MATLAB: In parfor , why a variable cannot be classified

parallel computingparfor

In the code below I get an error

Error: The variable Matrix in a parfor cannot be classified.See Parallel for Loops in MATLAB, "Overview".

    parfor i=1:n
    try
        query = 'SELECT * FROM TABLE WHERE Criterion = '' criteria(i) ''';
        curs = exec(conn, query);
        curs = fetch(curs);
        close(curs);
        results = curs.Data;
        A = f(results);
        B = g(results);
        C = h(results);
        D = k(results);
        Matrix(i,6:9) = horzcat(A,B,C,B);
    catch
        Matrix(i,6:9) = {0,0,0,0};
    end
end

Any ideas what I am missing?

Best Answer

write to

tMatrix(i,:) = horzcat(A,B,C,B);

or write {0,0,0,0} there. Then after the parfor, let

Matrix(:,6:9) = tMatrix;

The difficulty you are running into is that you are trying to update only part of one row of a matrix, which requires that the matrix be input to the parfor as well as output from the parfor. parfor can handle rows that are only input and rows that are only output, but it doesn't like input-update-output as your code is requiring. The adjusted version I show has the entire row of tMatrix be output, and that is fine with parfor.

Related Solutions

MATLAB: I am unable to perform database operations such as FETCH or EXEC inside the PARFOR LOOP in MATLAB 7.14 R2012a

Within PARFOR LOOP, database operations like exec and fetch are not possible even if database connection object is supplied to these workers.

The main reason is because, MATLABPOOL OPEN , opens new processes (not new functions). The client process (PCT) which holds the database connection object cannot share or send the object to other processes. This is analogous to file handles. In regular programming language like C++, two different C++ applications or processes (not functions) cannot share file handles as they are separate processes or applications in memory. Similarly database connection object which is created in one process (PCT Client process) cannot be shared with worker processes. Workers in PCT are not functions but are separate processes in the operating system.

The ultimate main reason for this is that database connection objects are TCP/IP ports and multiple processes cannot access the same port. That is why database connection object are not serialized. Ports cannot be serialized.

The workaround for this is to establish the database connection within the PARFOR LOOP. This means that each worker will open a new port to the database and communicate to the database through its individual port. But this also means the database will have to bear the overhead of maintaining multiple connections with different processes. This might be a performance issue based on the application.

MATLAB: How to modify non-consecutive rows of a matrix in a parfor loop

Extract the rows you want, parfor over the subset, write the results back after the parfor.

In your description you talk about "editing" but your code shows only writing new values. If new values are to be written without reference to the old values, then parfor producing a consecutive matrix and write the result into the sparse array afterwards. For example,

target_index = [5 10 15];
t_matrix = spalloc(length(target_index), size(my_Matrix,2));
parfor ii = 1 : 3
  i = target_index(ii);
  t_matrix(ii, :) = some values probably involving i
end
my_Matrix(target_index, :) = t_matrix;   %non-parallel write back of results

Best Answer

Related Solutions

MATLAB: I am unable to perform database operations such as FETCH or EXEC inside the PARFOR LOOP in MATLAB 7.14 R2012a

MATLAB: How to modify non-consecutive rows of a matrix in a parfor loop

Related Question