Why is assigning a new handle object in an array slow?

I have a handle class that contains an array of node handle objects. Generally I have upwards of 2 million nodes. Initially they're the array is full of pointers to the same sentinel node, and as my algorithm proceeds they're assigned to new node objects. For Example:

   classdef SlowExample < handle

         properties
             foo = Node();
         end

         methods
             function func(obj)
                 obj.foo(1:2000000) = Node();
                 for i = 1:800
                     bar = Node();
                     obj.foo(i) = bar;
                 end
             end
         end
     end

In this example I just used a handle class for the nodes:

classdef Node < handle
    %NODE a mock object

    properties
    end

    methods
    end

end

When run with the profiler, you can see that there is a bottle neck at:

obj.foo(i) = bar;

For me it spent about 300 seconds on that line, and little on anything else. I haven't been able to figure out why. It doesn't seem to happen if foo isn't the data of an object.

Any ideas why this is, and if possible if there's a way to solve it?

[Yasha]

Yes, it should be

 obj.foo(i) = bar;

Thanks per isakson, I missed that typing it here. The code I ran though didn't have that problem. Also, I'm definitely getting more than that, also using tic and toc as well.

I'm using R2011b, could it be that it's been fixed since my version?

Answer 1

[per isakson]

There is a performance penalty for using objects. That has been discussed here at Answer and in other places, e.g. Considering Performance in Object-Oriented MATLAB Code. The examples, e.g. linked list, in the documentation do not scale. The MathWorks are working on it and the performance has improved with each release. (I use R2012a.)

My execution with the code according to OP

    tic
    se = SlowExample();
    se.func
    toc

returns

    Elapsed time is 139.794373 seconds.

Modified code, based on a hint (avoid referencing properties in loops) I picked up here at Answer

    classdef SlowExample < handle
        properties
            foo = Node();
        end

        methods
            function func(obj)
                fun(1:2000000) = Node();
                for i = 1:800
                    bar = Node();
                    fun(i) = bar;
                end
                obj.foo = fun;
            end
        end
    end

Executions time

    Elapsed time is 0.188095 seconds.

" upwards of 2 million nodes" I doubt that Matlab is ready for an array of user defined objects that large.

The design approach of Tree data structure as a MATLAB class might be an alternative. It is faster and has potential to be much faster.

[Yasha]

Thank you, good to know. Unfortunately, I don't think I'll be able to make that modification, as my code isn't creating the new Nodes all in a loop like that.