Code covered by the BSD License

Demo files for "Parallel Computing with MATLAB on Multicore Desktops and GPUs" Webinar

Jiro Doke (view profile)

05 May 2011 (Updated )

Demo files for "Parallel Computing with MATLAB on Multicore Desktops and GPUs" Webinar (May 4, 2011)

Parameter Sweep of ODEs

Parameter Sweep of ODEs

This is a parameter sweep study of a 2nd order ODE system.

We solve the ODE for a time span of 0 to 25 seconds, with initial conditions and . We sweep the parameters and and record the peak values of for each condition. At the end, we plot a surface of the results.

Initialize Problem

m     =                    5;  % mass
bVals = linspace(0.1, 5, 50);  % damping values
kVals = linspace(1.5, 5, 70);  % stiffness values
[kGrid, bGrid] = meshgrid(bVals, kVals);
peakVals = nan(size(kGrid));

Parameter Sweep (Serial)

We'll first do the parameter sweep computation in a serial for loop.

disp('Computing in serial...');drawnow;

tic;
for idx = 1:numel(kGrid)
% Solve ODE
[T,Y] = ode45(@(t,y) odesystem(t, y, m, bGrid(idx), kGrid(idx)), ...
[0, 25], ...  % simulate for 25 seconds
[0, 1]) ;      % initial conditions

% Determine peak value
peakVals(idx) = max(Y(:,1));
end
t1 = toc;
fprintf('Elapsed time is %0.2f seconds.\n', t1);
Computing in serial...
Elapsed time is 16.98 seconds.

Parameter Sweep (Parallel)

Next, we convert the for loop to a parfor loop and start a pool or MATLAB workers.

disp('Computing in parallel...');drawnow;

matlabpool open

tic;
parfor idx = 1:numel(kGrid)
% Solve ODE
[T,Y] = ode45(@(t,y) odesystem(t, y, m, bGrid(idx), kGrid(idx)), ...
[0, 25], ...  % simulate for 25 seconds
[0, 1]) ;      % initial conditions

% Determine peak value
peakVals(idx) = max(Y(:,1));
end
t2 = toc;
fprintf('Elapsed time is %0.2f seconds.\n', t2);

% Close MATLAB Pool
matlabpool close
Computing in parallel...
Starting matlabpool using the 'local' configuration ... connected to 2 labs.
Elapsed time is 8.45 seconds.
Sending a stop signal to all the labs ... stopped.

Visualize

figure;
surf(bVals, kVals, peakVals);
xlabel('Damping (b)');
ylabel('Stiffness (k)');
zlabel('Peak Response');
view(50, 30)

Speed Up

We can see that we get a very good speed up. Since this loop has very minimal overhead in terms of data transfer, we may see a super-linear speed up, which sometimes happens due to the Just In Time acceleration.

fprintf('\n\nSpeed up (time serial / time parallel): %0.2f\n', t1/t2);

Speed up (time serial / time parallel): 2.01