Efficient use of Python

Ariel Lozano

CÉCI training

November 22, 2017

Outline

I Analyze our code with profiling tools:

I cpu: cProfile, line_profiler, kernprof
I memory: memory_profiler, mprof
I Being a highly abstract dynamically typed language, how to
make a more efficient use of hardware internals?
I Numpy and Scipy ecosystem (mainly wrappers to C/Fortran
compiled code)
I binding to compiled code: interfaces between python and
compiled modules
I compiling: tools to compile python code
I parallelism: modules to exploit multicores
Sieve of eratostenes
Algorithm to find all prime numbers up to any given limit.
Ex: Find all the prime numbers less than or equal to 25:
I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Cross out every number displaced by 2 after 2 up to the limit:
I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Move to next n non crossed, cross out each non crossed
number displaced by n:
I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
The remaining numbers non crossed in the list are all the
primes below limit.
2
√ to n to start crossing out.
Trivial optimization: jump directly
Then, n must loop only up to limit.
Simple python implementation

def primes_upto ( limit ) :

sieve = [ False ] * 2 + [ True ] * ( limit - 1 )
for n in xrange (2 , int ( limit ** 0 . 5 + 1 ) ) :
if sieve [ n ] :
i = n ** 2
while i < limit + 1 :
sieve [ i ] = False
i += n
return [ i for i , prime in enumerate ( sieve ) if prime ]

if __name__ = = " __main__ " :

primes = primes_upto ( 25 )
print ( primes )

$ python sieve01.py

[2, 3, 5, 7, 11, 13, 17, 19, 23]

Measuring running time

Computing primes up to 30 000 000:

I linux time command
$ time python sieve01.py

real 0m10.419s
user 0m10.192s
sys 0m0.217s
I using timeit module to average several runs
$ python -m timeit -n 3 -r 3 -s "import sieve01" \
> "sieve01.primes_upto(30000000)"

3 loops, best of 3: 10.2 sec per loop

CPU profiling: timing functions

cProfile: built-in profiling tool in the standard library. It hooks

into the virtual machine to measure the time taken to run every
function that it sees.
$ python -m cProfile -s cumulative sieve01.py
5 function calls in 10.859 seconds

Ordered by: cumulative time

ncalls tottime percall cumtime percall filename:lineno(function)

1 0.000 0.000 10.859 10.859 {built-in method builtins.exec}
1 0.087 0.087 10.859 10.859 sieve01.py:3(<module>)
1 9.447 9.447 10.772 10.772 sieve01.py:3(primes_upto)
1 1.325 1.325 1.325 1.325 sieve01.py:11(<listcomp>)
1 0.000 0.000 0.000 0.000 {method 'disable' of '_lsprof.Profiler

I Useful information but for big codes we will need extra

tools to visualize the dumps
CPU profiling: line by line details of a function
line_profiler: profiling individual functions on a line-by-line
basis, big overhead introduced. We must add the @profile
decorator on the function to be analyzed.
@profile
def primes_upto ( limit ) :
sieve = [ False ] * 2 + [ True ] * ( limit - 1 )
for n in xrange (2 , int ( limit ** 0 . 5 + 1 ) ) :
if sieve [ n ] :
i = n ** 2
while i < limit + 1 :
sieve [ i ] = False
i += n
return [ i for i , prime in enumerate ( sieve ) if prime ]

if __name__ = = " __main__ " :

primes = primes_upto ( 30000000 )

Then, we run the code with the kernprof.py script provided by

the package.
CPU profiling: line by line details of a function
$ kernprof -l -v sieve01_prof.py
Wrote profile results to sieve01_prof.py.lprof
Timer unit: 1e-06 s

Total time: 101.025 s

File: sieve01_prof.py
Function: primes_upto at line 2

Line # Hits Time Per Hit % Time Line Contents

==============================================================
2 @profile
3 def primes_upto(limit):
4 1 415906 415906.0 0.4 sieve = [False] * 2 + [True] *
5 5477 2307 0.4 0.0 for n in xrange(2, int(limit**0
6 5476 2362 0.4 0.0 if sieve[n]:
7 723 680 0.9 0.0 i = n**2
8 70634832 28740579 0.4 28.4 while i < limit+1:
9 70634109 33142484 0.5 32.8 sieve[i] = False
10 70634109 26776815 0.4 26.5 i += n
11 30000002 11943768 0.4 11.8 return [i for i, prime in enume

% Time is relative to each function only, not to total running time.

Memory profiling: line by line details of a function

memory_profiler: module to measure memory usage on a

line-by-line basis, runs will be slower than line_profiler. Is
also required the @profile decorator on the function to analyze.
$ python -m memory_profiler sieve01_prof.py
Filename: sieve01_prof.py

Line # Mem usage Increment Line Contents

================================================
2 32.715 MiB 0.000 MiB @profile
3 def primes_upto(limit):
4 261.703 MiB 228.988 MiB sieve = [False] * 2 + [True] * (limit - 1)
5 261.703 MiB 0.000 MiB for n in xrange(2, int(limit**0.5 + 1)):
6 261.703 MiB 0.000 MiB if sieve[n]:
7 261.703 MiB 0.000 MiB i = n**2
8 261.703 MiB 0.000 MiB while i < limit+1:
9 261.703 MiB 0.000 MiB sieve[i] = False
10 261.703 MiB 0.000 MiB i += n
11 return [i for i, prime in enumerate(sieve) if
Memory profiling: line by line details of a function

Why are 228 MB allocated on this line?

4 261.703 MiB 228.988 MiB sieve = [False] * 2 + [True] * (limit - 1)

I In a Python list each boolean variable has a size of 8 bytes.

The standard for a C long int in 64-bits.
I We are creating a list with 30000002 elements.
30000002∗8
I Doing the math: 1024∗1024 = 228.881
Memory profiling: analyzing the whole run vs time

I Line by line analysis introduces a huge overhead, can be up

to 100x slower
I We can miss information due to many
allocations/deallocations taking place on a single line
I The memory_profiler package provides the mprof tool to
analyze and visualize the memory usage as a function of
time
I It has a very minor impact on the running time
I Usage:
$ mprof run --python python2 mycode.py
$ mprof plot
Memory profiling: analyzing the whole run vs time
$ mprof run --python python2 sieve01.py
$ mprof plot
Memory profiling: analyzing the whole run vs time

We can add a @profile decorator and timestamps to introduce

details in the analysis.
@profile
def primes_upto ( limit ) :
with profile . timestamp ( " create_sieve_list " ) :
sieve = [ False ] * 2 + [ True ] * ( limit - 1 )
with profile . timestamp ( " cross_out_sieve " ) :
for n in xrange (2 , int ( limit ** 0 . 5 + 1 ) ) :
if sieve [ n ] :
i = n ** 2
while i < limit + 1 :
sieve [ i ] = False
i += n
return [ i for i , prime in enumerate ( sieve ) if prime ]
Memory profiling: analyzing the whole run vs time
$ mprof run --python python2 sieve01_memprof.py
$ mprof plot
Memory profiling: analyzing the whole run vs time
Why the 500 MB peak during the sieve list creation?
I Experimenting with the mprof tool can be verified that:
sieve = [ False ] * 2 + [ True ] * ( limit - 1 )

I is actually equivalent to something like:

sieve1 = [ False ] * 2
sieve2 = [ True ] * ( limit - 1 )
sieve = sieve1 + sieve2
del sieve1
del sieve2

I is allocated temporarily an extra ≈ 30E6 boolean list !!

I We can try to replace with:
sieve = [ True ] * ( limit + 1 )
sieve [ 0 ] = False
sieve [ 1 ] = False
Memory profiling: analyzing the whole run vs time
$ mprof run --python python2 sieve02_memprof.py
$ mprof plot
Numpy library

I Provides a new kind of array datatype

I Contains methods for fast operations on entire arrays
without having to write loops
I They are basically wrappers to compiled C/Fortran/C++
code
I Runs almost as quickly as C
I It is the foundation of many other higher-level numerical
tools
I Compares to MATLAB in functionality
Numpy library: matrix vector product

>>> import numpy as np

>>> a = np . array ( [ [ 5 , 1 ,3 ] ,
[ 1 , 1 ,1 ] ,
[ 1 , 2 ,1 ] ] )
>>> b = np . array ( [1 , 2 , 3 ] )
>>> c = a . dot ( b )
array ( [ 16 , 6 , 8 ] )
Numpy library: sieve revisited

We replace the sieve list with a Numpy boolean array:

import numpy as np

def primes_upto ( limit ) :

sieve = np . ones ( limit + 1 , dtype = np . bool )
sieve [ 0 ] = False
sieve [ 1 ] = False
for n in xrange (2 , int ( limit ** 0 . 5 + 1 ) ) :
if sieve [ n ] :
i = n ** 2
while i < limit + 1 :
sieve [ i ] = False
i += n
return [ i for i , prime in enumerate ( sieve ) if prime ]
Numpy library: sieve revisited

I In a Numpy array each boolean has a size of 1 byte

I Math now: 30000002∗1
1024∗1024 = 28.61
Numpy library: sieve revisited
I Timing did not improve with Numpy array and same loop
I Fully Numpy solution using slice indexing to iterate:

import numpy as np

def primes_upto ( limit ) :

sieve = np . ones ( limit + 1 , dtype = np . bool )
sieve [ 0 ] = False
sieve [ 1 ] = False
for n in xrange (2 , int ( limit ** 0 . 5 + 1 ) ) :
if sieve [ n ] :
sieve [ n * n : : n ] = 0
return np . nonzero ( sieve ) [ 0 ]

$ time python2 sieve04_np.py

real 0m0.552s
user 0m0.518s
sys 0m0.033s

I 22x gain in time!!

Numpy library: sieve line by line profiling
$ kernprof -l -v sieve04_np_prof.py
Wrote profile results to sieve04_np_prof.py.lprof
Timer unit: 1e-06 s

Total time: 0.482723 s

File: sieve04_np_prof.py
Function: primes_upto at line 3

Line # Hits Time Per Hit % Time Line Contents

==============================================================
3 @profile
4 def primes_upto(limit):
5 1 8785 8785.0 1.8 sieve = np.ones(limit
6 1 5 5.0 0.0 sieve[0] = False
7 1 0 0.0 0.0 sieve[1] = False
8 5477 2796 0.5 0.6 for n in xrange(2, int
9 5476 3119 0.6 0.6 if sieve[n]:
10 723 420784 582.0 87.2 sieve[n**2::n]
11 1 47234 47234.0 9.8 return np.nonzero(siev
Numpy library: sieve line by line profiling

I line_profiler helps to understand the massive gain

I Pure python solution:
6 5476 2362 0.4 0.0 if sieve[n]:
7 723 680 0.9 0.0 i = n**2
8 70634832 28740579 0.4 28.4 while i < limit+1:
9 70634109 33142484 0.5 32.8 sieve[i] = False
10 70634109 26776815 0.4 26.5 i += n
I Full Numpy solution:
9 5476 3119 0.6 0.6 if sieve[n]:
10 723 420784 582.0 87.2 sieve[n**2::n] = 0
I The loops to cross out the sieve are fully performed by
lower level implementations in Numpy
I Time and memory usage is the same as C or Fortran
compiled solutions !
CPU and Memory profiling: summary

I Line-by-line profiling introduces a huge overhead, they

must be used reducing the problem size and for specific
functions detected as bottlenecks
I The mprof tool is very dynamic, timestammping in a smart
way can be used both as a fast CPU and Memory profiler
I The cProfile dumps are great to detect bottlenecks on big
projects, but a visualization tool is almost mandatory.
Explore the KCachegrind package, usual workflow:
$ python -m cProfile -o prof.out sieve02.py
$ pyprof2calltree -i prof.out -k
Numpy library: SciPy ecosystem
Collection of open source software for scientific computing in
Python
I Core packages:
I NumPy: the fundamental package for numerical computation
I SciPy library: collection of numerical algorithms and
domain-specific toolboxes, including signal processing, fourier
transforms, clustering, optimization, statistics...
I Matplotlib: a mature plotting package, provides publication-quality
2D plotting as well as rudimentary 3D plotting
I Data and computation:
I pandas: providing high-performance, easy to use data structures
(similar to R)
I SymPy: symbolic mathematics and computer algebra
I scikit-image: algorithms for image processing
I scikit-learn: algorithms and tools for machine learning
I h5py and PyTables: can both access data stored in the HDF5
format
Python Bindings

I Interfacing python with compiled code can provide huge

performance gains
I f2py: project to provide a connection between Python and
Fortran languages
I weave: tools for including C/C++ code within Python code
I cffi (C Foreign Function Interface for Python): Interact with
almost any C code from Python.
I ctypes: foreign function library for Python. It provides C
compatible data types and allows calling functions in DLLs
or shared libraries.
Python Bindings: f2py example

subroutine foo ( a ) import hello

integer a
print * , " Hello from Fortran ! " if __name__ = = " __main__ " :
print * , " a = " ,a
end hello . foo ( 10 )

$ f2py2 -c -m hello hello.f90 $ python2 call_fhello.py

Hello from Fortran!
a= 10
Compiled Python
There are also tools to compile python code
I cython: C-Extensions for Python
I optimising and static compiler
I can compile Python code and Cython language
I can compile Python with Numpy code
I can do bindings with C code
I Pypy: Just-in-time compiler
I sometimes less memory hungry than Cython
I not fully compliant with Python with Numpy code
I Numba: a compiler specialized for numpy code using the
LLVM compiler
I Pythran: compiler for both numpy and non-numpy code.
Takes advantage of multi-cores and single instruction
multiple data (SIMD) units
I All, except pypy requires to modify or decorate the original
python code
Compiled Python: pypy

I We can directly run the original sieve01.py with pypy

$ time pypy sieve01.py

real 0m2.593s
user 0m2.222s
sys 0m0.294s
Parallel processing

I multiprocessing module
I allows to use process- and thread-based parallel processing
I allows to share memory among processes
I constrained to single-machine multicore parallelism
I mpi4py
I Python bindings to the MPI-1/2/3 interfaces
I if you know MPI on C/Fortran you already know mpi4py
I can make use equivalently of multiple cores on a single-machine
or distributed
I each process has a separate address space, no possibility to
share memory between them
I we covered it in the MPI session
Further information on the topic

I High Performance Python by By Micha Gorelick and Ian

Ozsvald
I Python in HPC Tutorial:
https://github.com/pyHPC/pyhpc-tutorial