Handle object array doesn't update when parfor is used within a class method

I have a class that is similar to this:

 classdef Example < matlab.mixin.Copyable

    properties
        data = [];
    end

    methods

        function obj = Example()
        end

        function process(obj)

            parfor ii = 1:numel(obj)
                objSlice = obj(ii);
                objSlice.process_sub()
                obj(ii) = objSlice;
            end

        end

        function process_sub(obj)
            obj.data = datestr(now, 'HH:MM:SS.FFF');
        end

    end

 end

As you can probably see, I would like to process the elements of an array of this class in parallel. This works as expected when the parallel pool is closed, e.g.:

% Initialise array of objects
num = 5;
for nn = 1:num
    exArray(nn) = Example();
end

% Process the object array in parallel
exArray.process();
exArray.data

ans =

11:36:56.387

ans =

11:36:56.387

% ... snip

However, when the parallel pool is open, the data is not updated correctly; e.g. (after processing):

exArray.data

ans =

       []

ans =

       []

% ... snip

Is this behaviour expected or a bug? And either way, how can I get the object array to process in parallel?

I have seen other similar questions (e.g. here , here), and the documentation ( here ), but I think my issue is different because the parfor loop is actually inside the handle object, and I haven't been able to find any ideas about this.

Thanks very much in advance for your help.

Answer 1

[Rahul Goel]

Hi,

As per the documentation, the workers do not propagate the changes back to the client. In the code you mentioned, you took care of this by:

obj(ii) = objSlice;

However, this change has to be propagated back to the client explicitly. This can be done by returning the changed objects from the function to the caller as:

function obj = process(obj)

and capturing this returned value in the caller function as:

exArray = exArray.process();

So, the code you provided works as expected with these modifications:

classdef Example < matlab.mixin.Copyable
    properties
        data = [];
    end
    methods
        function obj = Example()
        end
        function obj = process(obj)
            parfor ii = 1:numel(obj)
                objSlice = obj(ii);
                objSlice.process_sub()
                obj(ii) = objSlice;
            end
        end
        function process_sub(obj)
            obj.data = datestr(now, 'HH:MM:SS.FFF');
        end
    end
 end

When following objects are created and call the process function:

% Initialise array of objects
num = 5;
for nn = 1:num
    exArray(nn) = Example();
end
% Process the object array in parallel
exArray = exArray.process();
exArray.data

It gives the following output:

>> exArray.data

ans =

11:04:54.894

ans =

11:04:54.894

ans =

11:04:54.894

ans =

11:04:54.898

ans =

11:04:54.899

Note: Before executing the enhanced version of this class, clear the class from the memory by:

>> clear classes

Hope this helps.

[mbar170]

Thanks for your answer Rahul.

I can see that this approach works, but I think the fact that it is necessary is inconsistent with the expected behaviour of a handle object.

Even if something like the following worked it would be closer to what I expect:

classdef Example < matlab.mixin.Copyable
    properties
        data = [];
    end
    methods
        function obj = Example()
        end
        function process(obj)
            obj = obj.process_parfor();
        end
        function obj = process_parfor(obj)
            parfor ii = 1:numel(obj)
                objSlice = obj(ii);
                objSlice.process_sub()
                obj(ii) = objSlice;
            end
        end
        function process_sub(obj)
            obj.data = datestr(now, 'HH:MM:SS.FFF');
        end
    end
end

However, in the absence of another workaround, I guess I'll have to change my syntax to explicitly capture the modified object.

Thanks again, Matt