Random Sampler behavior

General Usage
1

Hello. I’m now rewriting some Monte Carlo simulation code from Java to Julia. It probably was a bad idea to implement it in Java in the first place, but now I use results calculated by Java implementation to verify the correctness of my Julia implementation. Classic Mersenne Twister (MT19937) is used in both variants. And I found a difference in results, let me show it.

Java code:

public static void main(String[] args) {
    MersenneTwister twister = new MersenneTwister(0);

    for (int i = 1; i <= 5; i++) {
        System.out.println( twister.random());
    }

    for (int i = 1; i <= 10; i++) {
        System.out.print((twister.chooseFrom(10) + 1) + " ");
    }
}

Julia code:

using Random
using RandomNumbers.MersenneTwisters

RNG = MT19937(0)
for i in 1:5 println(rand(RNG)) end

s = Random.Sampler(RNG, 1:10)
for i in 1:10 print(rand(RNG, s), " ") end

Java output:
0.548813502304256
0.5928446163889021
0.715189364971593
0.8442657440900803
0.6027633703779429
9 6 9 5 7 7 4 5 3 9

Julia output:
0.548813502304256
0.5928446163889021
0.715189364971593
0.8442657440900803
0.6027633703779429
6 5 7 5 9 10 4 8 6 6

So, as the first 5 lines of outputs prove, both implementations start from the same pseudo-random sequence. But Random.Sampler in Julia provides only even-numbered elements of the sequence:

9 6 9 5 7 7 4 5 3 9 
  6   5   7   5   9 10 4 8 6 6

So it looks like the Sampler makes extra calls to RNG. Is it so? Why this is needed? Does this affect the performance? Is there a way to get rid of that?

Thanks in advance.

2

this is because Java defaults to Int32:

julia> RNG = MT19937(0)
MT19937(UInt32[0x00000000, 0x00000001, 0x6c078967, 0x714acb41, 0x48077044, 0x93e7383e, 0xf0f54bb2, 0x141f95dc, 0xc0acdbd4, 0x3cb9cadc  …  0xad4200bc, 0xdaa1fb5d, 0x78ff7c7e, 0x71e51784, 0x499b76e3, 0xd522db95, 0x075dea9a, 0x8ccffb2f, 0xb5b4302f, 0x6b1a1930], 625)

julia> for i in 1:5 println(rand(RNG)) end
0.548813502304256
0.5928446163889021
0.715189364971593
0.8442657440900803
0.6027633703779429

julia> for i in 1:10 print(rand(RNG, Int32(1):Int32(10)), " ") end
9 6 9 5 7 7 4 5 3 9

so you want to change to:

s = Random.Sampler(RNG, Int32(1):Int32(10))
4 Likes
3

Thank you very much. Do I understand correctly that in my code the Sampler generates 64 random bits at each call but discards half of the result then? If so, it probably means that for the 1:10 range the type could be restricted even further (Int8). Do such things make any significant performance gain?

4

if you care about performance, don’t use MT to begin with, as the document would tell you: Mersenne Twisters - RandomNumbers.jl

5

I won’t. I just need it for the debugging time, for the reproducibility of the results. I’ll change it to Xoroshiro after that. Thanks again.