Can one do Parallel Computing coding in MATLAB ?
I want to implement some parallel algorithm in MATLAB.
To be more specific, I want to implement the parallel maximum weight bipartite matching algorithm (Hungarian method), and for that I am thinking to implement it in MATLAB.
Matlab offers some very constrained access to parallelism - you cannot write a threaded program, for instance. Then again, the whole point of Matlab is convenience at the expense of efficiency and expressive range. If your goal is to explore algorithm implementation, I'd say Matlab is a poor choice. If you just want to get some calculation done, leveraging parallelism, Matlab may be expedient.
Sure you can. There is a package called the Parallel Computing Toolbox that allows you to execute Matlab code in parallel using multiple cores, GPUs and compute clusters. If software licensing is an issue you can look at the "parallel" package for GNU Octave.
Hi Dr. Healy,
Thank you very much for the information and quick reply...
Matlab offers some very constrained access to parallelism - you cannot write a threaded program, for instance. Then again, the whole point of Matlab is convenience at the expense of efficiency and expressive range. If your goal is to explore algorithm implementation, I'd say Matlab is a poor choice. If you just want to get some calculation done, leveraging parallelism, Matlab may be expedient.
To Mark Hahn,
Thanks for your suggestion.
I want to implement the parallel maximum weight bipartite matching algorithm (Hungarian method), and for that I am thinking to implement it in MATLAB.
What do you think?
Hello, Shibsankar! I have worked with parallel computing in Matlab a few times. I have some examples of codes. If you wish, you may contact me directly inbox and I would send it for you.
I'm no graph expert, but reviewed this page: http://www.netlib.org/utk/lsi/pcwLSI/text/node223.html
my impression is that there is a lot of parallelism available, but it's "small", in the sense of not much work-per-data or work-per-phase. that means that there will be a lot of scatter/gather/synchronization in Matlab's parallel model. Threaded/shared-memory is much finer-grained, and might work better for this workload. I notice there are papers describing the use of GPUs for this, too, where each pseudo-thread effectively has shared-memory access to all the data.
To Mark Hahn
Thanks for all the information.
I could not able to see the review page provided by you as it is down for now (probably).
To Mark Hahn,
With reference to the my above comment; the page http://www.netlib.org/utk/lsi/pcwLSI/text/node223.html is working now.
Another approach: google for "call matlab code from C++" then use your "driver" program in C++ to call your Matlab routines, using thread forking, synchronization, muitex etc in your C++ code.
You can use parallel for loop, i.e., parfor. It is very easy to use.
- Badges
- Science topic