Avoid memory copy in thread-based environment (parfeval)

Is it possible to avoid memory copy of the input data to each individual worker when using a thread based environment (thus with shared memory)? Here a very simple example on what I'm intended to do:

function compMat()

m = rand(1000000000, 1);

c1 = f1(m);

c2 = f2(m);

fprintf("%f %f\n", c1, c2);

p1 = parfeval(backgroundPool, @f1, 1, m);

p2 = parfeval(backgroundPool, @f2, 1, m);

c1 = fetchOutputs(p1);

c2 = fetchOutputs(p2);

fprintf("%f %f\n", c1, c2);

end

function x = f1(m)

x = mean(sin(m));

end

function x = f2(m)

x = mean(cos(m));

end

The input data is quite large, but the two operations f1 and f2 on this data don't need a copy of the data. Is it possible to reimplement the above example such that all the time there is only one instance of the input vector in memory and the operations f1 and f2 run in parallel on it?

Answer 1

[Edric Ellis]

In this case, thread-based workers already avoid copying the large array m to the workers. See this section in the doc. Note that when you call sin(m) that will create a new temporary array of the same size as m for the output.

Perhaps somewhat counter-intuitively for MATLAB, you might be better off using a for loop (rather than vectorised operations as you have in your example code) to operate on the large array m to avoid making same-sized temporary arrays during the computation.

[Thomas Witkowski]

Thanks, I've completly overseen that sin(m) / cos(m) creates a temporary array. I've check the memory behavior with a for loop working on m. The memory consumption is than as expected in multi threaded environment.