Concatenation in Matlab (with or without copying elements)

Assume we have X = [A, B] where A and B are two vectors (or matrices). Is concatenation operation copies elements in A? If so is it possible to do concatenation without copying elements?

Possible answer to the second. What if I do the following in order to avoid copying elements: X(1:length(A)) = A; X(length(A)+1:length(A)+length(B)) = B; Does 'copy-on-write' work in this context?

Follow-up question: How can I benchmark these two codes with respect to memory usage? Any ideas?

[Mateusz]

Or perhaps X=A; X=[X,B] would avoid copying elements?

Answer 1

[Jan]

If you concatenate two arrays, the values of both must be copied. The memory used to store the values of an array need to be stored in a continuos block. Therefore even A = [A, B] has to copy all elements of A and B.

The copy-on-write strategy does not matter in this case.

If you have to store large arrays and really need to avoid a copy, use a CELL as a container of the subarrays. Of course any operations using this new object must be re-programmed and will be complicated (slow!), e.g. a matrix-multiplication.

This is even worse:

X(1:length(A)) = A;
X(length(A)+1:length(A)+length(B)) = B;

The indexed copy duplicates A at first. But in the 2nd line, a new larger array is allocated for X and the values of A are copied a second time before B is appended!

It is very hard to define a "memory footprint" exactly, and in consequence the measurement is not trivial also. The JIT might free the memory used by A if it is not used anymore in a function. But if this freed memory is reused to create X depends on the interaction with the operating system. E.g. the current processor load can influence if the memory is cleared and re-used, or if a new block is allocated and the garbage collection happens later.

Answer 2

[the cyclist]

This page will provide you with a lot of information about memory management in MATLAB: http://www.mathworks.com/support/tech-notes/1100/1106.html

When you do ...

 >> X = A;

MATLAB doesn't make a copy of A, until you change an element of X. (Then it copies the entire array.) However, I am pretty sure that that does not apply to concatenation. I don't think there is a way to carry out

 >> X = [A,B];

without actually creating the new array in memory.