c# - 在 linq 中，为什么后续调用 IEnumerable.Intersect 的速度要快得多

Question

一边看这个问题C# Similarities of two arrays有人指出，最初的 linq 调用比后续调用慢得多。缓存的是什么造成了如此大的不同？我感兴趣的是我们何时可以期望实现这种行为(也许这里只是因为反复使用相同的列表)。

    static void Main(string[] args)
    {
        var a = new List<int>() { 7, 17, 21, 29, 30, 33, 40, 42, 51, 53, 60, 63, 66, 68, 70, 84, 85, 91, 101, 102, 104, 108, 109, 112, 115, 116, 118, 125, 132, 137, 139, 142, 155, 163, 164, 172, 174, 176, 179, 184, 185, 186, 187, 188, 189, 192, 197, 206, 209, 234, 240, 244, 249, 250, 252, 253, 254, 261, 263, 270, 275, 277, 290, 292, 293, 304, 308, 310, 314, 316, 319, 321, 322, 325, 326, 327, 331, 332, 333, 340, 367, 371, 374, 403, 411, 422, 427, 436, 440, 443, 444, 446, 448, 449, 450, 452, 455, 459, 467, 470, 487, 488, 489, 492, 494, 502, 503, 505, 513, 514, 522, 523, 528, 532, 534, 535, 545, 547, 548, 553, 555, 556, 565, 568, 570, 577, 581, 593, 595, 596, 598, 599, 606, 608, 613, 615, 630, 638, 648, 661, 663, 665, 669, 673, 679, 681, 685, 687, 690, 697, 702, 705, 708, 710, 716, 719, 724, 725, 727, 728, 732, 733, 739, 744, 760, 762, 775, 781, 787, 788, 790, 795, 797, 802, 806, 808, 811, 818, 821, 822, 829, 835, 845, 848, 851, 859, 864, 866, 868, 875, 881, 898, 899, 906, 909, 912, 913, 915, 916, 920, 926, 929, 930, 933, 937, 945, 946, 949, 954, 957, 960, 968, 975, 980, 985, 987, 989, 995 };
        var b = new List<int>() { 14, 20, 22, 23, 32, 36, 40, 48, 63, 65, 67, 71, 83, 87, 90, 100, 104, 109, 111, 127, 128, 137, 139, 141, 143, 148, 152, 153, 157, 158, 161, 163, 166, 187, 192, 198, 210, 211, 217, 220, 221, 232, 233, 236, 251, 252, 254, 256, 257, 272, 273, 277, 278, 283, 292, 304, 305, 307, 321, 333, 336, 341, 342, 344, 349, 355, 356, 359, 366, 373, 379, 386, 387, 392, 394, 396, 401, 409, 412, 433, 437, 441, 445, 447, 452, 465, 471, 476, 479, 483, 511, 514, 516, 521, 523, 531, 544, 548, 551, 554, 559, 562, 566, 567, 571, 572, 574, 576, 586, 592, 593, 597, 600, 602, 615, 627, 631, 636, 644, 650, 655, 657, 660, 667, 670, 680, 691, 697, 699, 703, 704, 706, 707, 716, 742, 748, 751, 754, 766, 770, 779, 785, 788, 790, 802, 803, 806, 811, 812, 815, 816, 821, 824, 828, 841, 848, 853, 863, 866, 870, 872, 875, 879, 880, 882, 883, 885, 886, 887, 888, 892, 894, 902, 905, 909, 912, 913, 914, 916, 920, 922, 925, 926, 928, 930, 935, 936, 938, 942, 945, 952, 954, 955, 957, 959, 960, 961, 963, 970, 974, 976, 979, 987 };
        var s = new System.Diagnostics.Stopwatch();
        const int cycles = 10;
        for (int i = 0; i < cycles; i++)
        {
            s.Start();
            var z= a.Intersect(b);
            s.Stop();
            Console.WriteLine("Test 1-{0}: {1} {2}", i, s.ElapsedTicks, z.Count());
            s.Reset();
            a[0]=i;//simple attempt to make sure entire result isn't cached
        }

        for (int i = 0; i < cycles; i++)
        {
            var z1 = new List<int>(a.Count);
            s.Start();
            int j = 0;
            int b1 = b[j];
            foreach (var a1 in a)
            {
                while (b1 <= a1)
                {
                    if (b1 == a1)
                        z1.Add(b[j]);
                    j++;
                    if (j >= b.Count)
                        break;
                    b1 = b[j];
                }
            }
            s.Stop();
            Console.WriteLine("Test 2-{0}: {1} {2}", i, s.ElapsedTicks, z1.Count);
            s.Reset();
            a[0]=i;//simple attempt to make sure entire result isn't cached
        }

        Console.Write("Press Enter to quit");
        Console.ReadLine();
    }
}

根据某些人的要求 - 示例输出:

Test 1-0: 2900 45
Test 1-1: 2 45
Test 1-2: 0 45
Test 1-3: 1 45

(正常循环只显示连续运行之间的细微差别)

注意更改后调用 a.Intersect(b).ToArray(); 而不是 just a.Intersect(b); 正如 @kerem 所建议的结果成为:

Test 1-0: 13656 45
Test 1-1: 113 45
Test 1-2: 76 45
Test 1-3: 64 45
Test 1-4: 90 45 
...

Answer 1

我预计任何循环的第一次运行都会变慢，原因有以下三个:

1. 必须在第一次对代码进行 jit 处理，但以后不必。
2. 如果运行的可执行代码足够小以适合缓存，那么它就不会被逐出，CPU 加载速度也会更快。
3. 如果数据小到足以放入缓存，那么它就不会被逐出，CPU 加载速度也会更快。