與其想當然的 overdesign，不如自己動手做個試驗

Conmajia
Jan. 29th, 2019

早在2012年，我曾經針對 C# System.Random 不同的初始化方案專門做過一次試驗，得出了單次默認初始化即可獲得質量很好的隨機數的結論。可是這么多年過去，C# 從2.0升到了4.7，還能在網上看到很多新手（甚至是老鳥）被一些想當然的奇怪思想誤導，費時費力地脫褲子放屁。

▲ 萬千被誤導的靈魂

有些人總覺得用點額外的、生僻的玩意兒會顯得很炫技，很厲害。正如修真小說里，稀有的古代神器多半比量產的現代裝備牛逼，在編程的時候，用上那么幾個不常用，或者當前問題環境下一般沒人用的patterns or callbacks什么的，仿佛就擁有了高貴的血統，讓人不明覺厲。然而再華麗的古羅馬戰車，也比不上穿梭在街頭巷尾的五菱宏光；再炫酷的開發技巧，也敵不過沒頭沒腦的overdesign。

以前有種說法，覺得用隨機性相當大的GUID做種子來初始化Random就能得到比new Random()更隨機的輸出結果。這種不知從何而來的莫名自信一直延續到了今天。

▲ 不管合適不合適，先來一頓瞎幾把亂秀

且不說這堆一脈相承的智障操作對性能的影響，很顯然，他們對真隨機數、偽隨機數這些概念有點誤會，對計算機生成隨機數的原理也不甚了然^[1]。他們只是看到那一長串變幻無窮的GUID后，心中的虔誠感油然而生。然而回過神，迎面撲來的卻是現實的一盆刺骨冷水：對於 Random，任何多余的初始化都不過是拖后腿的累贅而已。

一頓操作猛如虎，一看戰績0-5

已經足夠好的 Random

一個隨機數發生器好壞的評判標准，首先看它在值域的分布概率是不是符合均勻分布（uniform distribution），也就是說它取得任何一個值的概率都是相同的。其次看它的性能。不管你的Random用到了多么炫酷亮瞎狗眼的神技，只要它的性能不夠，狂吃資源，那它一定是個辣雞。如果產生一個隨機數需要5分鍾，那么任何音樂軟件的“隨機播放”都將變得索然無味。而對於那些熱衷於使用花里胡哨的玩意兒來做隨機數種子的方案，性能永遠不可能超過默認構造函數。畢竟你每次生成隨機數的時候都必須把這幫家伙初始化一遍，否則就和默認初始化完全一樣，這些花里胡哨將變得毫無意義。數學理論不說了，先讓秀兒們看看默認的Random到底夠不夠隨機。

Case #1 的數據，是用默認構造函數初始化生成的數據統計概率直方圖。\(x\) 軸是輸出的隨機數數值，范圍從0-99，共100個數。\(y\) 軸是對應數值出現的概率。每次試驗生成 10,000,000（一千萬）個隨機數，每個case進行10次試驗，共生成100,000,000（一億）個隨機數，得到單次試驗平均用時0.76秒。令 \(\tau=0.76\)，后面的case都按類似的方法進行，並以 \(\tau\) 作為性能基准，排除計算平台的干擾。

Case #1  new Random()，用時：$\tau=0.76$

Loading...

平均值0.010,000,000,000,000,002（移動滑塊縮放顯示窗口）

// 函數定義
static Random r = new Random();
int GetRandomNumber() {
    return r.Next(0, 100);
}
// 主程序（單次試驗，1000 萬隨機數）
 ...
 for (int i = 0; i < 10000000; i++) {
 output[i] = GetRandomNumber();
 }
 ...
 

看，典型的均勻分布。在一千萬大數據量的支撐下，100個可能值每個的輸出概率都達到了完美的1%，帶內波動小於 ±0.000,05，還有什么可挑剔的呢？

那么，現在開始試驗備受推崇的GUID初始化隨機數發生器了。當然，這句話也可以拗口地說成隨機生成隨機數發生器（generate random generator randomly），反正都是秀嘛。

找個比較簡單的GUID例子：

▲ Case #2 算法

依然生成 10,000,000個隨機數，主程序內容不變，只需要修改 GetRandomNumber()：

Case #2  new Random(GUID)，用時：$52\tau$

Loading...

平均值0.009,999,999,999,999,998

int GetRandomNumber() {
    Random r = new Random(
        BitConverter.ToInt32(Guid.NewGuid().ToByteArray(), 0)
    );
    return r.Next(0, 100);
}

沒錯，這確實能得到輸出效果不錯的隨機數。但是，它的帶內波動達到了 ±0.000,2，差不多是 Case #1 的4倍，完全談不上更好。那么性能呢？在效果近似，波動更大的情況下，Case #2 用時達到了 #1 的52倍（39.5秒），這么辣雞的性能還好意思吹您 的 呢？

再來看看更秀的，本文最開始那張圖里的例子，用的是GUID×Time×計數器這種秀破天際的初始化方案：

Case #3  new Random(GUID * Time * count)，用時：$56\tau$

Loading...

平均值0.009,999,999,999,999,998

static randomCount = 0;
static int GetRandomNumber() {
    randomCount++;
    Guid guid = Guid.NewGuid();
    int key1 = guid.GetHashCode();
    int key2 = unchecked((int)DateTime.Now.Ticks);
    int seed = unchecked(key1 * key2 * randomCount);
    Random r = new Random(seed);
    return r.Next(0, 100);
}

您可省省吧！

這段代碼的作者甚至還想到了用unchecked略微優化一下代碼的健壯性。上手就來秀優化，全然不顧丫的壓根兒從算法上就有毛病。習慣成自然，可以猜測他平時在業務工作中沒少這么干。然后是hashcode、time tick各種key一頓花里胡哨得到一個seed來初始化Random。可是這又有什么卵用呢？朋友？為了這個和 Case #1 幾乎一樣效果的輸出結果花掉了 56倍（43秒）的計算時間，您覺得合適嗎？？

代碼的質量不是看它用了多少技巧，秀了多少知識，只要花點功夫，這並不難做到。恰恰相反，用最簡單的辦法實現適當功能和良好的性能，才是最困難的。一段代碼是不是實用，你也不可能靠它的字數來判斷，任何結論，要么理論推導，要么試驗驗證。那些被人奉為經典的半吊子大神的話，可能往往只是他們放的狗屁而已。

The End. \(\Box\)

function paintCanvas(canvas,data,title=''){var chart=echarts.init(document.getElementById(canvas));chart.setOption({title:{text:title},tooltip:{},legend:{data:['']},dataZoom:[{id:'dataZoomX',type:'slider',xAxisIndex:[0],filterMode:'filter'}],xAxis:{type:'category',data:data.x},yAxis:{scale:true},series:[{name:'',type:'bar',data:data.y}],grid:{top:10}})} paintCanvas('canvas-uniform',{x:[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99],y:[0.009999,0.0099791,0.0099851,0.010045,0.0099323,0.0100227,0.009973,0.0099823,0.0099812,0.0100117,0.0099555,0.0100219,0.0099808,0.0099235,0.0100256,0.0100195,0.0100347,0.0099649,0.0099957,0.0100012,0.0099983,0.0100443,0.0100185,0.0100028,0.0099762,0.0099998,0.010038,0.0099765,0.0100109,0.01005,0.01003,0.0100054,0.0099526,0.010001,0.0100172,0.0100067,0.0099779,0.0100547,0.0100519,0.009982,0.0100288,0.0100559,0.0100406,0.0099738,0.009963,0.0099715,0.0099753,0.0099745,0.0099862,0.0099928,0.0100211,0.009992,0.0100048,0.0099731,0.0100005,0.0100157,0.0100208,0.0099976,0.0099595,0.01004,0.0100246,0.0100253,0.0100169,0.0099769,0.0099607,0.0100206,0.010013,0.0099873,0.0099567,0.0099987,0.0099625,0.0100595,0.0099338,0.0100009,0.0100181,0.0099867,0.0100141,0.010015,0.0099953,0.0100089,0.0100287,0.0100257,0.0100045,0.0100001,0.0100012,0.0100357,0.0099458,0.0100448,0.0099926,0.0099496,0.0100401,0.0099849,0.0099666,0.0100041,0.0100289,0.0099873,0.009978,0.0099922,0.0099923,0.0100045]}); paintCanvas('canvas-guid-complex',{x:[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99],y:[0.0099588,0.0099963,0.0099867,0.0100279,0.0099967,0.010009,0.0100268,0.0099682,0.0100487,0.0099889,0.0099977,0.0100182,0.0099335,0.0100526,0.009973,0.0100066,0.0100051,0.0100047,0.0099966,0.0099876,0.0100194,0.0099806,0.0100175,0.0099505,0.0100212,0.0100264,0.009992,0.0100031,0.009983,0.010008,0.0099936,0.0099991,0.0099872,0.0100077,0.0099912,0.010037,0.0099905,0.010021,0.009914,0.0100222,0.0100325,0.0100102,0.0099669,0.0100526,0.0100275,0.0099605,0.009992,0.0100189,0.009977,0.0099784,0.0100042,0.0100493,0.0100071,0.0100028,0.0099897,0.0099899,0.0100025,0.0100122,0.0099531,0.0099807,0.0099904,0.0100486,0.0099704,0.0100065,0.010033,0.0100121,0.0099809,0.0100112,0.0099883,0.0100166,0.0100253,0.0100342,0.0100222,0.0099733,0.010014,0.0100402,0.0100118,0.0099748,0.0099949,0.0099435,0.0100269,0.0100096,0.0100035,0.0099935,0.0100527,0.010029,0.0099903,0.0100189,0.0099248,0.0099613,0.0100053,0.010031,0.0099276,0.0099528,0.0100068,0.0099687,0.0100655,0.0100228,0.0099488,0.0100112]}); paintCanvas('canvas-guid',{x:[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99],y:[0.009969,0.0099814,0.010007,0.0099734,0.0100348,0.0099411,0.010003,0.0099791,0.0100223,0.0100081,0.010043,0.0099691,0.0100782,0.0099857,0.0100318,0.0099852,0.0099031,0.0100022,0.0099955,0.0100051,0.0100461,0.0100305,0.0099732,0.010001,0.0100225,0.0099725,0.0099751,0.0099508,0.0100818,0.0100174,0.0099828,0.0099956,0.0100132,0.0099434,0.0100041,0.0100116,0.0100056,0.0099988,0.0099917,0.0099825,0.0100184,0.0100239,0.0100129,0.010003,0.0099602,0.0099697,0.0100037,0.0100334,0.0099445,0.0099953,0.0100062,0.0100498,0.0099619,0.0100386,0.0100099,0.0099716,0.009962,0.009943,0.010002,0.0100403,0.0100361,0.0099616,0.0100443,0.0100754,0.0099895,0.0100405,0.0100066,0.0100115,0.0099725,0.0100516,0.0099657,0.0099764,0.0100091,0.0100129,0.0100226,0.0099746,0.009995,0.0100378,0.0099761,0.010022,0.0099874,0.0099882,0.0099836,0.0100138,0.0100195,0.0100456,0.0099339,0.0099185,0.0099723,0.0099984,0.0100329,0.010027,0.0099859,0.0100038,0.0099408,0.0100047,0.0100058,0.0100318,0.010004,0.0100517]});

---

1. 所有的真隨機數發生器都需要專用硬件支持，它們中絕大部分受到發明專利保護。System.Random 基於 Donald E. Knuth 的減隨機數生成器算法實現，從實用角度而言，隨機程度已經足夠。 ↩︎