Best parallel computing option for heavily serialized code

Hello,

I have a MATLAB code that I want to speed up. However due to its serial (dependent on previous values) nature I cannot use parfor. Is there any other parallel computing option that might work for me. I have tried to implement communication jobs, however that does not seem to work.

Regards,

Shivangi

Answer 1

[Raymond Norris]

The very nature of a cascading algorithm (i.e., dependent on previous values) makes it impossible to parallellize. However, perhaps you have the need to run multiples of your code. That could be parallized. That is, with Parallel Computing Toolbox (and perhaps MATLAB Parallel Server) you could run the entire code across different data sets, all at the same time.

[Shivangi Giri]

Thank you for your reply Raymond. So the format of my code if

for loop %runs more than 300 times

call to func1;

some code

call to func2;

some code

end

So I need the for loop to execute in parallel. I cannot use parfor as each iteration is dependent on the previous one. Can you please give a little more insight into how can I parallelize the for loop and mainitain continuity in the execution. I do not have different datasets but few data files that I load before the for loop.

Please note I am at a beginner level with Matlab parallel computing toolbox. Also I am thinking if tranferring data to GPU can execute the code faster. Please help if you are familiar with GPU computing too.

Regards,

Shivangi

Answer 2

[Sulaymon Eshkabilov]

Have seen this fcn: spmd ... end

https://www.mathworks.com/help/parallel-computing/spmd.html

[Shivangi Giri]

Hello Sulaymon,

Thank you for your reply. spmd does not work with dot notations. That is the error I get when implementing spmd. So I guess this might not work for me. In case you have any other suggestions, please recommend.

Regards,

Shivangi