# Use Python for Scientific Computing

As computers become more and more powerful, scientific computing is becoming a more important part of fundamental research into how our world works. We can do more now than we could even imagine just a mere decade ago.

Most of this work has been done traditionally in more low-level languages, such as C or FORTRAN. Originally, this was done in order to maximize the efficiency of the code and to squeeze out every last bit of work from the computer. With computers now reaching multi-GHz speeds, this is no longer the bottleneck it once was. Other efficiencies come into play, with programmer efficiency being paramount. With this in mind, other languages are being considered that help make the most of a progammer's time and effort.

This article discusses one of these options: Python. Although Python is an interpreted language and suffers, unjustly, from the stigma that entails, it is growing in popularity among scientists for its clarity of style and the availability of many useful packages. The packages I look at in this article specifically are designed to provide fast, robust mathematical and scientific tools that can run nearly as fast as C or FORTRAN code.

The packages I focus on here are called numpy and scipy. They are both available from the main SciPy site (see Resources). But before we download them, what exactly are numpy and scipy?

numpy is a Python package that provides extended math capabilities. These include new data types, such as long integers of unlimited size and complex numbers. It also provides a new array data type that allows for the construction of vectors and matrices. All the basic operations that can be applied to these new data types also are included. With this we can get quite a bit of scientific work done already.

scipy is a further extension built on top of numpy. This package simplifies a lot of the more-common tasks that need to be handled, including tools such as those used to find the roots of polynomials, doing Fourier transformations, doing numerical integrals and enhanced I/O. With these functions, a user can develop very sophisticated scientific applications in relatively short order.

Now that we know what numpy and scipy are, how do we get them and start using them? Most distributions include both of these packages, making this the easy way to install them. Simply use your distribution's package manager to do the install. For example, in Ubuntu, you would type the following in a terminal window:

sudo apt-get install python-scipy

This installs scipy and all of its dependencies.

If you want to use the latest-and-greatest version and don't want to wait for your distribution to get updated, they are available through Subversion. Simply execute the following:

svn co http://svn.scipy.org/svn/numpy/trunk numpy svn co http://svn.scipy.org/svn/scipy/trunk scipy

Building and installing is handled by a setup.py script in the source directory. For most people, building and installing simply requires:

python setup.py build python setup.py install # done as root

If you don't have root access, or don't want to install into the system package directory, you can install into a different directory using:

python setup.py install --prefix=/path/to/install/dir

Other options also are available, which you can find out about by using:

python setup.py --help-commands

Take time to experiment and see whether you can use any of the extra options in your specific case.

Now that we have scipy and numpy installed, let's begin our tour by looking at some of the basic functions that are often used in scientific calculations. One of the most common tasks is matrix mathematics. This is greatly simplified when you use numpy. The most basic code to do a multiplication of two matrices using numpy would look like this:

import numpy a1=numpy.empty((500,500)) a2=numpy.empty((500,500)) a3=a1*a2

Contrast this to what we would need to write if we did it in C:

#include <stdlib.h> int main() { double a1[500][500]; double a2[500][500]; double a3[500][500]; int i, j, k; for (i=0; i<500; i++) { for (j=0; j<500; j++) { a3[i][j] = 0; for (k=0; k<500; k++) { a3[i][j] += a1[i][k] * a2[k][j]; } } } }

The Python code is much shorter and cleaner, and the intent of the code is much clearer. This kind of clarity in the code means that the programmer can focus much more on the algorithm rather than the gritty details of the implementation. There are C libraries, such as LAPACK, which help simplify this work in C. But, even these libraries can't match the simplicity of scipy.

“But what about efficiency?”, I hear you ask. Well, let's take a look at it with some timed runs. Taking our above example, we can put some calls around the actual matrix multiplication part and see how long each one takes. See Table 1 for the results.

Although your mileage will vary, because these times depend on your hardware and what other programs also are running on your machine, we can see a general trend. The Python code actually was about eight times faster than the C code compiled with no command-line options. That is actually quite surprising. Once we use the optimization command-line option, we see that the C code is now faster, by a factor of approximately 25. So, we can get faster code using optimized C, but we need to realize that multiplying two matrices with 250,000 elements each in one-quarter of a second is probably fast enough.

As well, we get a certain amount of protection when we use Python. What happens if we try to multiply two matrices where such multiplication doesn't make sense mathematically? When we try to multiply two matrices of different sizes, Python gives us:

ValueError: shape mismatch: objects cannot be broadcast to a single shape

In C, we get no error at all. This due to the fact that when we work with matrices, we actually are using pointer arithmetic. So pretty much anything we do is valid C, even if it makes no sense in the problem domain.

We also can work just as easily with complex numbers. If we wanted to create an array of 64-bit complex numbers, we would write:

a=zeros((500,500), dtype=complex64)

Joey Bernard has a background in both physics and computer science. This serves him well in his day job as a computational research consultant at the University of New Brunswick. He also teaches computational physics and parallel programming.

## Trending Topics

## Webinar

### Fast/Flexible Linux OS Recovery

On Demand Now

In this live one-hour webinar, learn how to enhance your existing backup strategies for complete disaster recovery preparedness using Storix System Backup Administrator (SBAdmin), a highly flexible full-system recovery solution for UNIX and Linux systems.

Join *Linux Journal*'s Shawn Powers and David Huffman, President/CEO, Storix, Inc.

Free to *Linux Journal* readers.

July 2016 Issue of Linux Journal | Jul 01, 2016 |

Tibbo Technology's Tibbo Project System | Jul 01, 2016 |

July 2016 Video Preview | Jun 30, 2016 |

Client-Side Performance | Jun 30, 2016 |

Peppermint 7 Released | Jun 29, 2016 |

Libarchive Security Flaw Discovered | Jun 28, 2016 |

- Download "Linux Management with Red Hat Satellite: Measuring Business Impact and ROI"
- Client-Side Performance
- Tibbo Technology's Tibbo Project System
- Sony Settles in Linux Battle
- July 2016 Issue of Linux Journal
- Peppermint 7 Released
- Libarchive Security Flaw Discovered
- The Giant Zero, Part 0.x
- Profiles and RC Files
- Git 2.9 Released

## Geek Guides

With all the industry talk about the benefits of Linux on Power and all the performance advantages offered by its open architecture, you may be considering a move in that direction. If you are thinking about analytics, big data and cloud computing, you would be right to evaluate Power. The idea of using commodity x86 hardware and replacing it every three years is an outdated cost model. It doesn’t consider the total cost of ownership, and it doesn’t consider the advantage of real processing power, high-availability and multithreading like a demon.

This ebook takes a look at some of the practical applications of the Linux on Power platform and ways you might bring all the performance power of this open architecture to bear for your organization. There are no smoke and mirrors here—just hard, cold, empirical evidence provided by independent sources. I also consider some innovative ways Linux on Power will be used in the future.

Get the Guide
## Comments

## The C code

I had the same thought as Rob Hooft about the timing when I read this article, and came here to comment. As I suspected, main() gets optimized away to nothing because the result is never used. I typed in the code and compiled it with gcc 4.3 -O3 -fverbose-asm -S:

objdump -d matmul

00000000004004b0 :

4004b0: 31 c0 xor %eax,%eax

4004b2: c3 retq

4004b3: 90 nop

4004b4: 90 nop

...

(I added a return 0 to main(), hence clearing the return value register with xor.)

0.01 seconds is typical for program that just return true without doing anything at all (other than dynamically linking libc, and other libc initialization, at startup, etc). On my system, (Ubuntu Intrepid with Linux 2.6.27, Ubuntu's "generic" AMD64 kernel image on a C2D: E6600 DDR2-800 g965), the program compiled with gcc -O3 takes 0.001 seconds. The reported time has more to do with kernel time resolution than anything else.

Berthold is also correct, although I didn't know what he meant at first, since I wasn't very familiar with numpy. In numpy, a1 * a2 is an element-wise product: in C:

a3[i][j] = a1[i][j] * a2[i][j];

instead of the inner loop over k. This is not the same as what C = A * B means in usual mathematical notation, hence the confusion. There, it means the matrix product, which is what the C routine calculates, and what dgemm() from BLAS calculates. In numpy, that is numpy.dot(a1, a2).

If you don't want gcc (or any other compiler) to optimize away a calculation, you have to use the result. Either with gcc tricks like __attribute__ ((used)) or by passing the address of the output array to a function that the compiler can't see while it's compiling the function you're testing. Calling a function in a file that you compile separately will defeat any optimizer except e.g. Sun Studio's cross-file optimizer mode that puts extra information in the .o files...

matmul.c:

matmul.py:

results:

gcc 4.3's auto-vectorized (SSE2) version is twice as fast as gcc 3.4's scalar version on a large array like this that doesn't fit in cache (process RSS ~= 450MB). gcc4 vectorizes the loop to process two columns at once, so it only triggers half the cache misses of a direct interpretation of the C code. going down a colum in a row-major array is slow, because each successive element is in a new cache line. Optimized libraries, like ATLAS's BLAS implementation, which numpy uses, can reduce memory access costs to ~n^2 instead of ~n^3, by working in blocks that fit in the cache. GCC's vectorization illustrates that there's a lot to be gained from improved memory access patterns. (it's also a win when you're FPU limited, on small cached arrays, but trust me, it's the memory access that's giving the factor of 2 speedup with an array of 5000. Use oprofile yourself if you want to see.)

So numpy vs. naive C is a huge win if your arrays are not tiny, since even a few decent-sized naive matmuls will dominate your run time regardless of language.

BTW, I used to work at Dalhousie as a Linux cluster sysadmin and all around geek (with the phylogenetics group), where I had heard of ACEnet, where Joey works. Hi!

## Error in exaple

Joey,

Your runtime comparison does compare apple with oranges. The C codes does not compute a1*a2 but numpy.dot(a1, a2) in numpy terms.

Regards

Berthold

## C optimization

Joey,

Thanks for a nice article expressing a lot of my feelings regarding the comparison of C/Fortran and Python.

I do have one remark: your 250000 element matrix multiplication requires 2x500^3 = 250M floating point operations. You are suggesting that the C program performed those in 0.01 second. That requires 25Gflops and a similar amount of integer operations, simultaneously. This is more than most of us have available in their desktops. I think this shows that "gcc -O3" optimized your entire block of code out of the program, leaving just the startup and termination of the program.

In fact, the Python program you show here performs very close to the optimum, since the total interpreter overhead is only a handful of lines of code. It is in small arrays, like 3x3 matrix multiplications, that C can be significantly faster than Python.