Parallel computing toolbox
By sunfrost at Wed, Mar 12 2008 2:30pm |
Hi forum,
I currently have some calculations running on Matlab on a quad core mac pro. However, as matlab does not support the SMP architecture only one core gets to do the work. I was wondering of anyone has experience with the parallel computing toolbox (http://www.mathworks.com/products/distribtb/) from matlab in an effort to use all the Mac pro's power.
Any input would be welkom as it could speed up things 4 times and 8 times in the future.
Cheers,
Koen
Parallel Matlab on a Quad Core Mac
Howdy,
I have a quad core PPC PowerMac on which I run extensive Monte Carlo simulations. Since you are using a shared memory system, a single logged in user may launch multiple instances of Matlab simultaneously. Thus, in computationally intensive portions of your code, you can launch multiple instances of Matlab to perform tasks in parallel, have them write temporary files, and then read the temporary files back into the initial instance of Matlab. Each instance of Matlab gets its own process ID and is sent to a free processor. Of course there is substantial overhead with launching each instance of Matlab. Nevertheless, I have run simulations with four such "threads" and have thereby reduced the duration from weeks to days. It is an inelegant hack, but it is free and it works quite well in extreme situations.
Cheers,
-asn
Matlab 2007b
Please be more specific about the MATLAB version and also your toolboxes.
I guess your best bet is 2007b with distribution toolbox AND engine.
Sounds like you're missing the ENGINE which is the core part.
Here is what you can do.
Call Mathworks and ask them you want to evaluate their new product which is 2007b.
Usually they send out people to install the engine so ask about it.
At the end of evaluation period, if you like it buy their product.
Of course you can install by yourself but I wouldn't recommend because it was long and painful process for me.
Maybe it was because I tested right after they launched the product. Gave up to use MATLAB for parallel computing. Now I use OpenMP on G5 Xserve.
TIP:
Please don't expect 4 times speed up just because you're using 4 cores. It doesn't work in that way.
You might get 1.5~2.0 depending on your code. Use matlab profiling to catch bottleneck.
Depending what you're doing you may waste your time and money. I use parallel computing if I need huge amount of memory.
Currently the lab has a
Currently the lab has a campus license for Matlab 2007b without the necessary toolboxes (otherwise I wouldn't ask for your experience with these tools).
I know I must not expect a 4x increase in performance but as the code has a lot of moving window analysis in it I should gain some speed by using a parfor() instead of the normal for() loop.
I looked into other alternatives like mentioned by asn, but I didn't get it up and running. A lot of the code written to implement the suggested method of running several threats of matlab has become obsolete in version R14 and up. And although speeding up things would be nice, optimalisation isn't that critical that I can spend to much time in writing such code myself.
Anyway thanks for the advice, I think I'm going to request an evaluation version of the toolboxes to see what gives.