issue with pmap: document?

[cecileane]

This issue seems related to #12, but here is an example. The NLopt object is part of a type, to be reused multiple times. All is well, unless we want to distributed the work across processors.

using NLopt

struct Foo
  data::Vector{Float64}
  guess::Vector{Float64}
  opt::NLopt.Opt
end
function Foo(data, frel,fabs,maxeval)
  guess = similar(data)
  opt = NLopt.Opt(:LD_SLSQP, 1)
  NLopt.ftol_rel!(opt,frel)
  NLopt.ftol_abs!(opt,fabs)
  NLopt.maxeval!(opt, maxeval)
  return Foo(data,guess,opt)
end

foo = Foo([4.0,1.0,-5.0], .01,.01,5)

function optimizeone!(obj::Foo,i)
    function f(x, g)
        d = (x[1]-obj.data[i])
        objective = d*d
        if length(g) > 0
            g[1] = 2d
        end
        return objective
    end
    opt = foo.opt
    min_objective!(opt, f)
    optf, optx, ret = NLopt.optimize(opt, [obj.guess[i]])
    obj.guess[i] = optx[1]
    return obj.guess[i]
end

for i in 1:3
  optimizeone!(foo,i)
end
foo.guess ≈ foo.data # true

so all is well here. The problem appears when we use multiple processors:

foo = Foo([4.0,1.0,-5.0], .01,.01,5)
using Distributed
addprocs(1) # 1 worker
@everywhere using NLopt
bar = Distributed.pmap(1:3) do i
  optimizeone!(foo,i)
end

I get an error starting with:

ERROR: On worker 2:
UndefVarError: #optimizeone! not defined
deserialize_datatype at /Users/julia/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.4/Serialization/src/Serialization.jl:1211

As said in #12, I can serialize the optimization parameters and create the Opt object by the worker. But still, it would be a nice feature to have, and if not, it would be nice to be documented.

[odow]

The error is not related to NLopt. Here is the relevant section of the Julia manual: https://docs.julialang.org/en/v1/manual/parallel-computing/#code-availability-1

To avoid issues with serializing the NLopt.Opt, you probably want to construct it on each process.

@everywhere function optimizeone!(i)
    foo = Foo([4.0,1.0,-5.0], .01,.01,5)
    # ...
    return obj.guess
end
pmap(optimizeone!, 1:3)

[cecileane]

Oh yes duh!! I am sorry, I read the error too fast. (it's an attempt to make a small example. In real life I'm running into this problem with things inside a module). So when I start julia with julia -p 1, say, and do @everwhere in front of almost everything above, the error is like this:

      From worker 2:	signal (11): Segmentation fault: 11
      From worker 2:	in expression starting at none:0
      From worker 2:	nlopt_get_errmsg at /Users/ane/.julia/artifacts/a98e645e32bc5ee2510d3fbf404e13941a7e2aa2/lib/libnlopt.0.10.0.dylib (unknown line)
      From worker 2:	errmsg at /Users/ane/.julia/packages/NLopt/KTzQh/src/NLopt.jl:198 [inlined]
      From worker 2:	_errmsg at /Users/ane/.julia/packages/NLopt/KTzQh/src/NLopt.jl:203
      From worker 2:	chk at /Users/ane/.julia/packages/NLopt/KTzQh/src/NLopt.jl:211
      From worker 2:	min_objective! at /Users/ane/.julia/packages/NLopt/KTzQh/src/NLopt.jl:406 [inlined]
      From worker 2:	optimizeone! at ./REPL[5]:11
      From worker 2:	#4 at ./REPL[9]:2
      From worker 2:	unknown function (ip: 0x109cc37a7)
      ...

To be replicate this:

@everywhere using NLopt

@everywhere struct Foo
  data::Vector{Float64}
  guess::Vector{Float64}
  opt::NLopt.Opt
end
@everywhere function Foo(data, frel,fabs,maxeval)
  guess = similar(data)
  opt = NLopt.Opt(:LD_SLSQP, 1)
  NLopt.ftol_rel!(opt,frel)
  NLopt.ftol_abs!(opt,fabs)
  NLopt.maxeval!(opt, maxeval)
  return Foo(data,guess,opt)
end
@everywhere function optimizeone!(obj::Foo,i)
    function f(x, g)
        d = (x[1]-obj.data[i])
        objective = d*d
        if length(g) > 0
            g[1] = 2d
        end
        return objective
    end
    opt = foo.opt
    min_objective!(opt, f)
    optf, optx, ret = NLopt.optimize(opt, [obj.guess[i]])
    obj.guess[i] = optx[1]
    return obj.guess[i]
end

foo = Foo([4.0,1.0,-5.0], .01,.01,5) # or @everywhere, same error
bar = Distributed.pmap(1:3) do i
  optimizeone!(foo,i)
end

(@v1.4) pkg> status
Status `~/.julia/environments/v1.4/Project.toml`
  [76087f3c] NLopt v0.6.0

[odow]

You need to follow the second half of my previous comment.

[cecileane]

oh yes, that's what I already did in my package. Documentation about the problem would have been helpful (and would have saved me time to diagnose the issue).

[odow]

You can make a PR to improve the documentation by clicking on the pencil Icon at the top of the README on the homepage:
https://github.com/JuliaOpt/NLopt.jl/edit/master/README.md

[cecileane]

yes! will do when I get the time.

[odow]

Closing as a duplicate of #186.