Parallel Computing Toolbox on local machine

Hi all, I'm working on a 24-core local computer with 64 GB RAM.

I would like to use the Parallel Computing Toolbox locally to improve the computing time. I've looked the different webinars available on mathworks.fr but have still some questions about it. Could someone help me ?

1 - For simply add parallellized comuting in my script, I've just to add something like "matlabpool open local", right ? But I've read that there can be no more than 8 workers. Does that imply that I will not use the full 24 cores at the same time ? How can I improve that ?

2 - The next step is to replace the "for" loops by "parfor" loops. Are there ohter improvements I can do in my script (namely with the others Matlab functions) ? I use a lot of matricial functions, are they automatically optimized by the "matlabpool open" ?

3 - I've heard that Matlab implicitly optimizes the computing regarding the number of core available. Is that true ? How can I control what kind of parallisation is done ?

Thank in advance for all your informations !

Best Regards,

Pierre

Answer 1

[Titus Edelhofer]

Hi Pierre,

1. More or less yes, although the current version (R2011b) supports up to 12 local workers instead of the previous eight. For more workers you will need to employ the MATLAB Distributed Computing Server, but I would definetely start with the local workers first.

2. Again more or less yes: of course it makes no sense just to write everywhere where a for loop is just parfor. Use it on outermost for loops.

3. Yes, MATLAB uses multithreading for so called implicit parallelization. But keep in mind, that the workers run usually single threaded so that they don't "disturb" each other.

Titus

Answer 2

[Daniel Shub]

These are the type of questions that TMW support staff can answer very well.

Answer 3

[Thomas]

Hi, here is a link for a workshop I conducted..

http://docs.uabgrid.uab.edu/wiki/MATLAB_workshop_2011#Workshop_Demo.27s

You want to look over the following sections in the demos.

1. Serial Job

3. Convert serial job to parallel and run in locally.

Other than that I think Titus has answered all your questions..