Problem with parallel run of Pattern Search (Global Optimization)

Hi All, I have a problem with running pattern search, using 4 available cpu cores. I change the 'UseParallel' default to 'on' with psoptimset, but it only uses 1 cpu core. matlabpool size returns 4 (so I am connected to 4 cpu cores.) If I use fmincon, with 'UseParallel' --> 'on' it uses 4 cpus, but it does not work with pattern search. Any Idea?

Answer 1

[Konrad Malkowski]

Hi Sina, Have you taken a look at the following:

http://www.mathworks.com/access/helpdesk/help/toolbox/gads/bsc7xh9-2.html

There are certain conditions that have to be met for patternsearch to take advantage of Parallel Computing Toolbox. It is hard to guess from your brief description.

cheers

Konrad

[Sina]

Hi Konrad; I tried a simple optimization problem. below you can find the code and the results. Using parallel toolbox makes the process much slower !!!

the code:

function test

clear all

clc

x0 = [10 10];

lb = [0 0];

ub = [20 20];

Options=psoptimset('PollMethod','GSSPositiveBasisNp1','UseParallel','always'...

,'CompleteSearch','on','CompletePoll', 'on', 'Vectorized', 'off','Display','off' ); %,'UseParallel','always','Vectorized','on'

% Solving regularly, without parallel toolbox

t=tic;

for j=1:100

x = patternsearch(@OBJ,x0,[],[],[],[],lb,ub,Options);

end

t=toc(t);

matlabpool size;

fprintf('ElapsedTime=%6.3f PoolSize=%2.1f\n',t,ans);

% Solving with parallel toolbox, with different matlabpool sizes

for n=1:4

command = ['matlabpool open ' num2str(n)];

eval(command)

t=tic;

for j=1:100

x = patternsearch(@OBJ,x0,[],[],[],[],lb,ub,Options);

end

t=toc(t);

matlabpool size;

fprintf('ElapsedTime=%6.3f PoolSize=%2.1f\n',t,ans);

matlabpool close

end

end

function y = OBJ(x)

y = ( 4 - 2.1*x(1)^2 + x(1)^(4/3) )*x(1)^2 ...

+ x(1)*x(2) + ( -4 + 4*x(2)^2 )*x(2)^2;

end

the printed results on Command Window:

ElapsedTime= 4.360 PoolSize=0.0

Starting matlabpool using the 'local' configuration ... connected to 1 labs.

ElapsedTime=51.672 PoolSize=1.0

Sending a stop signal to all the labs ... stopped.

Starting matlabpool using the 'local' configuration ... connected to 2 labs.

ElapsedTime=57.076 PoolSize=2.0

Sending a stop signal to all the labs ... stopped.

Starting matlabpool using the 'local' configuration ... connected to 3 labs.

ElapsedTime=54.798 PoolSize=3.0

Sending a stop signal to all the labs ... stopped.

Starting matlabpool using the 'local' configuration ... connected to 4 labs.

ElapsedTime=57.315 PoolSize=4.0

Sending a stop signal to all the labs ... stopped.

>>

[Konrad Malkowski]

Hi Sina,

This looks like a case where the overhead of parallel computation (i.e., communication, synchronization, etc) outweigh the performance benefits of running in parallel. Try a problem that requires a much longer execution time per iteration of the objective function.