Why do I want to use ur_optimize?

The version of ATLAS bundled with Ureka (http://ssb.stsci.edu/ureka) is not designed to work on a large number of CPUs, therefore certain operations requiring linear algebra may perform inadequately to your needs.

ur_optimize rebuilds the underlying ATLAS/LaPACK/BLAS stack to be tailored for your CPU's architecture. After that, the NumPy/SciPy stack is recompiled and linked against the new ATLAS implementation in order to take advantage of these optimizations.

License

BSD

Requirements

RHEL 6.0+ / Fedora 10+

sudo yum install gcc gfortran

Ubuntu 12.04+

sudo apt-get install build-essential

Testing ur_optimize

For this test I used the latest Ureka build (as of 09/04/2014):

Ureka_linux-rhe5_64_1.4.1.tar.gz

Optimizing Ureka

$ git clone https://bitbucket.org/jhunkeler/ur_optimize.git
$ cd ur_optimize
$ ./ur_optimize -p /path/to/Ureka

Consider running ur_optimize in a screen session. ATLAS has been known to take a very long time to build.

SYSTEM

Manufacturer: Dell Inc.
Product Name: OptiPlex 790

BIOS

Vendor: Dell Inc.
Version: A07
Release Date: 09/10/2011

CPU

vendor_id   : GenuineIntel
cpu family  : 6
model       : 42
model name  : Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
stepping    : 7
microcode   : 0x29
cpu MHz     : 3399.867
cache size  : 8192 KB
physical id : 0
siblings    : 8
core id     : 3
cpu cores   : 4

MEMORY INFORMATION

size            : 16GB
type            : DDR3
Mhz             : 1333

test_linalg.py module

#!/usr/bin/env python
import numpy
import scipy.interpolate

def interpolate(user_methods=None, points=1000):
    x = numpy.linspace(1.0, 1000.0, num=points)
    y = numpy.linspace(1.0, 1000.0, num=points)

    methods = ['linear', 'nearest', 'zero', 'slinear', 'quadratic', 'cubic']
    if user_methods is not None:
        methods = user_methods

    for method in methods:
        print("{}...".format(method)),
        try:
            scipy.interpolate.interp1d(x, y, kind=method, bounds_error=False, fill_value=0.0)
        except MemoryError:
            # Common ocurrance with quadratic
            print("failed")

        print("ok")

if __name__ == "__main__":
    interpolate()

Performance Comparison

Before ur_optimize

$ ipython
Python 2.7.5 (default, Jun 19 2014, 11:22:38) 
Type "copyright", "credits" or "license" for more information.

IPython 2.0.0 -- An enhanced Interactive Python.
?         -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help      -> Python's own help system.
object?   -> Details about 'object', use 'object??' for extra details.

In [1]: import test_linalg

In [2]: %timeit test_linalg.interpolate(['quadratic'])
quadratic... ok
quadratic... ok
quadratic... ok
quadratic... ok
1 loops, best of 3: 39 s per loop

39 seconds per operation.

After ur_optimize

$ ipython
Python 2.7.5 (default, Jun 19 2014, 11:22:38) 
Type "copyright", "credits" or "license" for more information.

IPython 2.0.0 -- An enhanced Interactive Python.
?         -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help      -> Python's own help system.
object?   -> Details about 'object', use 'object??' for extra details.

In [1]: import test_linalg

In [2]: %timeit test_linalg.interpolate(['quadratic'])
quadratic... ok
quadratic... ok
quadratic... ok
quadratic... ok
1 loops, best of 3: 1.17 s per loop

1.17 seconds per operation.

Found a bug?

If ATLAS fails to compile for your architecture, please contact the developers of ATLAS, not me: http://math-atlas.sourceforge.net/faq.html#help

Otherwise, feel free to submit a bug report on this project's issue tracker: http://bitbucket.org/jhunkeler/ur_optimize/issues