Speed comparison & benchmark of differents C++ random number generators (PRNG)

This repository provides an easy way to compare some common pseudo random number generators (PRNGS) used in C++. It also provides my results running the benchmarks for those wanting a quick answer. To reproduce my results and test for yourself, see section Test for yourself. As of now, the tested PRNG are: Mersenne twister, Linear Congruential Engine, xoshiro / xorshift*, PCG, rand(), Substract with carry and some other more exotic engines.

⚠️	Most of these algorithms should not be used in a context where security is required.

Results

All these results were compiled over 100 hundred runs (--benchmark_repetitions=100) using an Intel i5-3570 @ 3.40GHz…​.. The binary was compiled with clang using -O3 -DNDEBUG, without any architecture optimization. The code is written to avoid any clever trick from the compiler that could bias the results.

Table 1. Table of results

PRNG (link)	Mean time per generation (user)	Mean generation per second	Remarks
xoshiro128++	1.06ns ± 0.002 ns	946.757M/s ± 0.719M/s
xoshiro256++	1.06ns ± 0.002 ns	946.704M/s ± 0.581M/s	Fastest for 64 bits, good quality, 400 lines
pcg_cpp::pcg32	1.09ns ± 0.010ns	916.834M/s ± 8.424M/s
pcg_cpp::pcg32_fast	0.933ns ± 0.006ns	1071.970M/s ± 6.220M/s	Fastest for 32 bits
pcg_cpp::pcg64_oneseq	1.09ns ± 0.010ns	916.063M/s ± 8.155M/s
pcg_cpp::pcg64	1.98ns ± 0.013ns	504.325M/s ± 3.121M/s
pcg_cpp::pcg64_fast	1.31ns ± 0.012ns	763.500M/s ± 6.629M/s
pcg_cpp::pcg64_oneseq	1.98ns ± 0.008ns	504.750M/s ± 2.064M/s
wjakob pcg32	1.06ns ± 0.012ns	940.393M/s ± 10.598M/s	Not that much faster than pcg-cpp
wjakob pcg32_8	8.48ns ± 0.112ns	944.200M/s ± 12.215M/s	Fastest if your CPU have AVX, mine does not
burtle burtle C random	2.79ns ± 0.023ns	358.876M/s ± 2.627M/s	20 lines implementation
C rand()	15.4ns ± 0.123ns	65.005M/s ± 0.509M/s	Slow and poor quality, avoid
C++ STL std::mt19937 (MT)	3.70ns ± 0.048ns	270.652M ± 3.418M/s	Best STL generator for portability
C++ STL std::mt19937_64 (MT)	3.40ns ± 0.028ns	294.261M ± 2.352M/s	Best STL generator for x86_64
C++ STL std::minstd_rand (LCG)	3.72ns ± 0.038ns	269.078M ± 2.705M/s	Worse quality than mt19937 and slower
C++ STL std::minstd_rand0 (LCG)	3.72ns ± 0.033ns	268.768M ± 2.336M/s	Worse quality than mt19937 and slower
C++ STL std::ranlux24_base (SWC)	3.27ns ± 0.022ns	306.088M ± 2.013M/s	Poor quality
C++ STL std::ranlux48_base (SWC)	3.28ns ± 0.038ns	305.131M ± 3.484/s	Poor quality
C++ STL std::ranlux24 (DBE)	25.8ns ± 0.243ns	38.809M ± 0.356M/s	Awfully slow
C++ STL std::ranlux48 (DBE)	98.3ns ± 0.926ns	10.18M ± 0.093M/s	Awfully slow
C++ STL std::knuth_b (SOE)	14.9ns ± 0.153ns	66.990M ± 0.660M/s	Better quality than minstd_rand but as slow as rand

Conclusion

std::mt19937 and std::mt19937_64 are the best standard C pseudo random number generators in the STL. However, https://www.pcg-random.org/[PCG] is superior in all aspects **for non-cryptographic** uses. https://prng.di.unimi.it/[xoshiro256] is another very good choice for generating 64 bits random numbers, but it does not support multiple streams like PCG.

rand() need to be avoided at all costs, it is very slow and its quality is bad.

Test for yourself

CMake will automatically download all dependencies from other repositories. You just need to have CMake and C++ compiler installed.

1. Clone this repository using git

   git clone https://github.com/zeFresk/cpp-random-benchmarks.git

2. Compile using CMake :

   cmake .
   cmake --build .

3. Run and publish your results

   ./random_benchmarks