The Magic of Numba and Python

Python is a great programming language. It's primary merits are readability and the numerous packages that are available online. However, being an interpreted language the speed of execution is always an issue. Here is a simple example of a piece of code written in Python that tries to add two numbers from a grid.

#!/usr/bin/python

def overlap_pp(x,y):
    count = 0
    for i in range(x):
        for j in range(y):
            count += i + j
    return count

for _ in range(1000):
    q = overlap_pp(500,500)

If you run the above script (saved as n.py) on the command line terminal with the time command you should see some numbers like the below

time ./n.py

real 0m22.379s
user 0m22.363s
sys 0m0.407s

The process running on one processor took about 22 seconds to complete the run. Now lets do something seemingly magical. Lets add a decorator. Decorators are python speak for a mechanism to do wrapper functions. The decorator we are going to add is '@jit'. The code looks like as it is shown below

#!/usr/bin/python

from numba import jit

@jit
def overlap_pp(x,y):
    count = 0
    for i in range(x):
        for j in range(y):
            count += i + j
    return count

for _ in range(1000):
    q = overlap_pp(500,500)

Now let's time this as before.

time ./n.py

real 0m0.420s
user 0m0.370s
sys 0m0.443s

Magic! The exact same code now runs 55x faster! This happens because the Just-In-Time (jit) feature of the numba package compiles the function to machine code on the fly. The timing you see is almost in line with what you would expect from a similar code done in C or Fortran. In fact you will be surprised that in this particular example, the numba version runs faster than the C version! The C code is shown below.

#include < stdio.h >

int overlap_pp(int x,int y){
  int i,j,q;
  for(i=0;i<= x;i++){
    for(j=0;j<= y;j++){
      q += i + j;
    }
  }
  return(q);
}


int main(int argc,char** argv){
  int i,q;
  for(i = 0;i<=1000;i++){
    q = overlap_pp(500,500);
  }
  return 0;
}

The timing results after compiling the above code 'gcc t.c -o t'

time ./t

real 0m0.774s
user 0m0.774s
sys 0m0.000s

Unfortunately there isn't a clear description online on how to get Numba installed on Ubuntu 14.04. After some hacking and poking around stackoverflow the following worked for me.

Step 1:
Numba uses llvm. You need to remove all llvm, llvm-config and llvmlite installations that you already have on your system. Numba needs llvm-3.8 upwards to work properly.

Step 2:
Follow instructions in this stackoverflow thread. But when you get installing do the following
 sudo apt-get install zlib1g zlib1g-dev libedit2 libedit-dev llvm-3.8 llvm-3.8-dev
Notice change in version number to 3.8.

Step 3:
Next, install llvmlite. You also want to create a symbolic link on '/usr/bin/' to point to the latest 3.8 version of llvm-config.
"sudo ln -s /usr/bin/llvm-config-3.8 /usr/bin/llvm-config"
This step is important because the installation of numba appears to use llvm-config. After this you can install Numba directly using pip.

If you are interested in learning more about python programming, data science and probability here are a list of books that are worthwhile to buy.