11

As the title suggests numpy.dot (I think numpy is general) performs much slower after updating my system. The example code that I'm running to compare is:

from numpy import *
import time
A=random.random((1000,1000))
B=random.random((1000,1000))
st=time.time();dot(A,B);end=time.time();print end-st

The following code takes ~0.09s on my other computer (The computer with the problem used to run as fast as the other one), but the code takes ~0.26 on the computer with the problem.

Here are my attempts to solve this problem. My first guess was, 1: ATLAS is not connected to numpy. So, I installed atlas packages using synaptic, libatlas3-bas, libatlas-doc, libatlas-dev, libatlas3gf-bas, libatlas-bas-dev.

After installing this, I checked whether numpy is linked to atlas or not by typing

import numpy.distutils.system_info as sysinfo
sysinfo.get_into('atlas')

The return message is the following:

ATLAS version 3.10.1 built by buildd on Sat Jul 27 19:04:50 UTC 2013:
   UNAME    : Linux roseapple 3.2.0-37-generic #58-Ubuntu SMP Thu Jan 24 15:28:10 UTC 2013 x86_64 x86_64 x86_64 GNU/Linux

   INSTFLG  : -1 0 -a 1 -l 1
   ARCHDEFS : -DATL_OS_Linux -DATL_ARCH_x86SSE2 -DATL_CPUMHZ=1596 -DATL_SSE2 -DATL_SSE1 -DATL_USE64BITS -DATL_GAS_x8664
   F2CDEFS  : -DAdd_ -DF77_INTEGER=int -DStringSunStyle
   CACHEEDGE: 1048576
   F77      : /usr/bin/x86_64-linux-gnu-gfortran-4.8, version GNU Fortran (Ubuntu/Linaro 4.8.1-8ubuntu1) 4.8.1
   F77FLAGS : -fomit-frame-pointer -mfpmath=sse -O2 -msse2 -fPIC -m64
   SMC      : /usr/bin/c99-gcc, version gcc (Ubuntu/Linaro 4.8.1-8ubuntu1) 4.8.1
   SMCFLAGS : -fomit-frame-pointer -mfpmath=sse -O2 -msse2 -fPIC -m64
   SKC      : /usr/bin/c99-gcc, version gcc (Ubuntu/Linaro 4.8.1-8ubuntu1) 4.8.1
   SKCFLAGS : -fomit-frame-pointer -mfpmath=sse -O2 -msse2 -fPIC -m64
Out[12]: 
{'define_macros': [('ATLAS_INFO', '"\\"3.10.1\\""')],
 'include_dirs': ['/usr/include/atlas'],
 'language': 'f77',
 'libraries': ['lapack', 'f77blas', 'cblas', 'atlas'],
 'library_dirs': ['/usr/lib/atlas-base/atlas', '/usr/lib/atlas-base']}

So, I assume that it is linked. I also went into numpy/core/ path and typed

$ldd _dotblas.so

and the output is

linux-vdso.so.1 =>  (0x00007fff16ffa000)
libcblas.so.3 => /usr/lib/libcblas.so.3 (0x00007fa913908000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fa9136eb000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fa913322000)
libatlas.so.3 => /usr/lib/libatlas.so.3 (0x00007fa912d8f000)
libgfortran.so.3 => /usr/lib/x86_64-linux-gnu/libgfortran.so.3 (0x00007fa912a77000)
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fa912860000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fa91255c000)
/lib64/ld-linux-x86-64.so.2 (0x00007fa913d4f000)
libquadmath.so.0 => /usr/lib/x86_64-linux-gnu/libquadmath.so.0 (0x00007fa912320000)

Then, to make sure libcblas.so.3 is linked to ATLAS, I typed

$ /usr/sbin/update-alternatives --config libblas.so.3
There are 4 choices for the alternative libblas.so.3 (providing /usr/lib/libblas.so.3).

  Selection    Path                                     Priority   Status
------------------------------------------------------------
  0            /usr/lib/openblas-base/libopenblas.so.0   40        auto mode
* 1            /usr/lib/atlas-base/atlas/libblas.so.3    35        manual mode
  2            /usr/lib/libblas/libblas.so.3             10        manual mode
  3            /usr/lib/openblas-base/libopenblas.so     30        manual mode
  4            /usr/lib/openblas-base/libopenblas.so.0   40        manual mode

Press enter to keep the current choice[*], or type selection number: 1

The following steps didn't fix my problem...

  1. And I realized my CPU throttling was on, and that ATLAS doesn't work well when CPU throttling is on. So, I turned it off using rcconf, and reinstalled Atlas as above and reinstalled Numpy as well... But it doesn't solve the problem...

  2. Now I'm running python 2.7.5+, so I tried reinstalling it as well.. didn't solve the problem.

  3. I checked my memory ..

             total       used       free     shared    buffers     cached

    Mem: 5959 2345 3614 0 86 730

    -/+ buffers/cache: 1528 4431

    Swap: 6133 0 6133

And my memory is free.. and my matlab code runs as fast as before.. so I don't think it is memory problem..

Can anyone help, please? It seems like linear algebra package of numpy is slow... Now, It delays my simulation a lot, so I can't really run my codes. Thank you very much in advance! Let me know if there's any confusion in my question.!

Annie Jihyun Park
  • 111
  • 3
  • Are other parts of NumPy slow? If not, it might be that the new Ubuntu is shipping an ATLAS that is worse than their old one. ATLAS auto-tunes at compile time, so it might have auto-tuned to some slow build bot if the Ubuntu packagers weren't careful enough. – Fred Foo Oct 15 '14 at 16:25

1 Answers1

6

When you say "and reinstalled Numpy as well", did you use pip or apt-get/synaptic?

The Ubuntu (or Debian) packaged version of NumPy available through apt-get install python-numpy is pre-complied. On the other hand, installing NumPy using pip install numpy or python setup.py within the NumPy source directory uses the available compilers to build NumPy from source.

You may then prefer to compile NumPy using ATLAS on your own computer. ATLAS will optimize the NumPy library using the characteristics of your computer (effectively requires turning off CPU throttling).

apt-get install -y python-pip python-dev build-essential
pip install -U cython

apt-get install -y gcc gfortran libatlas-base-dev liblapack-dev
pip install numpy

You may wish to read the ATLAS installation guide or the instructions to build NumPy and SciPy from source on Linux

It may be what you've already done, but then someone else could find it useful!

francois
  • 81
  • 5