How to reduce the time for data collection from workers in parallel computing?

Hello everyone. I am having a parallel for loop 'parfor' that iterates for N times. In each iteration, a new vector with size (M x 1) is created and accumulated to a matrix, resulting in a M x N matrix (very big matrix) after finishing the loop. The parfor works efficiently and performs the job without problems, but it takes very long time to collect the data back from the workers. For example, if the total consumed time is 10 minutes, I can say that parfor takes about 1 minute, and remaining 9 are consumed in the data collection. How to reduce this time efficiently?

parfor i=1:N      
    stuff ...
    k(i,:) = b            % b is a temporary vector (M x 1) inside the parfor loop
end

[Walter Roberson]

Are you pre-allocating k?

parfor deliberately runs loops in reverse order to attempt to cut down the pre-allocation costs, but it wouldn't hurt to pre-allocate to be sure.

[Ahmad Gad]

No, I am not pre-allocating k. My problem usually happens after finishing the loop. I never thought that pre-allocating it will save me time.

I pre-allocated it and this reduced the time to a reasonable amount. After I got the k, I need to use sparse function for it to get the matrix completed. Any Ideas about a faster alternative?

Thanks and best Ahmad