[julia-users] Re: @everywhere and memory allocation

[Andre Bieler]

I think you are looking for shared arrays?
http://docs.julialang.org/en/latest/manual/parallel-computing/#id2



On Tuesday, December 1, 2015 at 2:55:59 PM UTC-5, Pieterjan Robbe wrote:
>
> does the @everywhere macro allocate extra memory to make local copies of a 
> matrix for every processor?
>
> A = sprandn(1,1,0.7)
> @time A = sprandn(1,1,0.7)
>
> gives 2.422259 seconds (23 allocations: 1.565 GB, 3.77% gc time)
>
> @everywhere A = sprandn(1,1,0.7)
> @time @everywhere A = sprandn(1,1,0.7)
>
> gives 16.495639 seconds (1.31 k allocations: 1.565 GB, 6.14% gc time).
>
> However, I know that there are local copies of the matrix on each 
> processor:
>
> @everywhere println(A[1,1])
>
> -1.2751101862102039
>
> From worker 5: 0.0
>
> From worker 4: 0.0
>
> From worker 2: 0.853669869355948
>
> From worker 3: 0.0
>
> Is there a way to use the @everywhere macro without allocating extra 
> memory? Suppose A was created using
>
> A = speye(1,1,0.7)
>
> is there also a copy of the matrix A for all of the workers?
>

[julia-users] Re: @everywhere and memory allocation

[Pieterjan Robbe]

No, not really. SharedArrays do only support bitstypes. I have an application 
where the function to be executed in parallel depends on some fixed data (some 
constants, a dict, some arrays, etc.) I would like to replace my @parallel for 
loop by @everywhere (because of the reuse of my cholesky factorization), but 
I'm worried that this will explode my memory usage.