Class with static variables in parallel global optimization algorithm

Hello,

I have a global optimization program dealing with large matrices (several gigabytes of data), so in order to save memory, a class with static variables was implemented similar to the implementation in Static Data, and then one object of this class is created, initialized and passed as an argument to a function handle acting as the objective function of global multistage optimization algorithm (Particle Swarm + Pattern Search). When parallelization in optimoptions is true:

optimoptions( ...
'UseParallel', true);

The optimization always yields false results, but when parallelization is turned off, it works correctly.

Thanks in advance!

Answer 1

[Matt J]

All variables are cloned when parallelization is used. Each parallel worker operates with an independent copy of any variable sent to it.

[Omar Kamel]

So as they are persistent variables, they are cloned/created with their initial null value?

In other words, how can I force Matlab to copy the object with its data to each parallel worker?

[Matt J]

We need to see code...

[Omar Kamel]

This is the code of the class:

classdef cDampingConstData

    properties
        
    end
    
    methods (Static )
        
        function out = setget_A_0_0(data)
            persistent Var;
            if nargin
                Var = data;
            end
            
            out = Var;
            
        end
        
        function out = setget_delta_A_to_delta_d(data)
            persistent Var;
            if nargin
                Var = data;
            end
            
            out = Var;
            
        end
    end
end
    

[Matt J]

Okay, but we also need to see how it is used in the optimization.

[Omar Kamel]

Sorry for the late reply.

%create object and initialize
dampingModelDataObject = cDampingConstData;   %create object of type cDampingConstData
dampingModelDataObject.setget_delta_A_to_delta_d(delta_A_to_delta_d);    %initialize object with fields
dampingModelDataObject.setget_A_0_0(A_0_0);    %initialize object with fields
dampingModelDataObject.setget_states_num(A_0_0);    %initialize object with fields

%objective function handler, object is passed to it with another constant inputs and d as variable
fun = @(d) DampIT.SumModalDamping_sparse_oo(reshape(d,1,1,num_damping_coefficients), dampingModelDataObject,targetModalDamping);


%optimization options
options = optimoptions('particleswarm', 'Display', 'iter', ...
                'HybridFcn', {@patternsearch, patternsearch_hybrid_options}, ... %hybrid optimization
                'MaxTime', obj.settings.pp_swarm_maxTime, ...
                'MaxIterations', obj.settings.pp_swarm_maxIter, ...
                'FunctionTolerance', obj.settings.pp_swarm_tolFun, ...
                'PlotFcn', { @pswplotbestf}, ...
                'SwarmSize', obj.settings.pp_swarm_swarmSize, ...
                'OutputFcn', memLog_swarm, ...
                'UseParallel', useparallel_flag, ... %true
                'UseVectorized', false, ...
                'MaxStallIterations', obj.settings.pp_swarm_maxStallIter, ...
                'InitialSwarmMatrix', InitialSwarmMatrix);
            
                
%problem settings
problem.solver = 'particleswarm';
problem.objective = fun;
problem.nvars = num_damping_coefficients;
problem.lb = x_start;
problem.ub = x_end;
problem.options = options;

%start optimization
[d_optimized, fval, exitflag, output] = particleswarm(problem);
                

Answer 2

[Walter Roberson]

Look again at the link you provided . Notice the point about static data not being saved with an object . The process of sending variables to parallel workers involves save and load.

[Omar Kamel]

I couldn't understand your answer.

What I want to implement is that: save a matrix in memory (constant matrix), pass this matrix to a function that is evaluated with parfor, so every worker shoud use only the instance that is saved only 1 time in the memory

parfor i = 1:1000
    out = some_function(constant_matrix, variable_data)    
end

This constant_matrix is created before this parfor command.

[Walter Roberson]

https://www.mathworks.com/help/parallel-computing/parallel.pool.constant.html

However, that saves the matrix once per worker, not "only 1 time in the memory". If you strictly need "only 1 time in the memory" then you should look in the File Exchange for https://www.mathworks.com/matlabcentral/fileexchange/28572-sharedmatrix which uses operating system shared memory.

[Matt J]

Interesting.