MATLAB: How to Pin Workers to Specific Cores

Parallel Computing Toolboxparpool

I would like to know how to pin parpool workers to specific cores. I have a dual socket system, so each socket has faster access to certain resources, like GPUs.

Therefore, say I have 8 workers in my pool and 8 GPUs. I want to pin each worker to a specific core on my dual socket system.

Best Answer

What OS are you using? If Windows, you could adapt the code from this answer on Stack Overflow. One way to pick the affinitization is to use an spmd block after opening the pool and using labindex to choose the CPU.

On Linux, you can use this approach to set affinity. I would adapt that approach and use the undocumented feature('getpid') means of accessing the current process' PID, something like this

spmd
  pid = feature('getpid');
  system(sprintf('taskset -pc %d %d', labindex-1, pid));
end

Related Solutions

MATLAB: When starting local workers can I specify an affinity mask for each worker in Parallel Computing Toolbox R2013b

The ability to automatically change the affinity mask for local workers is not available in Parallel Computing Toolbox R2013b.

On Windows in general it is possible however to change the affinity mask of a process by calling the SetProcessAffinityMask Windows API function; in MATLAB you could do this from a MEX-file; an example MEX-file is attached to this article. You will need to compile the MEX-file using the MEX function. Once the file has been compiled you should be able to use something like the following to assign an unique core to each worker:

% Open the pool
matlabpool open local
% On all workers run the SetAffinity function
spmd
    SetAffinity(uint64(2^(labindex-1)))
end
% Your own PARFOR/SPMD code goes here
% Clean-up/close
matlabpool close

On Linux you could use the taskset command which can be called through the SYSTEM function. Basically you can use the same approach as above only instead of calling SetAffinity you could use something like:

system(sprintf('taskset -p %d %d',2^(labindex-1),feature('getpid')))

*Important Note *when working with Intel CPUs with Hyperthreading enabled, you will only want to assign your workers to physical cores and not the logical hyperthreading cores. This means you will need to make some changes to how you compute the masks for the workers..

MATLAB: In parallel processing, are multiple workers used for anything else than parfor loops

Using SVD under a parfor will likely slow the performance of SVD. SVD is inherently multithreaded and make use of multiply cores calling it as it is. Here are a list of function that are inherently multithreaded:

http://www.mathworks.com/support/solutions/en/data/1-4PG4AN/

To comment further on how we could speed up your code, we may need to see more of what you have within the parfor loop.

Best Answer

Related Solutions

MATLAB: When starting local workers can I specify an affinity mask for each worker in Parallel Computing Toolbox R2013b

MATLAB: In parallel processing, are multiple workers used for anything else than parfor loops

Related Question