Question

I’m using a block processing approach to handle a calculation between two large matrices.

The code significantly speeds up when using a larger block size. But if I go too large, then I get an Out of Memory error. Currently I hand-tune my code to find the largest working block size for a given input.

My question: how can I automate the process of finding the largest possible block size?

I’ve toyed with wrapping everything in a try/catch block and looping with progressively smaller block sizes till it succeeds. I’m hoping there is a more elegant or idiomatic way.

Answer 1

Before doing the block processing, you can use the MEMORY function to see how much memory is already being used and how much is left available for any additional variables the block processing may need to create. If you can estimate the total amount of memory the block processing steps will need as a function of the block size, you can figure out how large the block size can be before you run out of available memory. This may be easier said than done, since I don’t know exactly how you are doing the block processing.

Here’s a simple example. I’ll start by clearing the workspace and creating 2 large matrices:

>> clear all
>> mat1 = zeros(8000);  %# An 8000-by-8000 matrix of doubles
>> mat2 = zeros(8000);  %# Another 8000-by-8000 matrix of doubles

Now, let’s say I know I will have to allocate an N-by-N matrix of doubles, which will require 8*N*N bytes of memory (8 bytes per double). I can do the following to find out how large I can make N:

>> uV = memory  %# Get the memory statistics

uV = 

    MaxPossibleArrayBytes: 314990592
    MemAvailableAllArrays: 643969024
            MemUsedMATLAB: 1.2628e+009

>> maxN = floor(sqrt(uV.MaxPossibleArrayBytes/8))  %# Compute the maximum N

maxN =

        6274

>> mat3 = ones(maxN);    %# Works fine
>> mat3 = ones(maxN+1);  %# Tanks! Too large!
??? Out of memory. Type HELP MEMORY for your options.

If you are routinely having trouble with running out of memory, here are a couple of things you can do:

• Use single precision (or integer types) for large matrices instead of the default double precision.
• Be sure to clear variables you don’t need anymore (especially if they are large).