Using Parfor for two separate matrix multiplication

Hello,

I am trying to use parfor for independent matrix multiplication inside a time loop. I am measuring the cpu time and actual calculation time. I have a dual core computer and I aspect cpu time being double the execution time but they are not. In addition I don't gain any time with parfor. I don't know what I am doing wrong or if I am doing something wrong. A simplified version of my code would look like

if matlabpool('size') == 0 % checking to see if my pool is already open
matlabpool open 2
end

D=rand(1000,1000);
D2=rand(1000,1000);
R=rand(1000,1);

qum=cputime;   
tickron=tic; 

parfor (t=1:1000,1)
u1=D*R;
u2=D2*R;
end

Taking=toc(tickron)
muq=cputime-qum

if I use parfor

Taking = 1.76698075488077  (execution time)
muq = 0.062400400000115    (cpu time)

if I use for

Taking = 1.26792301952307
muq =    2.51161610000008

why there is no gain when I use parfor and also why cpu time is much smaller in parfor?

Thank you

Erdem

Answer 1

[Edric Ellis]

I think the problem here is that your parfor syntax explicitly asks the loop to run on only one worker:

parfor (t=1:1000,1)

I would suggest removing the trailing ,1 from that. See the parfor reference page for more about the syntax.

Your cputime call is measuring the CPU time on the MATLAB client, not on the workers. If you wanted to measure the CPU time on the workers, you'd need to do something like this:

spmd, cpu0 = cputime(); end
parfor idx = 1:100, rand(1000)*rand(1000); end
spmd, cpu1 = cputime() - cpu0, end

This records the initial CPU time in cpu0 which is a different value on each worker. After the parfor loop, each worker then gets a value for cpu1 which is the elapsed CPU time. You can calculate the total CPU time like so:

spmd, gplus(cpu1), end

Answer 2

[Kamuran]

Thank for the reply. Lets forget about the CPU time and simply work with tic and toc. In the following code is it possible to make the C=B*B and D=A*A in parallel.

matlabpool open 2

A=rand(2000);
B=rand(2000);
C=zeros(2000,2000);
D=zeros(2000,2000);
tic
parfor idx = 1:10,
    D=A*A;
    C=B*B;
    E=A*B;
end
toc

Thanks

[Edric Ellis]

That code does run C=B*B and D=A*A in parallel, so I'm not sure what your question is...

[Kamuran]

I thought what this code does is run: for example idx=1 and idx=2 ... in parallel but not when idx=1 run calculation of D and calculation of C in parallel.

[James Tursa]

Are you only trying to get multiple matrix multiplies to happen in parallel? Or is your concern really more general than that for your real problem? E.g., an OpenMP mex routine could easily be written to call the BLAS library in parallel for multiple matrix multiplies, but it is not clear from your examples what you are really trying to do and how the parallel calculations would occur for your real problem. I.e., a mex routine can be written that does the D=A*A and the C=B*B matrix multiplies in parallel on each iteration, but I am not sure how this would apply to your real problem.