PHP 对象数组不是线性缩放的，而全局数组却可以？

使用对象内数组作为属性与使用全局 php 数组变量相比，存在一个主要的性能问题，为什么？

为了对这个问题进行基准测试，我创建了以下基准测试，它存储一个越来越大的数组，并以 stdClass 作为节点，运行了两个测试，一个使用类中的数组属性，另一个使用全局数组。

测试代码

ini_set('memory_limit', '2250M');
class MyTest {
    public $storage = [];
    public function push(){
        $this->storage[] = [new stdClass()];
    }
}

echo "Testing Objects".PHP_EOL;
for($size = 1000; $size < 5000000; $size *= 2) {
    $start = milliseconds();
    for ($a=new MyTest(), $i=0;$i<$size;$i++) {
        $a->push();
    }
    $end = milliseconds();
    echo "Array Size $size".PHP_EOL;
    echo $end - $start . " milliseconds to perform".PHP_EOL;
}
echo "================".PHP_EOL;
echo "Testing Array".PHP_EOL;
for($size = 1000; $size < 5000000; $size *= 2) {
    $start = milliseconds();
    for ($a=[], $i=0;$i<$size;$i++) {
        $a[] = [new stdClass()];
    }
    $end = milliseconds();
    echo "Array Size $size".PHP_EOL;
    echo $end - $start . " milliseconds to perform".PHP_EOL;
}

以及令人震惊的结果：

Testing Objects
Array Size 1000
2 milliseconds to perform
Array Size 2000
3 milliseconds to perform
Array Size 4000
6 milliseconds to perform
Array Size 8000
12 milliseconds to perform
Array Size 16000
35 milliseconds to perform
Array Size 32000
97 milliseconds to perform
Array Size 64000
246 milliseconds to perform
Array Size 128000
677 milliseconds to perform
Array Size 256000
2271 milliseconds to perform
Array Size 512000
9244 milliseconds to perform
Array Size 1024000
31186 milliseconds to perform
Array Size 2048000
116123 milliseconds to perform
Array Size 4096000
495588 milliseconds to perform
================
Testing Array
Array Size 1000
1 milliseconds to perform
Array Size 2000
2 milliseconds to perform
Array Size 4000
4 milliseconds to perform
Array Size 8000
8 milliseconds to perform
Array Size 16000
28 milliseconds to perform
Array Size 32000
61 milliseconds to perform
Array Size 64000
114 milliseconds to perform
Array Size 128000
245 milliseconds to perform
Array Size 256000
494 milliseconds to perform
Array Size 512000
970 milliseconds to perform
Array Size 1024000
2003 milliseconds to perform
Array Size 2048000
4241 milliseconds to perform
Array Size 4096000
14260 milliseconds to perform

现在，除了对象调用自身的明显开销之外，当数组变大时，对象数组属性有时会花费 3 - 4 倍的时间来扩展，但标准全局数组变量的情况并非如此。

关于这个问题有什么想法或答案吗？这是否是 PHP 引擎可能存在的错误？

我在 PHP 5.3.9 上测试了你的代码。为此我必须翻译[] to array()，我还必须纠正你的第 12 行：来自$a=new MyTest($size), to $mytest=new MyTest($size)（顺便说一句，构造函数参数被默默地忽略，有趣）。我还添加了这段代码：

echo "================".PHP_EOL;
echo "Testing Function".PHP_EOL;
for($size = 1000; $size < 1000000; $size *= 2) {
    $start = milliseconds();
    for ($a=array(), $i=0;$i<$size;$i++) {
        my_push($a);
    }
    $end = milliseconds();
    echo "Array Size $size".PHP_EOL;
    echo $end - $start . " milliseconds to perform".PHP_EOL;
    echo "memory usage: ".memory_get_usage()." , real: ".memory_get_usage(true).PHP_EOL;
}

function my_push(&$a)
{
   $a[] = array(new stdClass());
}

我在同一点将内存使用情况添加到循环中，添加了unset($mytest);在对象情况之后（以获得更一致的内存日志），并且还将您的 5000000 替换为 1000000，因为我只有 2GB RAM。这就是我得到的：

Testing Objects
Array Size 1000
2 milliseconds to perform
memory usage: 1666376 , real: 1835008
Array Size 2000
5 milliseconds to perform
memory usage: 2063280 , real: 2097152
Array Size 4000
10 milliseconds to perform
memory usage: 2857008 , real: 2883584
Array Size 8000
19 milliseconds to perform
memory usage: 4444456 , real: 4718592
Array Size 16000
44 milliseconds to perform
memory usage: 7619392 , real: 8126464
Array Size 32000
103 milliseconds to perform
memory usage: 13969256 , real: 14417920
Array Size 64000
239 milliseconds to perform
memory usage: 26668936 , real: 27262976
Array Size 128000
588 milliseconds to perform
memory usage: 52068368 , real: 52690944
Array Size 256000
1714 milliseconds to perform
memory usage: 102867104 , real: 103546880
Array Size 512000
5452 milliseconds to perform
memory usage: 204464624 , real: 205258752
================
Testing Array
Array Size 1000
1 milliseconds to perform
memory usage: 18410640 , real: 20709376
Array Size 2000
4 milliseconds to perform
memory usage: 18774760 , real: 20709376
Array Size 4000
7 milliseconds to perform
memory usage: 19502976 , real: 20709376
Array Size 8000
13 milliseconds to perform
memory usage: 20959360 , real: 21233664
Array Size 16000
29 milliseconds to perform
memory usage: 23872176 , real: 24379392
Array Size 32000
61 milliseconds to perform
memory usage: 29697720 , real: 30146560
Array Size 64000
124 milliseconds to perform
memory usage: 41348856 , real: 41943040
Array Size 128000
280 milliseconds to perform
memory usage: 64651088 , real: 65273856
Array Size 256000
534 milliseconds to perform
memory usage: 111255536 , real: 111935488
Array Size 512000
1085 milliseconds to perform
memory usage: 204464464 , real: 205258752
================
Testing Function
Array Size 1000
357 milliseconds to perform
memory usage: 18410696 , real: 22544384
Array Size 2000
4 milliseconds to perform
memory usage: 18774768 , real: 22544384
Array Size 4000
9 milliseconds to perform
memory usage: 19503008 , real: 22544384
Array Size 8000
17 milliseconds to perform
memory usage: 20959392 , real: 22544384
Array Size 16000
36 milliseconds to perform
memory usage: 23872208 , real: 24379392
Array Size 32000
89 milliseconds to perform
memory usage: 29697720 , real: 30146560
Array Size 64000
224 milliseconds to perform
memory usage: 41348888 , real: 41943040
Array Size 128000
529 milliseconds to perform
memory usage: 64651088 , real: 65273856
Array Size 256000
1587 milliseconds to perform
memory usage: 111255616 , real: 111935488
Array Size 512000
5244 milliseconds to perform
memory usage: 204464512 , real: 205258752

正如您所看到的，在函数调用内附加到数组的成本几乎与在原始方法调用内执行附加操作的成本相同（并且具有相同的非线性行为）。可以肯定地说一件事：

正是函数调用消耗了 CPU 时间！

关于非线性行为，只有在超过一定阈值时才会变得非常明显。虽然所有三种情况都具有相同的内存行为（由于不完整的垃圾收集，这仅在本日志中的“普通数组”和“数组内部函数”情况中很明显），但它是“数组内部方法”和“函数内的数组”具有相同执行时间行为的情况。这意味着函数调用本身会导致时间非线性增加。在我看来，可以这么说：

函数调用期间的数据量会影响其持续时间。

为了验证这一点，我替换了所有$a[] with $a[0]以及所有 1000000 和 5000000 （以获得类似的总执行时间）并获得以下输出：

Testing Objects
Array Size 1000
2 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 2000
4 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 4000
8 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 8000
15 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 16000
31 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 32000
62 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 64000
123 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 128000
246 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 256000
493 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 512000
985 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 1024000
1978 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 2048000
3965 milliseconds to perform
memory usage: 1302672 , real: 1572864
Array Size 4096000
7905 milliseconds to perform
memory usage: 1302672 , real: 1572864
================
Testing Array
Array Size 1000
1 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 2000
3 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 4000
5 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 8000
10 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 16000
20 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 32000
40 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 64000
80 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 128000
161 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 256000
322 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 512000
646 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 1024000
1285 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 2048000
2574 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 4096000
5142 milliseconds to perform
memory usage: 1302464 , real: 1572864
================
Testing Function
Array Size 1000
1 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 2000
4 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 4000
6 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 8000
14 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 16000
26 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 32000
53 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 64000
105 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 128000
212 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 256000
422 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 512000
844 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 1024000
1688 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 2048000
3377 milliseconds to perform
memory usage: 1302464 , real: 1572864
Array Size 4096000
6814 milliseconds to perform
memory usage: 1302464 , real: 1572864

请注意，现在时间几乎是完美线性的。当然，现在数组大小固定为 1。另请注意，这三种情况的执行时间差异不如以前那么明显。请记住，最里面的操作在所有情况下都是相同的。

我不会尝试完全解释所有这些（函数调用时的垃圾收集？内存碎片？...？），但我认为我仍然为这里的每个人和我自己收集了一些有用的信息。