The upper limit of core usage in numpy vector operation

Question

I love numpy because multi cores are automatically used on vector operation.

But, I recently noticed that only 4 cores are used although my machine has 8 cores.

Why doesn't numpy use all cores in the machine? Is it possible to let numpy use more cores on numpy's vector operation?

I'm using Mac OSX 10.8.

Update

np.show_config() shows below:

lapack_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'iomp5', 'pthread']
    library_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/include']
blas_opt_info:
    libraries = ['mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'iomp5', 'pthread']
    library_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'iomp5', 'pthread']
    library_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/include']
blas_mkl_info:
    libraries = ['mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'iomp5', 'pthread']
    library_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/include']
mkl_info:
    libraries = ['mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'iomp5', 'pthread']
    library_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/lib']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['/Users/john/.pyenv/versions/anaconda-2.0.1/include']

"I love numpy because multi cores are automatically used on vector operation." - this is only true for certain functions that call BLAS/LAPACK library functions (e.g. matrix multiplication). This isn't a feature of numpy itself, but rather depends on which BLAS/LAPACK libraries numpy is linked against at runtime. What is the output of numpy.show_config()? — ali_m
– ali_m, Commented Apr 27, 2016 at 9:26

armatita · Accepted Answer · 2016-04-27 11:48:45Z

2

MKL, ideally already uses, an optimized number of threads for your system. You can check those like this:

python -c "import ctypes; mkl_rt = ctypes.CDLL('mkl_rt.dll'); print mkl_rt.mkl_get_max_threads()"

If you need to change it you might want to change MKL_NUM_THREADS environment variable:

python -c "import ctypes; mkl_rt = ctypes.CDLL('mkl_rt.dll'); print mkl_rt.mkl_get_max_threads()"

For more on MKL multi-threading check this, and the API.

I'm quoting Anton Malakhov, from Intel. You might want to check that question (Configure number of processors available to MKL NumPy) since it's very similar to yours.

answered Apr 27, 2016 at 11:48

armatita

13.6k9 gold badges54 silver badges53 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

Light Yagmi Over a year ago

Thank you very much! But, OSError: dlopen(mkl_rt.dll, 6): image not found happens. I've tried mkl_rt.dylib, mkl_rt.so, and mkl_rt.lib. Both of them occurs the same error.

armatita Over a year ago

@rkjt50r983 I think it's because it is failing in finding them. The path will depend on your system but you should be able to locate it in your Python installation than here: Lib\site-packages\numpy\core

armatita Over a year ago

@rkjt50r983 Just noticed you are on Anaconda. You might want to take a look at this if you are getting Fatal Error with the above recipe (I have the same problem in WinPython). Link: github.com/numpy/numpy/issues/6923

Collectives™ on Stack Overflow

The upper limit of core usage in numpy vector operation

Update

1 Answer 1

3 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

Update

1 Answer 1

3 Comments

Your Answer

Sign up or log in

Post as a guest

Related