rand function is slow with a parfor loop

I'm trying to multithread a stochastic simulation using the parallel computing toolbox. I've profiled the program extensively to ensure that it is optimized. However, when I'm using PCT with 4 local workers, I find that the rand function consumes an inordinate amount of computational time far beyond what it uses in a single threaded program. Here is a representation of the code in question.

if(~(rand <= Pr))

I am stumped as to why this would be 10-15x slower than the single threaded case, unless rand is a synchronized function. I'm using a MacPro with 16 GB RAM and 2 quad core processors (MATLAB R2008b) for reference.

Answer 1

[Alex]

All computer based random number generators need a 'seed'. I don't know about Matlab, but this seed, if not provided directly by the user, is based off of the current system time. (by using the same 'seed', you can then get repeatable results, even in a 'random' environment).

I think, 2008b does not have the updated random generator methods.

http://www.mathworks.com/help/techdoc/math/bsn94u0-1.html This link talks about some of the older changes that have been made to the rand function.

Answer 2

[Walter Roberson]

2008b does have the updated random number generator methods.

http://blogs.mathworks.com/loren/2008/11/13/new-ways-with-random-numbers-part-ii/

MATLAB uses a constant seed for each session, but a different seed for each parallel worker. This is described more in Peter Perkin's posting at http://www.mathworks.com/matlabcentral/newsreader/view_thread/299841

Answer 3

[James]

Each worker uses a provided seed, not the default. Even if they used the worker ID as a seed, that wouldn't explain why calling rand is much slower in a parfor loop than a for loop.

[Konrad Malkowski]

James could you provide a more complete example that reproduces the issue?