How can I solve memory leak in fft?

When using fft function many times with different input size it causes memory leak (as seen in the physical memory).

For example, the following matlab code :

for i=1:1000, y=fft([1:i*1000]); end

MATLAB Version: 7.14.0.739 (R2012a)

How can I solve it?

Answer 1

[Steve Eddins]

It is not an actual memory leak. MATLAB does cache (save for reuse) tables of trigonometric values for use in subsequent calls to FFT with vectors of the same length. (There is a significant performance benefit in doing so.) However, this cache does have a finite size; it will not continue to grow indefinitely.

[Shmuel]

Thank you for your fast replay. However when running the full program the memory reaches 99% (16Gb in total) and it slows dramatically (due to running many times FFT with vectors in different size). Only by closing MATLAB the memory is released. From your answer I understand I need to release the cache somehow or to restrict it from growing. Can you direct me how to do that?

[David]

I am performing ffts and iffts on large arrays >4GB. However, because the matlab memory footprint grows after the initial calls to fft and ifft, the usable memory for variables decreases. Because the arrays are very large to begin with, this causes issues maxing out the computer's memory. Is it possible to clear this fft cache without restarting matlab? For cases like this, low memory overhead is more important than execution speed.

[John D'Errico]

This would suggest the value of a 'nocache' option in fft.

[Noah Prisament]

To clear this cache, you can utilize the "fftw" command which is used to modify the settings and behavior of the "fft" function. You can execute the following commands in your scripts or in the command window to clear the cache:

fftw('swisdom', []);
fftw('dwisdom', []);

These will clear the single precision and the double precision caches respectively. Another important command to minimize the memory usage of the "fft" function is:

fftw('planner','estimate');

as the other planner options employ more memory in order to improve their precision. More information on the "fftw" command can be found here.

These uses of "fftw" should help with limiting the size of the cache and releasing it periodically.