Stream: helpdesk (published)

Topic: ✔ `Xoshiro` vs. `MersenneTwister`

Júlio Hoffimann (Jul 04 2025 at 15:21):

Do you have a general rule of thumb for picking one over the other?

Benchmarks are inconsistent depending on how many random numbers are generated:

julia> rng1 = MersenneTwister(123)
MersenneTwister(123)

julia> rng2 = Xoshiro(123)
Xoshiro(0xfefa8d41b8f5dca5, 0xf80cc98e147960c1, 0x20e2ccc17662fc1d, 0xea7a7dcb2e787c01, 0xf4e85a418b9c4f80)

julia> @btime rand($rng1, 1);
  27.570 ns (2 allocations: 64 bytes)

julia> @btime rand($rng2, 1);
  19.695 ns (2 allocations: 64 bytes)

julia> @btime rand($rng1, 100);
  118.183 ns (2 allocations: 928 bytes)

julia> @btime rand($rng2, 100);
  133.110 ns (2 allocations: 928 bytes)

Mason Protter (Jul 04 2025 at 15:29):

I just always use Xoshiro. MT is kinda just for backwards compatibility

Júlio Hoffimann (Jul 04 2025 at 15:31):

That is what I was told. I am curious why it is slower in the n=100 case above.

Mason Protter (Jul 04 2025 at 15:32):

FWIW, I don't see Xoshiro lose at any size on my machine:

julia> using Random

julia> rng1 = MersenneTwister(123)
MersenneTwister(123)

julia> rng2 = Xoshiro(123)
Xoshiro(0xfefa8d41b8f5dca5, 0xf80cc98e147960c1, 0x20e2ccc17662fc1d, 0xea7a7dcb2e787c01, 0xf4e85a418b9c4f80)

julia> @btime rand($rng1, 1);
  20.245 ns (2 allocations: 64 bytes)

julia> @btime rand($rng2, 1);
  15.435 ns (2 allocations: 64 bytes)

julia> @btime rand($rng1, 100);
  71.995 ns (2 allocations: 928 bytes)

julia> @btime rand($rng2, 100);
  66.703 ns (2 allocations: 928 bytes)

julia> @btime rand($rng1, 10000);
  3.859 μs (3 allocations: 78.20 KiB)

julia> @btime rand($rng2, 10000);
  2.639 μs (3 allocations: 78.20 KiB)

Mason Protter (Jul 04 2025 at 15:33):

I wouldn't be surprised if this sort of thing was pretty architecture dependent.

Júlio Hoffimann (Jul 04 2025 at 15:35):

Interesting. I will move to Xoshiro in that case.

Notification Bot (Jul 04 2025 at 15:35):

Júlio Hoffimann has marked this topic as resolved.

Jakob Nybo Nissen (Jul 04 2025 at 16:26):

IIRC, MT buffers a bunch of random numbers which xoshiro doesn't. So when you time with @btime, which gives the minimum time, you don't benchmark the actual time to generate the numbers, only that to copy it from the buffer

Júlio Hoffimann (Jul 04 2025 at 16:32):

Thanks for sharing.

Brenhin Keller (Jul 04 2025 at 22:07):

There's a whole funny rivalry between the Xoshiro and PCG people, but Mersenne is old and either is probably an upgrade over that IMHO

Last updated: Jul 22 2025 at 04:56 UTC