Using Groups of Rows in a Parfor Loop

Below is my code that is attempting to populate a variable using section of rows. The original ContractFile is hundreds of thousands of rows - the thinking is I can populate the variable on different workers using the parfor loop which will populate sections of 10,000 rows at a time on a different worker. Example: Rows 1:10,000 to one worker, rows 10,001:20,000 to a different worker, etc. This code works as a regular for loop, but breaks as a parfor loop and I can't figure out why. Thanks!

parfor i = 1:Contracts
    Rows = (i-1)*10000+(1:10000);
    Var1(Rows,:) = ContractFile(Rows,2) .* ContractFile(Rows,8);
end

[Matt J]

Why is the loop necessary? Why not simply,

Var1=ContractFile(:,2) .* ContractFile(:,8);

Answer 1

[Matt J]

As mentioned in my comment, your example does not make it clear why a loop is necessary at all. However, the reason for your difficulty is that your parfor code violates these rules. One way to fix it is as follows:

Var1=nan(10000,Contracts);
A=reshape(ContractFile(1:Contracts*10000,2),10000,[]);
B=reshape(ContractFile(1:Contracts*10000,8),10000,[]);

parfor i = 1:Contracts
    Var1(:,i) = A(:,i).*B(:,i);
end

Var1=Var1(:);

[Derek De Vries]

Thanks! I know what I posted can be performed without a parfor loop - it was just a major simplification of what I'm actually attempting so I'll see what I can do with the code you've posted. Thanks for the link - is there an easier way to explain why it wasn't working for the original code I posted. Where am I violating the parfor code conditions specifically?

[Derek De Vries]

It doesn't appear this code breaks up the rows into groups to be utilized.

What I'm hoping is that I could solve for Var1 by having:

Worker 1: Take Rows 1:10,000 of Column 2 of the Contracts File and Multiply by Rows 1:10,000 of Column 8 of the Contracts File

Worker 2: Take Rows 10,001:20,000 of Column 2 of the Contracts File and Multiply by Rows 10,001:20,000 of Columns 8 of the Contracts File

Etc.

Would this be faster for several different calculations over time since the calculations are broken out over several workers instead of just one worker multiplying for example 4 million rows by 4 million rows all on its own? I understand what's happening in your suggested code, however it doesn't seem to parse out groups of rows to split the calculations among different workers. Thanks!

[Matt J]

is there an easier way to explain why it wasn't working for the original code I posted.

You are attempting to treat Var1 and ContractFile as "sliced variables". When you index sliced variables, the indexing has to be a simple expression involving the loop variable. Using arbitrary index vectors like "Rows" is not allowed. From the documentation,

"Form of Indexing. Within the first-level of indexing for a sliced variable, exactly one indexing expression is of the form i, i+k, i-k, or k+i. The index i is the loop variable and k is a scalar integer constant or a simple (non-indexed) broadcast variable. Every other indexing expression is a positive integer constant, a simple (non-indexed) broadcast variable, a nested for-loop index variable, colon, or end."

It doesn't appear this code breaks up the rows into groups to be utilized.

By reshaping the data into 10000xN matrices, the batches of data you are looking for become separate matrix columns, and can be indexed in the form A(:,i), B(:,i) and Var(:,i), which is legal for sliced variables according to the above rule.

What I'm hoping is that I could solve for Var1 by having:

This would happen if the number of workers N is equal to the number of loop iterations "Contracts". Otherwise, it will split the loop into N smaller loops and each worker will process a consecutive chunk of Contracts/N iterations.

[Derek De Vries]

I apologize, I misunderstood how this code was working and it works great - so thank you!

A couple of follow up questions:

1. How does the final line of code execute? It appears to me that the code is just saying make Var1 equal all rows of the current Var1 variable. So I would think it would remain a 10,000xN matrix. How did it know to bring all of those columns in 2:N to the row 10,001 then 20,001, etc?
2. If the last "block" of rows isn't the full amount of exactly 10,000 (Example: There are 123,456 total rows in the Contract File) is there a way to code that the final grouping should be 3,456 rows instead of 10,000 inside of the parfor loop or would that piece have to just be performed outside of the loop at the end?

[Matt J]

1. Var1=Var1(:) is equivalent to Var1=reshape(Var1,[],1). See also

https://www.mathworks.com/help/matlab/ref/colon.html#description

2. You could pre-pad the array (e.g., with NaNs) to have an even multiple of 10000 elements and then proceed as above. Or, instead of reshaping, you could also partition your input data into cells and use cell array indexing. To do this splitting, I recommend mat2tiles (Download) because it will handle non-even multiples of 10000 rows, as demonstrated in the code below.

 CF=mat2tiles(ContractFile,[10000,Inf]);
 N=numel(C);
 
 Var1=cell(numel(C),1);
 
 parfor i=1:N
     
     Var1{i}=CF{i}(:,2).*CF{i}(:,8);
     
 end
 
 Var1=cell2mat(Var1);