Parallel pool constant with a parfor and for loop

Hi,

as far as I understood the usage of parallel pool constants, it really helps when a parfor loop is called several times, like in the example provided in the Help

data = rand(1000);
c = parallel.pool.Constant(data);
for ii = 1:10
    % Run multiple PARFOR loops accessing the data.
    parfor jj = 1:10
        x(ii,jj) = c.Value(ii,jj);
    end
end

However, what if I have the opposite, meaning, I have an external parfor, and the inner loop is a for loop. For example, I have something like this

parfor m = 1 : M
    for n = 1 : N
        matrix(m,n) = w1(n);
    end
end

Matlab complains that w1 is a broadcast variable.

Would it make sense if I would convert w1 to a parallel.pool.Constant?

Answer 1

[Edric Ellis]

Yes, parallel.pool.Constant is generally most useful when there are multiple parfor loops. (There are other use-cases such as when you need to construct some data directly on the workers)

In your specific case, if you have only a single parfor loop as you show, then there is no benefit to using parallel.pool.Constant for w1, and you can safely ignore the code analyzer message.

[Maria]

when you mention "other use-cases", could the construction of a matrix be such a case?

The matrix I am filling in is quite big. It is actually a 3D matrix that easily occupies 15 GB. As I am using parfor, I cannot use distributed array. I am running some tests and I see that only the first node in my cluster gets its memory full (16 GB total, so I am at the limit), while the other 7 don't.

I am wondering, would it help if I would declare the matrix as parallel constant?

[Edric Ellis]

Yes, building a large matrix directly on the workers can be a reason to use parallel.pool.Constant as it allows you to avoid transferring it from the client. So, you might do:

c = parallel.pool.Constant(@() ones(10000)) % calls ones(10000) directly on the worker

[Maria]

okay that's good to know! Just another question. When I send a batch job, and the matrix is created inside the function that is called from the batch, the matrix should be automatically created on one of the worker (if I do not use distributed), correct? Or also in this case I should use a parallel constant?

[Edric Ellis]

The function that you evaluate using batch is run on worker 1. If you use the 'Pool' parameter to batch, then parallel language constructs such as parfor run on workers 2:N. So yes - a matrix created by your batch function will be created on worker 1.

[Maria]

very clear, thanks!