Is randperm() consistent under different hardware platform, OS, MATLAB versions?

I use randperm() to generate sets of numbers repeatly using a for loop. I always make sure I keep the RandStream consistent over different running environment by using such as,

RandStream.setDefaultStream(RandStream('mt19937ar','seed',1); or RandStream.getDefaultStream();

Recently, I've notice that sometimes I get different randomization sequences across the sequential randperm sets, although I was using an identical seed. The only difference was 1) Physical processor (I was using number of heterogeneous cluster nodes)

Should I get the same sequences across different processors?

Answer 1

[Walter Roberson]

Exactly which point you seed at could make a difference. When you use multiprocessing operations, the different tasks are seeded differently, so you would need to seed inside the parallel operation.

Answer 2

[Matt Fig]

Let's test it out!

What do you get? Post in a comment...

% r2011b, Windows 7, 64bit maxNumCompThreads=4
rng(678)
randperm(10)
randperm(10)
ans =

     6     5     9    10     1     4     2     3     8     7
  ans =

     7     4     3     6     8     5     1     2     9    10

Now, on the same system:

% r2007b, Windows 7, 64bit maxNumCompThreads=1
rand('twister',678)
randperm(10) %  Same as above...
randperm(10) %  Same as above...

[Walter Roberson]

I get the same pair of results on R2012a, Mac OS-X Lion, 64 bit

[Matt Fig]

Now we just need a linux person to chime int!

[Daniel Shub]

I thought Walter was a Linux person... Same in Linux (R2011a, Arch, 64-bit).

[Matt Fig]

Is Mac OS a linux thing? I don't know much about it.

[Daniel Shub]

No, Mac and Linux are different. I just thought Walter used Linux and not a Mac.

[Walter Roberson]

I was on Linux where I used to work.

Answer 3

[Peter Perkins]

John, there is no difference in the random number generators between different hardware.

You seem to be saying that this was a distributed environment. Without knowing more, it's hard to say anything concrete, but you have to be careful about your random stream initialization on the workers. Especially if this is a parfor loop, since the way the iterations get doled out is unpredictable.