issue with Repeating Random Numbers in parfor-Loops

I want to get reproducible random number sequence in parfor-loops and tried the example in the documentation. However, the results are switching from two choices, which conflicts with the statement in the documentation.

what is the problem?

Here is the pilot code:

n = 3;
r = zeros(1,n);
parfor i = 1:n
  stream = RandStream.getGlobalStream();
  stream.Substream = i;
  r(i) = rand;
end
r

one result is

r =

0.2646    0.2537    0.2276

another is

r =

0.2646    0.2537    0.3111

BTW, the results are produced by 2015a.

Creating random number stream explicitly can solve this problem, but it causes a significant increase in computing time when there are many parfor-loops. Is there a solution which is not expensive concerning computing time?

Answer 1

[Edric Ellis]

I think you missed out a crucial piece of the example. In your code, you're accessing the current global stream without having explicitly set it first. You need the section:

spmd
    rng(0,'combRecursive');
end

This ensures that RandStream.getGlobalStream() returns the same stream on each worker, and then setting the substream will act as it should.

[Eli4ph]

Thanks! I thought that piece is another example and irrelevant with the parfor example. I should have checked the documentation again before asking. Sorry for that. Thanks again.

Answer 2

[Gareth Thomas]

Why do Random Numbers start the same after startup?

[Eli4ph]

Thanks, but this does not help.