Hi Shahriar,

have a look at my post here

https://www.researchgate.net/post/What_tricks_do_you_know_in_programming_by_MATLAB_to_perform_programs_faster

I have explained how to use matlab mixed programming capabilities to mix openmp with matlab scripts

also parfor can be used in matlab which is not as flexible

good luck - sina

There are lots of good answers for the question that Sina Haeri refers to.

Extra Matlab toolboxes are expensive. What a lot of people don't realize is that the core Matlab engine already makes very good use of parallelism on computers with multi-core CPUs. In the 90s, everyone knew that loops in Matlab should be avoided because they are handled by the interpreter and not by the low-level binaries. Over the years those binaries (for matrix/vector operations) have been continually optimized. Nowadays, those low-level binaries include multi-threading, where possible. Check your CPU monitor while your program runs to assess utilization.  

If you can't represent your problem using matrix/vector algebra, then I expect you would also have troubles getting your code to run faster on a GPU.

If you can't use use matrix/vector algebra, and your problem is readily factored into non-interacting sub-problems, that do a non-trivial amount of work, and that take roughly the same time to execute, then parallel for loops may help.

But keep this in mind ...

Parallelism may actually make your code run slower, due to the various overheads involved (communications, synchronization/coordination, data transfers, spawning/killing threads etc). If it does run faster, it may not be anywhere near as fast as you were expecting. It all depends on the nature of your problem and the approach that you take. So prepare to be disappointed.

Is it better to plough a field with one ox, 2 horses or 1024 chickens [1]?

[1] https://www.pearsonhighered.com/program/Stone-High-Performance-Computer-Architecture-3rd-Edition/PGM27168.html

the link may help you 

http://www.bu.edu/tech/support/research/training-consulting/online-tutorials/matlab-pct/implicit-parallelism/ 

Hello,

Yes, Matlab does implicitly parallelize your code as suggested and yes you have to avoid loops as much as possible and rely on vectorization. However, these techniques are not flexible and sometimes simply because of the nature of the algorithm or the type of access to array element these techniques are not simply applicable. 

I am not exactly sure what algorithm Shahriar is trying to implement or what type of calculation it involves but does not seem that it involves only simple matrix operation that can simply speed them up using vectorisation and implicit parallelisation. Besides, a very limited number of Matlab functions are GPU-enabled so it is not again a general approach

That's why I suggested mixed programming. 

Cheers

Sina

MATLAB parallelism methodology is an analog of MPI

https://www.researchgate.net/publication/283288686_Parallel_programming_in_MATLAB_for_modeling_an_economy

Chapter Parallel programming in MATLAB for modeling an economy

Using Matlab for this is a terrible idea.  Yes, Matlab will take advantage of threads for basic matrix operations (if sufficiently large).  But it's just using Intel's MKL for that - you can use MKL directly.  Similarly with parfor and DCE - Matlab will expose capabilities provided by other standard/open-source packages like MPI, but will charge you for it.

These days, compilation is basically instantaneous, and performance will be better than Matlab if you write using a high-level interface like Boost.  Matlab is really only for the weak (who can't imagine leaving a GUI) or those who insist on the convenience of Matlab's library of toolboxes (which, again, are just repackaging techniques freely available elsewhere.)

Hello everybody and thank you all.

Mr Hahn

Why do you think it's a terrible idea? do you mean using Matlab parallelism for HMM computations is a terrible idea or using the Parallel Computing Toolbox itself?

yes I agree it's easier than MPI. 

 you said "Matlab will expose capabilities provided by other standard/open-source packages like MPI, but will charge you for it."

how can I find this capabilities ( I dont know toolbox or functions or ...) for MPI?  This is my masters thesis and I'm really stuck. 

Thank you for your time.