about parfor and spmd speedup

hi,

I have tested the speedup of parfor and spmd as

1)for

tic
for n=1:1000
A=B*C;
end
toc;

2)parfor spmd in one worker in the same pc:

// I have stored B C in the  worker before computing
tic
parfor n=1:1000
A=B*C;
end
toc;

I find that the speed of parfor and spmd is only 2/3 of single matlab. is it possible to make it faster?

[Walter Roberson]

how big are B and C ?

Answer 1

[Walter Roberson]

If your matrices are not very big, then the time required to set up workers and communicate data in to them can overwhelm any saving of using multiple workers.

[Xiaochun]

Actually, the matrix is very large. The B is 70,000*70,000;

And I also test other code, and it needs nearly 4000seconds to finish for the single pc;the speedup is less than 2/3*p(p is processor numbers).Why?

[Walter Roberson]

With a matrix that big, your "for" matrix multiplication would be farmed out to the highly optimized and multithreaded BLAS or LAPACK, using multiple cores or hyperthreading.

I gather that each "parfor" or "smpd" worker requires a new process, and my understanding is that those would only have access to the core or hyperthread they are allocated to.

[Xiaochun]

I see. Thank you very much.

Answer 2

[Knut]

So the spot where parfor can really shine is loops that are : a)Simple enough that parfor can do them in parallell b)complex/non-general enough that MATLAB does not have a low-level library that does it in paralell anyways

?

[Titus Edelhofer]

As a rule of thumb: yes. Another way to formulate b): when you have e.g. a quad core (with hyperthreading), i.e., your task manager shows you 8 CPU usages, and your (loop) code runs at exactly 12.5% average CPU usage ...