Why is the new random library better than std::rand()?

Question

So I saw a talk called rand() Considered Harmful and it advocated for using the engine-distribution paradigm of random number generation over the simple std::rand() plus modulus paradigm.

However, I wanted to see the failings of std::rand() firsthand so I did a quick experiment:

Basically, I wrote 2 functions getRandNum_Old() and getRandNum_New() that generated a random number between 0 and 5 inclusive using std::rand() and std::mt19937+std::uniform_int_distribution respectively.
Then I generated 960,000 (divisible by 6) random numbers using the "old" way and recorded the frequencies of the numbers 0-5. Then I calculated the standard deviation of these frequencies. What I'm looking for is a standard deviation as low as possible since that is what would happen if the distribution were truly uniform.
I ran that simulation 1000 times and recorded the standard deviation for each simulation. I also recorded the time it took in milliseconds.
Afterwards, I did the exact same again but this time generating random numbers the "new" way.
Finally, I calculated the mean and standard deviation of the list of standard deviations for both the old and new way and the mean and standard deviation for the list of times taken for both the old and new way.

Here were the results:

[OLD WAY]
Spread
       mean:  346.554406
    std dev:  110.318361
Time Taken (ms)
       mean:  6.662910
    std dev:  0.366301

[NEW WAY]
Spread
       mean:  350.346792
    std dev:  110.449190
Time Taken (ms)
       mean:  28.053907
    std dev:  0.654964

Surprisingly, the aggregate spread of rolls was the same for both methods. I.e., std::mt19937+std::uniform_int_distribution was not "more uniform" than simple std::rand()+%. Another observation I made was that the new was about 4x slower than the old way. Overall, it seemed like I was paying a huge cost in speed for almost no gain in quality.

Is my experiment flawed in some way? Or is std::rand() really not that bad, and maybe even better?

For reference, here is the code I used in its entirety:

#include <cstdio>
#include <random>
#include <algorithm>
#include <chrono>

int getRandNum_Old() {
    static bool init = false;
    if (!init) {
        std::srand(time(nullptr)); // Seed std::rand
        init = true;
    }

    return std::rand() % 6;
}

int getRandNum_New() {
    static bool init = false;
    static std::random_device rd;
    static std::mt19937 eng;
    static std::uniform_int_distribution<int> dist(0,5);
    if (!init) {
        eng.seed(rd()); // Seed random engine
        init = true;
    }

    return dist(eng);
}

template <typename T>
double mean(T* data, int n) {
    double m = 0;
    std::for_each(data, data+n, [&](T x){ m += x; });
    m /= n;
    return m;
}

template <typename T>
double stdDev(T* data, int n) {
    double m = mean(data, n);
    double sd = 0.0;
    std::for_each(data, data+n, [&](T x){ sd += ((x-m) * (x-m)); });
    sd /= n;
    sd = sqrt(sd);
    return sd;
}

int main() {
    const int N = 960000; // Number of trials
    const int M = 1000;   // Number of simulations
    const int D = 6;      // Num sides on die

    /* Do the things the "old" way (blech) */

    int freqList_Old[D];
    double stdDevList_Old[M];
    double timeTakenList_Old[M];

    for (int j = 0; j < M; j++) {
        auto start = std::chrono::high_resolution_clock::now();
        std::fill_n(freqList_Old, D, 0);
        for (int i = 0; i < N; i++) {
            int roll = getRandNum_Old();
            freqList_Old[roll] += 1;
        }
        stdDevList_Old[j] = stdDev(freqList_Old, D);
        auto end = std::chrono::high_resolution_clock::now();
        auto dur = std::chrono::duration_cast<std::chrono::microseconds>(end-start);
        double timeTaken = dur.count() / 1000.0;
        timeTakenList_Old[j] = timeTaken;
    }

    /* Do the things the cool new way! */

    int freqList_New[D];
    double stdDevList_New[M];
    double timeTakenList_New[M];

    for (int j = 0; j < M; j++) {
        auto start = std::chrono::high_resolution_clock::now();
        std::fill_n(freqList_New, D, 0);
        for (int i = 0; i < N; i++) {
            int roll = getRandNum_New();
            freqList_New[roll] += 1;
        }
        stdDevList_New[j] = stdDev(freqList_New, D);
        auto end = std::chrono::high_resolution_clock::now();
        auto dur = std::chrono::duration_cast<std::chrono::microseconds>(end-start);
        double timeTaken = dur.count() / 1000.0;
        timeTakenList_New[j] = timeTaken;
    }

    /* Display Results */

    printf("[OLD WAY]\n");
    printf("Spread\n");
    printf("       mean:  %.6f\n", mean(stdDevList_Old, M));
    printf("    std dev:  %.6f\n", stdDev(stdDevList_Old, M));
    printf("Time Taken (ms)\n");
    printf("       mean:  %.6f\n", mean(timeTakenList_Old, M));
    printf("    std dev:  %.6f\n", stdDev(timeTakenList_Old, M));
    printf("\n");
    printf("[NEW WAY]\n");
    printf("Spread\n");
    printf("       mean:  %.6f\n", mean(stdDevList_New, M));
    printf("    std dev:  %.6f\n", stdDev(stdDevList_New, M));
    printf("Time Taken (ms)\n");
    printf("       mean:  %.6f\n", mean(timeTakenList_New, M));
    printf("    std dev:  %.6f\n", stdDev(timeTakenList_New, M));
}

That is pretty much why this advice exists. If you don't know how to test the RNG for sufficient entropy or whether or not it matters for your program then you should assume that std::rand() isn't good enough. en.wikipedia.org/wiki/Entropy_(computing) — Hans Passant
The bottom line on whether rand() is good-enough depends largely on what you are using the collection of random numbers for. It you need a particular type of random distribution, then, of course the library implementation will be better. If you simply need random numbers and don't care about the "randomness" or what type of distribution is produced, then rand() is fine. Match the proper tool to the job at hand. — David C. Rankin
possible dupe: stackoverflow.com/questions/52869166/… I just don't want to hammer this one, so I refrain from actually voting. — bolov
for (i=0; i<k*n; i++) a[i]=i%n; produces same exact mean and standard deviation as the best RNG out there. If this is good enough for your application, just use this sequence. — n. 1.8e9-where's-my-share m.
"standard deviation as low as possible" - no. That's wrong. You expect the frequencies to be a little bit different - about sqrt(frequency) is about what you expect the standard deviation to be. The "incrementing counter" that n.m. produced will have a much lower s.d. (and is a very bad rng). — Martin Bonner supports Monica

Matteo Italia Matteo Italia · Accepted Answer · 2018-10-29T08:49:15

Pretty much any implementation of "old" rand() use an LCG; while they are generally not the best generators around, usually you are not going to see them fail on such a basic test - mean and standard deviation is generally got right even by the worst PRNGs.

Common failings of "bad" - but common enough - rand() implementations are:

low randomness of low-order bits;
short period;
low RAND_MAX;
some correlation between successive extractions (in general, LCGs produce numbers that are on a limited number of hyperplanes, although this can be somehow mitigated).

Still, none of these are specific to the API of rand(). A particular implementation could place a xorshift-family generator behind srand/rand and, algoritmically speaking, obtain a state of the art PRNG with no changes of interface, so no test like the one you did would show any weakness in the output.

Edit: @R. correctly notes that the rand/srand interface is limited by the fact that srand takes an unsigned int, so any generator an implementation may put behind them is intrinsically limited to UINT_MAX possible starting seeds (and thus generated sequences). This is true indeed, although the API could be trivially extended to make srand take an unsigned long long, or adding a separate srand(unsigned char *, size_t) overload.

Indeed, the actual problem with rand() is not much of implementation in principle but:

backwards compatibility; many current implementations use suboptimal generators, typically with badly chosen parameters; a notorious example is Visual C++, which sports a RAND_MAX of just 32767. However, this cannot be changed easily, as it would break compatibility with the past - people using srand with a fixed seed for reproducible simulations wouldn't be too happy (indeed, IIRC the aforementioned implementation goes back to Microsoft C early versions - or even Lattice C - from the mid-eighties);
simplistic interface; rand() provides a single generator with the global state for the whole program. While this is perfectly fine (and actually quite handy) for many simple use cases, it poses problems:
- with multithreaded code: to fix it you either need a global mutex - which would slow down everything for no reason and kill any chance of repeatability, as the sequence of calls becomes random itself -, or thread-local state; this last one has been adopted by several implementations (notably Visual C++);
- if you want a "private", reproducible sequence into a specific module of your program that doesn't impact the global state.

Finally, the rand state of affairs:

doesn't specify an actual implementation (the C standard provides just a sample implementation), so any program that is intended to produce reproducible output (or expect a PRNG of some known quality) across different compilers must roll its own generator;
doesn't provide any cross-platform method to obtain a decent seed (time(NULL) is not, as it isn't granular enough, and often - think embedded devices with no RTC - not even random enough).

Hence the new <random> header, which tries to fix this mess providing algorithms that are:

fully specified (so you can have cross-compiler reproducible output and guaranteed characteristics - say, range of the generator);
generally of state-of-the-art quality (from when the library was designed; see below);
encapsulated in classes (so no global state is forced upon you, which avoids completely threading and nonlocality problems);

... and a default random_device as well to seed them.

Now, if you ask me I would have liked also a simple API built on top of this for the "easy", "guess a number" cases (similar to how Python does provide the "complicated" API, but also the trivial random.randint & Co. using a global, pre-seeded PRNG for us uncomplicated people who'd like not to drown in random devices/engines/adapters/whatever every time we want to extract a number for the bingo cards), but it's true that you can easily build it by yourself over the current facilities (while building the "full" API over a simplistic one wouldn't be possible).

Finally, to get back to your performance comparison: as others have specified, you are comparing a fast LCG with a slower (but generally considered better quality) Mersenne Twister; if you are ok with the quality of an LCG, you can use std::minstd_rand instead of std::mt19937.

Indeed, after tweaking your function to use std::minstd_rand and avoid useless static variables for initialization

int getRandNum_New() {
    static std::minstd_rand eng{std::random_device{}()};
    static std::uniform_int_distribution<int> dist{0, 5};
    return dist(eng);
}

I get 9 ms (old) vs 21 ms (new); finally, if I get rid of dist (which, compared to the classic modulo operator, handles the distribution skew for output range not multiple of the input range) and get back to what you are doing in getRandNum_Old()

int getRandNum_New() {
    static std::minstd_rand eng{std::random_device{}()};
    return eng() % 6;
}

I get it down to 6 ms (so, 30% faster), probably because, unlike the call to rand(), std::minstd_rand is easier to inline.

Incidentally, I did the same test using a hand-rolled (but pretty much conforming to the standard library interface) XorShift64*, and it's 2.3 times faster than rand() (3.68 ms vs 8.61 ms); given that, unlike the Mersenne Twister and the various provided LCGs, it passes the current randomness test suites with flying colors and it's blazingly fast, it makes you wonder why it isn't included in the standard library yet.

Why is the new random library better than std::rand()?

4 Answers