MATLAB DualCPU Hyper-Threading support

Hello everyone.

I'm going to buy a dual CPU workstation and would like to know if a vectorized MATLAB code uses

all the cores of both CPUs without the need to

use the Parallel Computing Toolbox.

Thanks!

[Guillaume]

Some functions of matlab rely on libraries that are optimised to use parallelisation when it makes sense, even if you don't have the parallel toolbox.

It's not documented though and subject to change from version to version. Your m script/function itself will always use just one thread.

[Emiliano Rosso]

Thanks,I refer to the use of " : " instead of " for " loop because I have independent process for my matrix. This is a general question not specific for any particular code. So,if I have 2 CPU with 4 (8 thread) cores each I'll use 4 (8) or 8 (16) cores when I use the vectorization function " : " ?

I verified that using 1 intel xeon w3680 6 cores 12 thread I can use all 12 thread even if not 100% powerfull because my RAM is slow. That's the same using 2 cpu (nodes)?

Thanks!

[Jan]

I do not understand what

use of " : " instead of " for " loop

means. Of course it depends on the specific code, how it is executed. Calling the colon operator "." a "vectorization function" is strange.

If your code exhausts 12 threads, you can expect that it let 8 threads work efficiently also. So what exactly is your question?

[Emiliano Rosso]

[m,n]=size(matrix);
for i=1:m
for l=1:n
matrix(i,l)=matrix(i,l)^2;
end
end

I vectorize the code :

[m,n]=size(matrix)
matrix(:,:)=matrix(:,:)^2

if I use 1 cpu xeon w3550 4 cores 8 thread:

1200 sec.

if I use 1 cpu xeon w3680 6 cores 12 thread:

800 sec.

The 2 cpu are identical except for the number of cores and I can verify that all thread are

working in both cases.

Now :

if I use 2 cpu xeon w3680 6 cores 12 thread does my code will use the second cpu?

Answer 1

[Jan]

Yes. Matlab will use the available cores. At least most likely. Matlab might split the array columnwise and if the input has 5 columns, there is a chance that only 5 cores are used. Many functions are multi-threaded, but this is applied for "large" arrays only.

I'm not sure, if hyper-threading is used in Matlab. Maybe only the physical cores are used for the reason of efficiency.

Note that:

[m,n]=size(matrix);
for i=1:m
  for l=1:n
    matrix(i,l) = matrix(i,l)^2;
  end
end

and

[m,n]=size(matrix)
matrix(:,:)=matrix(:,:)^2;

do not compute the same result. The 2nd code works for square matrices only and is a matrix multiplication. Equivalent to the first code would be:

% No need to determine the size...
matrix(:,:) = matrix(:,:) .^ 2;
%                         ^ elementwise squaring

It would be even more efficient, to omit the (:,:):

matrix = matrix .^ 2;

[Emiliano Rosso]

Thanks for your answer.

The " ^ " instead of ".^ " was a mistake.