BoundsError when using DistributedArrays

user_14548917 0
Viewed 47

Apparently I must have a fundamental misunderstanding about DistributedArrays.jl. I have set up a MWE of something similar to what I have to do:

using Distributed
using DistributedArrays

addprocs()
@everywhere using Distributed, DistributedArrays

a = distribute(zeros(5))

@sync @distributed for i in 1:5
    a_l = localpart(a)
    a_l[i] = 100 * i
end

And then I run into the following Error:

ERROR: TaskFailedException:
On worker 2:
BoundsError: attempt to access 1-element Array{Float64,1} at index [2]
setindex! at ./array.jl:847
macro expansion at /home/user/test.jl:36 [inlined]
[inlined]
#17 at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.5/Distributed/src/macros.jl:301
#160 at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.5/Distributed/src/macros.jl:87
#103 at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.5/Distributed/src/process_messages.jl:290
run_work_thunk at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.5/Distributed/src/process_messages.jl:79
run_work_thunk at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.5/Distributed/src/process_messages.jl:88
#96 at ./task.jl:356

...and 3 more exception(s).

Stacktrace:
 [1] sync_end(::Channel{Any}) at ./task.jl:314
 [2] (::Distributed.var"#159#161"{var"#17#18",UnitRange{Int64}})() at ./task.jl:333
Stacktrace:

sync_end(::Channel{Any}) at ./task.jl

top-level scope at task.jl

Using a = dzeros((5,1), workers()) also gives the same Error. Any help is appreciated!

julia multiprocessing
1 Answers

There are two problems:

  • localpart is indexed starting from 1
  • the number of your workers is greater than loop size. Since workers are randomly assigned your loop elements get empty localparts.

Let us consider this code:

a = distribute(zeros(5));

@sync @distributed for i in 1:5
    for j in keys(a[:L])
        a[:L][j] = 100 * i+myid()
    end
end

While it solves the first issue the second is still there:

julia> a
5-element DArray{Float64, 1, Vector{Float64}}:
 402.0
 503.0
   0.0
   0.0
   0.0

Why it does not work as expected? because addprocs is adding all processes so I have now 8 workers and the size of loop is 5.

Perhaps the simplest solution is to replace the range from 1:5 to 1:max(5,nworkers()). This makes sure that each localpart is going to get processed.

julia> @sync @distributed for i in 1:max(5,nworkers())
           @show i, myid(), length(a[:L])
           for j in keys(a[:L])
              a[:L][j] = 100 * i+myid()
           end
       end
      From worker 9:    (i, myid(), length(a[:L])) = (6, 9, 0)
      From worker 7:    (i, myid(), length(a[:L])) = (4, 7, 0)
      From worker 2:    (i, myid(), length(a[:L])) = (7, 2, 1)
      From worker 8:    (i, myid(), length(a[:L])) = (5, 8, 0)
      From worker 3:    (i, myid(), length(a[:L])) = (8, 3, 1)
      From worker 4:    (i, myid(), length(a[:L])) = (1, 4, 1)
      From worker 5:    (i, myid(), length(a[:L])) = (2, 5, 1)
      From worker 6:    (i, myid(), length(a[:L])) = (3, 6, 1)
Task (done) @0x0000000073e09f50

This code run shows clearly what is happening when you loop over 5 elements and use 8 workers.

The result is now as expected (with regard that tasks are randomly allocated around workers):

julia> a
5-element DArray{Float64, 1, Vector{Float64}}:
 702.0
 803.0
 104.0
 205.0
 306.0
user_99577100
Related