Dr Donald Kinghorn (HPC and Scientific Computing)

Intel Python Preview

Written on December 23, 2015 by Dr Donald Kinghorn
Share:

Yes, Intel is doing their own Python build! It is still in beta but I think it’s a great idea. Python is a pretty important programming language. It has a large and growing number of useful libraries for mathematical/scientific computing and machine learning, NumPy, SciPy, pandas, Scikit-learn, PySpark, theano, ...etc… Python is often used as a frontend or “glue code” language so most important API’s get a nice Python interface at some point. You can find Python modules for just about everything. Yeah, it’s important!

Python is not my favorite scripting language for mathematical programming. It is "C" like which makes it awkward for mathematical computing in my opinion. Yes, I like Fortran! To me a vector acts like a “column”, n x 1 matrix, and its “first” element is v[1] not v[0]! Math is hard enough without having to fight your programming language to get indexing and logic correct in your algorithms. Languages that get things “right” are Julia, Lua, Mathematica, Matlab and Fortran. At least that’s my opinion/rant! I do actually like Python but I just have to give that rant when I get a chance :-)

People are using Python for serious computing work so Intel’s effort to do an optimized Python is significant. You can build Python with Intel compilers yourself and link stuff like NumPy to the Intel MKL. However, Intel has some really great compiler guys and I expect that they will get a highly optimized Python put together.

Asside: It makes good sense for Intel to work on a Python build but really it would be great if they were looking at Julia! Julia is the most promising and, possibly, the most important new language to come out in over a decade! http://julialang.org/

Intel Python

You can register to get early access for Intel Python at, https://software.intel.com/en-us/python-distribution

You will need to register for the preview release and you will need to login to your Intel dev account, i.e. if you don’t have an account at Intel you will need to create one. There are Python 2.7 and 3.5 builds for Linux and Windows. The email sent after registration contains a serial number and as says “The software license expires on April 29, 2016”. The note in the email says you will need the serial number but the installer never asked for it. Be aware that this is “preview” stuff! The version I downloaded is tagged 1.0.0tp1 but the release notes are tagged v0.1. Be sure to read the release notes!

The version I downloaded untared to pythoni-2.7.20150803_184913. I ran the install script and installed in my home directory instead of doing a root install. There is an environment setup script bin/pythonvars.sh that you can source to set paths and such.

Lets try it

I did a simple matrix multiply and computed the norm of the result using loops and with numpy functions. The code at the end of this post was run using the CentOS 7.2 default python and the Intel python preview.

Here’s the results:

Note: Using loops in python is ridiculously slow so I used a small matrix size for that and a much larger size for numpy. Keep in mind that matrix multiplication complexity increases as n cubed!

Note: I expected Intel’s numpy to be fast but it is significant that plain old python code is much faster with the Intel version too.

CentOS 7.2 default Python

[kinghorn@i7 python-test]$ python simple-py-test.py 

2.7.5 (default, Nov 20 2015, 02:00:19) 
[GCC 4.8.5 20150623 (Red Hat 4.8.5-4)] 

matrix multiply n = 400    with loops  68.427585125 seconds
matrix norm     n = 400    with loops  0.135160923004 seconds
norm of C is                           40007.6355286 

matrix multiply n = 10000  with numpy  36.3187189102 seconds
matrix norm     n = 10000  with numpy  0.324057102203 seconds
norm of C is                           25001868.2292

Intel Python preview

[kinghorn@i7 python-test]$ source ~/opt/intel/python/2.7.2015.08.03/bin/pythonvars.sh 
[kinghorn@i7 python-test]$ python simple-py-test.py 

2.7.10 (default, Aug  3 2015, 18:21:13) 
[GCC Intel(R) C++ gcc 4.8 mode] 

matrix multiply n = 400    with loops  41.7018299103 seconds
matrix norm     n = 400    with loops  0.0752050876617 seconds
norm of C is                           40007.6355286 

matrix multiply n = 10000  with numpy  11.1052548885 seconds
matrix norm     n = 10000  with numpy  0.0345070362091 seconds
norm of C is                           25001868.2292

Here’s the code;

# simple-py-test.py
# quick python matrix multiply and norm test with loops and numpy

import numpy as np
from   math  import sqrt
import time
import sys

print "\n",sys.version,"\n"

np.random.seed([1234])

# simple loops
n = 400
A = np.random.rand(n,n)
B = np.random.rand(n,n)
C = np.zeros_like(A)

t0 = time.time()
for i in xrange(n):
	for k in xrange(n):
		for j in xrange(n):
			C[i][j] += A[i][k] * B[k][j]
print "matrix multiply n =", n, "   with loops ", time.time() - t0, "seconds"

t0 = time.time()
norm = 0
for i in xrange(n):
	for j in xrange(n):
		norm += C[i][j] * C[i][j]
norm = sqrt(norm)
print "matrix norm     n =", n, "   with loops ", time.time() - t0, "seconds"
print "norm of C is                          ", norm, "\n"

# numpy dot
n = 10000
A = np.random.rand(n,n)
B = np.random.rand(n,n)
C = np.zeros_like(A)

t0 = time.time()
C = np.dot(A,B)
print "matrix multiply n =", n, " with numpy ", time.time() - t0, "seconds"

t0 = time.time()
norm = np.linalg.norm(C)
print "matrix norm     n =", n, " with numpy ", time.time() - t0, "seconds"
print "norm of C is                          ", norm

Happy computing! --dbk

Tags: Intel, Python
Mark Madsen

Don - how does the optimized Intel python stack up to Anaconda with the MKL optimizations? I have to imagine that would be fairly close since Anaconda is using MKL already. Wait, I have a great Puget Systems box you guys built me with CentOS and a Xeon 2650, and I have Anaconda 2.4.1 with MKL Optimizations installed. Here's your code on Anaconda:

[mark@ewens tmp]$ python kinghorn-intel-python-benchmar...

2.7.11 |Anaconda 2.4.1 (64-bit)| (default, Dec 6 2015, 18:08:32)

[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]

matrix multiply n = 400 with loops 60.0356929302 seconds
matrix norm n = 400 with loops 0.104431867599 seconds
norm of C is 40007.6355286

matrix multiply n = 10000 with numpy 9.63591194153 seconds
matrix norm n = 10000 with numpy 0.0205500125885 seconds
norm of C is 25001868.2292

Looks to me like the Intel preview is better on smaller problems, and Anaconda has a bit of an edge on the larger size. It will be interesting to see how this improves towards release! Cheers!

Posted on 2016-01-30 03:19:58
Donald Kinghorn

Hi Mark, Thanks for posting your results, I was wondering the same thing! I was figuring an MKL linked numpy would be about the same as the Intel built version. The thing that surprised me the most in my (short) test was that the plain loops code was faster, I take that as a good sign of things to come. I'll be sure to post any updates from Intel. ... should have something new before the end of Feb.. Best regards --Don

Posted on 2016-02-01 17:54:59
Python is not my favorite scripting language for mathematical

programming. It is "C" like which makes it awkward for mathematical
computing in my opinion. Yes, I like Fortran! To me a vector acts like a
“column”, n x 1 matrix, and its “first” element is v[1] not v[0]! Math
is hard enough without having to fight your programming language to get
indexing and logic correct in your algorithms. Languages that get things
“right” are Julia, Lua, Mathematica, Matlab and Fortran.

I complained about this nearly 3 years ago: https://jdrch.wordpress.com/20...

Posted on 2016-02-19 15:30:22
Mark Madsen

Intel is now inviting folks (presumably those with an active compiler license) to preview the 2017 compiler products, which includes the Python beta. I'm installing it and will be able to see some of my simulation and analysis codes head to head. I'll try to post a link to some results and Jupyter notebooks for folks to replicate in the next week or two....

Posted on 2016-04-06 04:22:33
MIchal

Mark, your approach is great -- would you, kindly, include also as a bottom line test the plain python standard instruction-mix contained in pystone.py? That may seem as an obscure idea, however, there could hardly be a better reproducible plain-pythonesque piece of code, so for assumptions on how [ Intel Python ] may increase the execution-performance, this seems to me as a good base for any non-numpy / non-MKL et al part of the python projects we live with day by day. Thanks for you kind consideration.

Posted on 2016-06-12 18:52:23