Codistributed array as class property

Hi all,

I have a large array (a database let's say, which stays fixed), which I load into a codistributed array directly to avoid communication overhead. I do this as follows:

spmd
   % Allocate memory in codistributed array
   database = codistributed.zeros(some_size);

     % Get the local part on the worker
     lp = getLocalPart(database);

     % Load the data from files into the local part
     lp = loadDataFromFiles();

     % Put the local part back
     database = codistributed.build(lp,getCodistributor(database));
  end

I do this to allow several workers to process parts of the database simultaneously. This works fine, as long as I perform this in the 'main' script of the program, and the database variable is available in the workspace. However, when I want to put all functionality into a class, things stopped working. Let's say I have something as follows:

   classdef database

    properties
        data        
    end

    methods

        function obj = database()
            % Load the data
            spmd     
                % Allocate memory in codistributed array
                database = codistributed.zeros(100);                
            end

            obj.data = database;

        end

        function queryData(obj)
            spmd
                % Perform some operations on the database, which is stored
                % as a distributed array in obj.data.
                size(getLocalPart(obj.data))
            end
        end
    end 
  end

Matlab let's me create an object of the class, and the property data is initialised. When checking the properties of the class object, it says that obj.data is a distributed array, as I would expect.

However, when I try to execute the queryData method, I get the following error:

Error using distributed/getLocalPart (line 195)
It is not possible to call "getLocalPart" directly on a distributed array.  To call
"getLocalPart", you must enter an SPMD block and operate on the corresponding codistributed
array.

I can't seem to find any related topics. So my question obviously is: what am I doing wrong? And is it even possible to 'hide' a distributed array in a class property?

Thanks!

Answer 1

[Edric Ellis]

spmd blocks can only transform correctly between the distributed and codistributed representation when the distributed array is a separate variable, i.e. not a field of a struct or object. So, I think you need to work around this limitation like so:

function queryData(obj)
    % pull out the distributed array into its own variable
    tmpData = obj.data;
    spmd
        % Perform some operations on the database, which is stored
        % as a distributed array in tmpData
        size(getLocalPart(tmpData))
    end
end

If your method modifies the underlyin distributed array, then you'd also need to assign back into obj.data after the spmd block.

[philip.joris]

Thank you for your answer Edric. This is indeed possible.

I do have one question regarding this approach: the distributed array is rather large. If I pull obj.data into tmpData, and then call tmpData in the spmd block, does this mean that the whole matrix tmpData is transferred to workers? Or is this the case anyhow, because I'm storing the codistributed array as a distributed array in a variable on the client side?

[Edric Ellis]

tmpData is a distributed array - this means the data is not transferred back and forth between the client and workers. The client-side view of tmpData is effectively a series of "pointers" (or something like that) which refer to the codistributed parts that are stored on the workers.

[philip.joris]

Okay, this clears things up. Thanks Edric!