reduction_func and :threads error

[ivborissov]

Hi,

the issue is also mentioned here, but still present in Julia 0.7:
https://github.com/JuliaDiffEq/DiffEqMonteCarlo.jl/issues/23

It can be illustrated by the following example.

using DiffEqBase, OrdinaryDiffEq, DiffEqMonteCarlo, DiffEqCallbacks, Distributions

function ode_sys(du,u,p,t)
  v1 = p[1]*u[1]*p[2]
  du[1] = -v1/p[2]
  du[2] = v1/p[2]
end

prob = ODEProblem(ode_sys, [10.,0.], (0.,100.),  [1.5,1.0])

function prob_func(prob, i, repeat)
  @. prob.u0 = rand(Normal(prob.u0, 0.3))
  @. prob.p = rand(Normal(prob.p, 0.3))
  prob
end

function reduction(u,data,I)
  # do smth, calculate OnlineStats
  append!(u,data),false
end

monte_prob = MonteCarloProblem(
  prob,
  prob_func=prob_func,
  reduction=reduction
)

sol = solve(
  monte_prob,
  Tsit5(),
  batch_size=1,
  parallel_type=:threads,
  num_monte = 100
)

It works fine with parallel_type=:none but fails with parallel_type=:threads with

UnefRefError: access to undefined reference.
Error thrown in threaded loop on thread 0: BoundsError(a=Array{Any, (1,)}[#<null>], i=(2,))

Is this issue related to my code or general julia multithreading?

[ivborissov]

I guess it is the same issue
SciML/DifferentialEquations.jl#241

[ChrisRackauckas]

No, that issue is only with batch sizes > 1. This, I have no idea why this is coming up and will need to dig in.

[ChrisRackauckas]

Is this required now? I was planning on looking at this library again after PARTR merges JuliaLang/julia#22631 since that'll be a big change to the threading constructs.

[ivborissov]

Thanks! If it is not 2.0 but some upcoming 1.x merges, it is reasonable to wait. I'm really looking forward to this batching functionality as it will allow me to fully test OnlineStats with Parallel DiffEqMonteCarlo:
https://github.com/JuliaDiffEq/DiffEqMonteCarlo.jl/issues/23

[ChrisRackauckas]

It was supposed to be 1.1 but it missed that mark. Can't be too far off