 100 tests
Testing with a 500*500 grid area. Total time 1851. Avg 18.511 for 100 tests
Testing with a 1000*1000 grid area. Total time 8020. Avg 80.202 for 100 tests
Testing with a 3000*3000 grid area. Total time 78350. Avg 783.502 for 100 tests

（希望是最终版）更新 #6（Cython 相关优化，包括在 2 个内核上并行化 | 更新 #5 后的总体额外增益 ~88% | 与初始代码相比总增益 ~97%）：强>

Testing with a 50*50 grid area. Total time 0.668. Avg 0.007 for 100 tests
Testing with a 100*100 grid area. Total time 1.584. Avg 0.016 for 100 tests
Testing with a 500*500 grid area. Total time 57.374. Avg 0.574 for 100 tests
Testing with a 1000*1000 grid area. Total time 521.210. Avg 5.212 for 100 tests
Testing with a 3000*3000 grid area. Total time 10113.633. Avg 101.136 for 100 tests

与原始实现相比，当前版本平均快 30 到 35 倍。所以暂时不叫了。

最终更新（在超线程 8 核机器上测试（不再是 localhost，也不再是苹果之间的比较）| 进一步的 cython 相关优化，并缓存 y 的值 | 更新 #6 后的总体额外收益 ~78% | 与初始代码相比的总体收益 ~99.3%）：

Testing with a 50*50 grid area. Total time 0.498. Avg 0.005 for 100 tests
Testing with a 100*100 grid area. Total time 1.146. Avg 0.011 for 100 tests
Testing with a 500*500 grid area. Total time 22.856. Avg 0.229 for 100 tests
Testing with a 1000*1000 grid area. Total time 113.819. Avg 1.138 for 100 tests
Testing with a 3000*3000 grid area. Total time 2228.098ms. Avg 22.281 for 100 tests
Testing with a 10000*10000 grid area. Total time 29407.874ms. Avg 294.79 for 100 tests
Testing with a 20000*20000 grid area. Total time 157185.469ms. Avg 1571.855 for 100 tests

【问题讨论】：

Python 处理全局变量很慢，先把你的代码放在一个函数中。
另见stackoverflow.com/questions/3033329/…
将 range(..) 更改为 xrange(..) 可以让您的速度再提高 2-3%。
此时，切换到 Python 3 可能也是一个好主意；它在某些方面较慢，但在其他方面较快，而且与 Python 2 不同，它不会在不到一年的时间内停止支持。然后你不必切换到xrange（因为Python 3 range 已经是Python 2 的xrange 的改进版本）。
如果性能是一个问题，您想使用最新的 Python 3 版本。他们在 3.6、3.7 中添加了许多加速。大幅提速。

标签： javascript python python-2.7 performance benchmarking

【解决方案1】：

为 Python 使用更合适的数据类型（元组而不是创建 {x: .., y: ..} dict，使用列表而不是具有连续整数索引的 dict），使用 xrange 而不是 range，并将所有内容包装到一个函数中，给出以下内容代码：

import time


def foo():
    timeTotal = 0
    benchmarkTests = 100

    testSizes = [50, 100, 500, 1000, 3000]

    for a in range(len(testSizes)):
        minX = 0
        maxX = testSizes[a]
        minY = 0
        maxY = testSizes[a]
        cellDimensionsX = 0.991 # cell width
        cellDimensionsY = 1.652 # cell height
        cellDimensionsXHalf = cellDimensionsX / 2 # half cell width
        cellDimensionsYHalf = cellDimensionsY / 2 # half cell height

        maxCellsXCount = int(maxX / cellDimensionsX)
        maxCellsYCount = int(maxY / cellDimensionsY)

        print("X: %s | Y %s | Total %s" % (maxCellsXCount, maxCellsYCount, maxCellsXCount * maxCellsYCount))

        for k in xrange(benchmarkTests):
            start = time.time()
            # cellCoords = {}
            cellCoords = []

            for i in xrange(maxCellsXCount):
                xCoord = (i * cellDimensionsX) + cellDimensionsXHalf
                for d in xrange(maxCellsYCount):
                    # cellCoords[index] = {'x': xCoord, 'y': (d * cellDimensionsY) + cellDimensionsYHalf}
                    # cellCoords[index] = (xCoord, (d * cellDimensionsY) + cellDimensionsYHalf)
                    cellCoords.append((xCoord, (d * cellDimensionsY) + cellDimensionsYHalf))
                    # index += 1

            thisTime = (time.time() - start) * 1000;
            timeTotal = timeTotal + thisTime;
            # print("Elapsed Time: %s for grid with %s cells" % (thisTime, index))

        print("Testing with a %s*%s grid area. Total time %s. Avg %s for %s tests" % (testSizes[a], testSizes[a], timeTotal, (timeTotal / benchmarkTests), benchmarkTests))

foo()

提供大约 50% 的加速：

X: 50 | Y 30 | Total 1500
Testing with a 50*50 grid area. Total time 33.9999198914. Avg 0.339999198914 for 100 tests
X: 100 | Y 60 | Total 6000
Testing with a 100*100 grid area. Total time 191.999912262. Avg 1.91999912262 for 100 tests
X: 504 | Y 302 | Total 152208
Testing with a 500*500 grid area. Total time 4790.99988937. Avg 47.9099988937 for 100 tests
X: 1009 | Y 605 | Total 610445
Testing with a 1000*1000 grid area. Total time 24529.9999714. Avg 245.299999714 for 100 tests
X: 3027 | Y 1815 | Total 5494005
Testing with a 3000*3000 grid area. Total time 201085.000038. Avg 2010.85000038 for 100 tests

与我机器上的原始代码相比：

X: 50 | Y 30 | Total 1500
Testing with a 50*50 grid area. Total time 94.0001010895. Avg 0.940001010895 for 100 tests
X: 100 | Y 60 | Total 6000
Testing with a 100*100 grid area. Total time 495.000123978. Avg 4.95000123978 for 100 tests
X: 504 | Y 302 | Total 152208
Testing with a 500*500 grid area. Total time 10732.0001125. Avg 107.320001125 for 100 tests
X: 1009 | Y 605 | Total 610445
Testing with a 1000*1000 grid area. Total time 49074.0001202. Avg 490.740001202 for 100 tests
X: 3027 | Y 1815 | Total 5494005
Traceback (most recent call last):
  File "pyperf-orig.py", line 31, in <module>
    cellCoords[index] = {'x': xCoord, 'y': (d * cellDimensionsY) + cellDimensionsYHalf}
MemoryError

【讨论】：

感谢一百万为我指引正确的方向。

【解决方案2】：

一个改进是将所有内容都放入一个函数中，因为显然 Python 对全局变量不太好：

运行您的代码：

Testing with a 50*50 grid area. Total time 52.17409133911133. Avg 0.5217409133911133 for 100 tests
Testing with a 100*100 grid area. Total time 262.12358474731445. Avg 2.6212358474731445 for 100 tests
Testing with a 500*500 grid area. Total time 6206.289052963257. Avg 62.06289052963257 for 100 tests
Testing with a 1000*1000 grid area. Total time 30345.27611732483. Avg 303.4527611732483 for 100 tests

将代码放入函数中时，速度会快很多：

Testing with a 50*50 grid area. Total time 37.11581230163574. Avg 0.3711581230163574 for 100 tests
Testing with a 100*100 grid area. Total time 176.71799659729004. Avg 1.7671799659729004 for 100 tests
Testing with a 500*500 grid area. Total time 4659.825325012207. Avg 46.59825325012207 for 100 tests
Testing with a 1000*1000 grid area. Total time 23246.346712112427. Avg 232.46346712112427 for 100 tests

【讨论】：