OMP: Error #134: Cannot set thread affinity mask.

[JohnNapier]

This issue has been reported in the past, however there were some problems reproducing it. (#6720). I'm providing some code that hopefully helps catch the bug (finally!).

The error appears to be related to the division of a large Series or DataFrame that was built by pd.read_hdf(). To see this, please run the following sample code:

import pandas as pd,numpy as np
# Works
df=pd.DataFrame(np.arange(10000).reshape(-1,4),columns=['a','b','c','d'])
df.to_hdf('hdf5.h5','test')
df=pd.read_hdf('hdf5.h5','test')
x=(df['a']-df['b'])/(df['c']-df['d'])
print 'Done for 10000'

# Does not work
df=pd.DataFrame(np.arange(100000).reshape(-1,4),columns=['a','b','c','d'])
df.to_hdf('hdf5.h5','test')
df=pd.read_hdf('hdf5.h5','test')
x=(df['a']-df['b'])/(df['c']-df['d'])
print 'Done for 100000'

The first operation is carried out, however the second one produces this error when the division is attempted.

OMP: Error #134: Cannot set thread affinity mask.
OMP: System error #87: The parameter is incorrect.

The issue can be avoided by:

1. operating with Numpy arrays, e.g.

x=(df['a']-df['b']).values/(df['c']-df['d']).values 

2. or by setting

pd.computation.expressions.set_use_numexpr(False)

3. or by creating the DataFrame from pd.read_csv().

I'm running Pandas 0.15.2-1, Numpy 1.8.1-2. Enthought's Canopy 1.5.0.2717.

Thanks

[jreback]

@JohnNapier This works for me on windows 7, with the conda distribution (as well as macosx / debian). I recall this is specific to Enthought's packaging AFAICT.

[JohnNapier]

Thank you, @jreback. This seems to be related to an older MKL build. Enthought is aware of it and it should be corrected in a future Canopy version.

[jreback]

closing as related to the Enthought distro

[cournape]

Just to mention this has finally been fixed in Canopy when you update to MKL 11.1.4. I could reproduce the bug w/ MKL 10.3, and could not w/ the 11.1.4

[jreback]

fantastic
thanks @cournape