Execute parfor iterations in order!

Hi,

I want to use Matlab parallelization to do this parfor in order like 1, 2, 3, ... , and I have limitted number of workers/cores, say 4.

parfor cnt = 1:20

do_something(cnt)

end

This is important for me to run cnt = 1:4 first, and once one of them finished, I need to start running next one which will be 5. I would appreciate your suggestions.

Thanks!

[James Tursa]

Why do they need to be run in order? Does the result of the 1:4 iterations serve as inputs to the 5:8 iterations, etc.?

[Sina Gharebaghi]

1- Actually, results of 1:4 does not serve as inputs of 5:8. However, based on results of 1:4, I may decide that running 5:8 is not required. I just need to stop running 5:end to decrease runtime.

2- If I want to use results of 1:4 as inputs of 5:8, how can I do that?

Thank you!

Answer 1

[Walter Roberson]

This is not possible with parfor(), and it is unlikely that Mathworks will ever implement an option to support this.

parfor divides the available range up into chunks depending on the number of cores, and starts some of the chunks going, leaving other chunks in the pool to be allocated to whichever workers finish fastest. parfor also leaves a number of individual iterations to the end, to be allocated one-by-one as workers finish.

Furthermore, when parfor starts allocating the initial (larger) chunks, the first chunk it starts is the last iterations. This gives some advantages in pre-allocating arrays.

In your situation, instead of using parfor, you should use parfeval() . That allows you to control exactly which work gets sent out, in which order, and allows you to cancel chunks that have not finished yet.

[Sina Gharebaghi]

Many thanks Walter! I will check "parfeval".

[Steven Lord]

You may be able to use two parfor loops, one after the other. For the first one:

parfor k = 1:4
% body of parfor
end

If after evaluating the results of that parfor loop you decide you want to run the rest of the iterations:

parfor k = 5:20
% body of parfor
end

[Sina Gharebaghi]

Thank you, Steven!

Answer 2

[Jaya]

You need not run 4 times first and then start with 5. You can directly do the parfor for 1:20. Assuming yours is a quad core, then each core will take the task of doing 5 iterations. How it splits that is not a thing for us to worry.

I asked this similar question some time ago.Here is the link.

https://www.mathworks.com/matlabcentral/answers/1581124-parfor-iterations-and-number-of-cores

[Sina Gharebaghi]

Jaya, thanks for your response.

Actually, I am not repeating "parfor". I have only one command "parfar cnt=1:20". I need to run them in order. First, I need to run 1-4. However, it is not happening in parfor.

[Jaya]

By 'run them in order', do you mean all the operations inside the for loop aren't independent?

From my knowledge, if output of one iteration is required for next iteration then I don't think parfor can be used for that. Since, logically, the cores working at the 'same' time cannot get the previous output of another core that performed (say) 1st iteration.

If this isn't the case for you then maybe you can elaborate on that 1-4 thing in your next comment and I or someone else can better answer then.

[Sina Gharebaghi]

I may not need to continue simulations for all 20, I may stop somewhere in between, like 8. Then, running 9:20 will not be required. That is why I need to run them in order. I understand that parfor may not be the best option for me in Matlab's parallel computing!

[Jaya]

Then I don't really think or don't know if or how can you still use parfor for this case. Since I understand your situation as something that requires simulation stopping based on some result you get in between the iterations. And as you might know, parfor cannot use a break statement.

[Sina Gharebaghi]

Thanks for your help. You are right. Probably, I will not be able to use "parfor".