MATLAB: Parallel Random Number Generator

parallel computingrandom number generator

I would like to run simulations requiring random numbers on each worker. First, I create a random stream of numbers on each worker:

    parpool
    spmd
    s = RandStream.create(’mrg32k3a’,’NumStreams’,numlabs,’StreamIndices’,labindex); 
    RandStream.setGlobalStream(s);
    end

Then, I perform the simulations:

    spmd
    output = some_function(input); %%%this function uses rand() and randn(); 
    end

Am I correctly using RandStream to ensure statistically independent random numbers across workers? Can I combine the two spmd commands?

Best Answer

Yes, I believe your code is correct, and is basically the same initialization that is already carried out by Parallel Computing Toolbox.

You can safely merge the two spmd blocks.

Related Solutions

MATLAB: Obtaining same values at avery simulation using rand function

I think you should use the newer rng

    rng default;
    x = randn(100,1);
    rng default
    y = randn(100,1);
    max(abs(x-y))

x and y are the same sequence.

In general, you can capture the random generator settings and resuse them

   S = rng;
   x = randn(100,1);
   rng(S)
   y = randn(100,1);

MATLAB: Am I receiving the same results when I run a simulation on a Linux cluster that uses the RAND command in MATLAB 7.13 (R2011b)

The reason why the results are frequently the same is because the different jobs at the different nodes of the Linux cluster are starting at the same time. Therefore, the seed given by the clock to the random number generator algorithm is the same.

A possible way to work around this issue is to select the seed of the random number generator in such a way that not only it depends on the clock, but it depends on another variable which is different for every MATLAB session running in parallel.

If you are using Parallel Computing Toolbox (PCT) a simple solution can be the following:

s = RandStream('name_of_algorithm', 'seed', sum(100*(1+labindex/numlabs)*clock));
RandStream.setGlobalStream(s)

where NUMLABS returns the total number of labs currently operating on the current job and LABINDEX returns the index of the lab currently executing the function.

If you are not using PCT but you are starting different MATLAB sessions from a script, a simple solution can be the following:

x = feature('getpid');
s = RandStream('name_of_algorithm', 'seed', sum(x*clock));
RandStream.setGlobalStream(s)

where the command FEATURE('GETPID') returns a different identification number for each MATLAB session running.

Best Answer

Related Solutions

MATLAB: Obtaining same values at avery simulation using rand function

MATLAB: Am I receiving the same results when I run a simulation on a Linux cluster that uses the RAND command in MATLAB 7.13 (R2011b)

Related Question