CODEWITHSUNDEEP
arraysjuliacalculation
universe
universe

Reputation: 13

Distributed Arrays initialization in Julia

I'm looking for help with julia programming. I'm a newbie in programming, and I know little about the structure of computer and programming, so please pardon me if I'm asking some stupid questions.

I have to do heavy calculation, so I wanna use parallel computing for the double for statement. The code is following:

using Distributed
@everywhere using DistributedArrays
addprocs(4)

function FreeSpace_2D(profile, x_prime, y_prime, d, x, y) 
   # profile is a 2D array, and x_prime and y_prime are 1D arrays. d, x, y is real numbers.
   Nx = length(x_prime)
   Ny = length(y_prime)
   array = dzeros((Nx, Ny), workers()[1:4], [1, 4])
   @distributed for i in 1:Nx
      @distributed for j in 1:Ny
         localpart(array)[i, j] = ( profile[i, j]*exp(-1im*0.5*k/d*((x-x_prime[i])^2+(y-y_prime[j])^2)) )
      end
   end

   return array 
end

but the code does not initialize the 'array'.

When I search internet, there are a few methods for initialization of 1D array using distributed, but they did not work on a 2D array.

I would appreciate for any help.

Upvotes: 1

Views: 969

Answers (1)

Przemyslaw Szufel
Przemyslaw Szufel

Reputation: 42194

There are four problems with your code:

  • you need first to addprocs and then call @everywhere. Otherwise your @everywhere will be called only on the master process
  • you need to synchronize @distributed loops by putting @sync macro before @distributed or provide an aggregator function that is: @distributed (some_func) for i in 1:1000
  • @distributed loop distributes code across all available workers. Hence it does not make sense to include one @distributed loop within another. I recommend putting @distributed only in front of the external loop
  • you need to use a local part of your array on each worker node

Here is a sample code that I tried to make similar to your goals and that works correctly:

using Distributed
addprocs(4)

@everywhere using Distributed
@everywhere using DistributedArrays

a = dzeros(Int, (4, 8), workers()[1:4], [1, 4])

@sync @distributed for j in 1:8
    a_local = localpart(a)
    for i in 1:4
        a_local[i, ((j+1) % 2)+1 ] = 100j + 10i + myid() 
    end
end

Looking at the Array a reveals what has happened:

julia> a
4×8 DArray{Int64,2,Array{Int64,2}}:
 112  212  313  413  514  614  715  815
 122  222  323  423  524  624  725  825
 132  232  333  433  534  634  735  835
 142  242  343  443  544  644  745  845

Upvotes: 2

Related Questions