Making a memory efficient analysis script
This tutorial gives advice about how to do data-analysis in a memory-efficient way when large amount of data is analyzed. It does not provide information about any FieldTrip functions or any analysis steps.
Neurophysiological data can become quite large with the result that disk space, RAM and processing time can become compromised. It is important to pay attention on how much memory is used because a non-efficient analysis can result in “out of memory” errors in MATLAB. However, how to program MATLAB in a memory-efficient way is also not always obvious. Some common issues are discussed below.
Memory efficient coding tips
- Work on your programming style. Take a look here for a concise summary of the recommended style of programming.
- Downsample your data (but backup your original data), e.g. using ft_resampledata
- Change data to single-precision (after preprocessing by using ft_struct2single or by using “cfg.precision = 'single'” in certain functions)
- Check if you really have to “cfg.keeptrials = 'yes'” in ft_freqanalysis.
- If you are working on a single subject, make sure other subjects are no longer in memory. This might seem trivial, but many people assign unique variables to subjects and forget to clear them.
- Perhaps most importantly – once in a while let someone else go through your scripts to see if they can be optimized.
- Within a script or function make sure you clear large variables that you don’t need anymore using the clear statement. Note that MATLAB’s memory use might not be intuitive. For instance, reloading a large dataset into the same variable may result in MATLAB allocating twice the memory you actually need.
Save your data to disk
Remember to always backup your original data that was acquired on external hard disks, CDs or DVDs for long-term storage. If possible, the backup should include the presentation code that was used in the experiment and a small ascii *.txt file with the recording details.
When using FieldTrip for large analyses, it is recommended to save one MATLAB variable to a single file. That will result in a lot of files in your data directory and in first instance you may consider that to look messy. However, the advantage is that you can easily manage the data, delete results that you don't need any more, check that the results are complete for all subjects, check that the timestamps of the files with certain results are consistent for all subjects (e.g. after you have updated some parameters and rerun part of the analysis), …
When writing intermediate results, consider if you really need to save all intermediate steps in your analysis pipeline. For instance with MEG data it might not take that much time to calculate your planar gradient. It will save you a lot of disk space if you only have to write your axial data to disk. When using a high sampling frequency during acquisition, you may be able to downsample your data to save disk space and speed up all subsequent processing steps.
Do make sure you save the important parameters (e.g. rejected trials) so you can always rerun your script. Subject specific information should be added to the subject specific script.
Load only as much data as you need
Only import into MATLAB as much of a large data set as you need for the problem you are trying to solve. Many users are tempted to try and load the entire file first, and then process it with MATLAB. This is not always necessary. Use the whos function with the -file option to preview the file. This command displays each array in the MAT-file that you specify and the number of bytes in the array:
whos -file session1.mat Name Size Bytes Class Attributes S2 1x1 723 struct x 100x200 72 double sparse Mat4 4x20 640 double A 3151872x1 3151872 uint8 Seq 1x912211 912211 int8
If there are large arrays in the MAT-file that you do not need for your current task, you can selectively import only those variables that you want using load, for instance:
seq = load(‘session1.mat’,’Seq’).
Avoid creating temporary arrays
Avoid creating large temporary variables, and also make it a practice to clear those temporary variables you do use when they are no longer needed. For example, when you create a large array of zeros, instead of saving to a temporary variable A, and then converting A to a single:
A = zeros(1e6,1); As = single(A);
use just the one command to do both operations:
A = zeros(1e6,1,'single');
Using the repmat function, array preallocation and for loops are other ways to work on nondouble data without requiring temporary storage in memory.
Use nested functions to pass fewer arguments
When working with large data sets, be aware that MATLAB makes a temporary copy of an input variable if the called function modifies its value. This temporarily doubles the memory required to store the array. One way to use less memory in this situation is to use nested functions. A nested function shares the workspace of all outer functions, giving the nested function access to data outside of its usual scope. In the example shown here, nested function setrowval has direct access to the workspace of the outer function myfun, making it unnecessary to pass a copy of the variable in the function call. When setrowval modifies the value of A, it modifies it in the workspace of the calling function. There is no need to use additional memory to hold a separate array for the function being called, and there also is no need to return the modified value of A:
function myfun A = magic(500); function setrowval(row, value) A(row,:) = value; end setrowval(400, 0); disp('The new value of A(399:401,1:10) is') A(399:401,1:10) end
Using Appropriate Data Storage
MATLAB provides you with different sizes of data classes, such as double and uint8, so you do not need to use large classes to store your smaller segments of data. For example, it takes 7,000 KB less memory to store 1,000 small unsigned integer values using the uint8 class than it does with double.
Preallocate contiguous memory when creating arrays
In the course of a MATLAB session, memory can become fragmented due to dynamic memory allocation and deallocation. For and while loops that incrementally increase the size of a data structure each time through the loop can add to this fragmentation as they have to repeatedly find and allocate larger blocks of memory to store the data. When memory is fragmented, there may be plenty of free space, but not enough contiguous memory to store a new large variable. To make more efficient use of your memory, preallocate a block of memory large enough to hold the matrix at its final size before entering the loop. When you preallocate memory for an array, MATLAB reserves sufficient contiguous space for the entire full-size array at the beginning of the computation. Once you have this space, you can add elements to the array without having to continually allocate new space for it in memory.
The following code creates a scalar variable x, and then gradually increases the size of x in a for loop instead of preallocating the required amount of memory at the start:
x = 0; for k = 2:1000 x(k) = x(k-1) + 5; end
Change the first line to preallocate a 1-by-1000 block of memory for x initialized to zero. This time there is no need to repeatedly reallocate memory and move data as more values are assigned to x in the loop:
x = zeros(1, 1000); for k = 2:1000 x(k) = x(k-1) + 5; end
Clear old variables from memory when no longer needed
When you are working with a very large data set repeatedly or interactively, clear the old variable first to make space for the new variable. Otherwise, MATLAB requires temporary storage of equal size before overriding the variable. For example,
a = rand(100e6,1) % 800 MB array a = rand(100e6,1) % New 800 MB array ??? Error using ==> rand Out of memory. Type HELP MEMORY for your options. clear a a = rand(100e6,1) % New 800 MB array
Summary and suggested further readings
It is much advised to also read the following information:
These are FAQs that are related to memory issues:
|2010/03/31 15:59||Robert Oostenveld|
|2011/01/20 14:23||Robert Oostenveld|
|2009/06/17 16:48||Robert Oostenveld|
|2010/01/26 09:54||Robert Oostenveld|