control random number generator in async programming

My problem is very simple. I am using asynchronous programing to generate random number dependent data.

I would like to have reproducible results, but I can't.

For example:

rng(1);
fA = parfeval(backgroundPool,@rand,1,10000);
B = rand(10000);
wait(fA)
C = max(fetchOutputs(fA),B);
B(1:5)
C(1:5)

If I repeat this code multiple times, only array B is reproducible. I want array C to be reproducible as well. Seems a different random number generator is used there. How can I control that generator?

Answer 1

[Walter Roberson]

parpool are defined to seed random number generators differently. See https://www.mathworks.com/help/parallel-computing/control-random-number-streams-on-workers.html

[André A]

I already did what explained there, but doesn't work:

rng(1, 'twister');
fA = parfeval(backgroundPool,@rand,1,10000);
B = rand(1,10000);
wait(fA)
C = max(fetchOutputs(fA),B);
B(1:5)
C(1:5)

I still get different results for C.

Calling rng('default') does not solve the problem.

[Walter Roberson]

backgroundPool = backgroundPool();

F = parfevalOnAll(backgroundPool, @() rng(1,'twister'), 0);
fetchOutputs(F);

fA = parfeval(backgroundPool,@rand,1,1,10000);
rng(1,'twister');
B = rand(1,10000);
wait(fA)
C = max(fetchOutputs(fA),B);
B(1:5)

ans = 1×5

0.4170 0.7203 0.0001 0.3023 0.1468

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

C(1:5)

ans = 1×5

0.4170 0.7203 0.0001 0.3023 0.1468

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

[Walter Roberson]

rng(1,'twister');

That sets the seed for the client process

fA = parfeval(backgroundPool,@rand,1,10000);

Note that command asks to run parfeval() on the backgroundpool, asking to run the function @rand, expecting 1 output, and passing parameter 10000 to rand. rand(10000) is the same as rand(10000,10000)

By default, each pool member is initialized with a different seed, ignoring the current seed for the client process. So you need to set the seed for the workers by using parfevalOnAll.

B = rand(1,10000);

That asks to run rand producing a 1 x 10000 output. not a 10000 x 10000 output.

So you were asking for diffferent sizes of output from rand.

[André A]

I see. Then if I want to specify a different seed for each worker, how do I proceed? parfevalOnAll seems to only allow to specify the same seed for all workers.

[Walter Roberson]

@André A

NumWorkers = backgroundPool.NumWorkers;
random_seeds = [appropriate list];

F = parfevalOnAll(backgroundPool, @() getCurrentTask().ID, 1);
IDs = fetchOutputs(F);
D = dictonary(IDs, random_seeds);
F = parfevalOnAll(backgroundPool, @() rng(D(getCurrentTask.ID), 'twister'), 0);
fetchOutputs(F);

Now you can

fA = parfevalOnAll(backgroundPool,@rand,1,1,10000);
A = cell(NumWorkers,1);
for idx = 1 : NumWorkers
    [~, A{idx}] = fetchNext(fA);
end