Skip to content

Carefully handle splitting of RNG seeds #110

Open
@penelopeysm

Description

@penelopeysm

There are several instances in the codebase where we have to create N pseudo-random number generators from one, for example:

"""
update_keys!(pc::ParticleContainer)
Create new unique keys for the particles in the ParticleContainer
"""
function update_keys!(pc::ParticleContainer, ref::Union{Particle,Nothing}=nothing)
# Update keys to new particle ids
nparticles = length(pc)
n = ref === nothing ? nparticles : nparticles - 1
for i in 1:n
pi = pc.vals[i]
k = split(state(pi.rng.rng))
Random.seed!(pi.rng, k[1])
end
return nothing
end

Right now, this is handled by split, which generates new seeds to be used with new PRNGs.

AdvancedPS.jl/src/rng.jl

Lines 33 to 42 in d0d180f

"""
split(key::Integer, n::Integer=1)
Split `key` into `n` new keys
"""
function split(key::Integer, n::Integer=1)
T = typeof(key)
inner_rng = Random.MersenneTwister(key)
return rand(inner_rng, T, n)
end

This is not always a reliable method as the new seeds may end up accidentally generating highly correlated sequences of numbers. See, e.g., https://gee.cs.oswego.edu/dl/papers/oopsla14.pdf

We should therefore switch to an underlying RNG type that supports splitting. (Note that right now, we are using Philox2x [paper, library] but the library does not implement a split function, even though it should theoretically be possible. Compare e.g. with Numpy's implementation, which provides a jumped method for this purpose.)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions