Python for Scientific and High

Performance Computing
SC09
Portland, Oregon, United States
Monday, November 16, 2009 1:30PM - 5:00PM
http://www.cct.lsu.edu/~wscullin/sc09python/
 Introductions

Your presenters:
William R. Scullin
[email protected]
James B. Snyder
[email protected]
Nick Romero
[email protected]
Massimo Di Pierro
[email protected]
 Overview

We seek to cover:
Python language and interpreter basics
Popular modules and packages for scientific applications
How to improve performance in Python programs
How to visualize and share data using Python
Where to find documentation and resources
Do:
Feel free to interrupt
the slides are a guide - we're only successful if you learn
what you came for; we can go anywhere you'd like
Ask questions
Find us after the tutorial
 About the Tutorial Environment

Updated materials and code samples are available at:

http://www.cct.lsu.edu/~wscullin/sc09python/
we suggest you retrieve them before proceeding. They should
remain posted for at least a calendar year.

You should have login instructions, a username and password

for the tutorial environment on the paper on your slip. Accounts
will be terminated no later than 6:30PM USPT today. Do not
leave any code or data on the system you would like to keep.

Your default environment on the remote system is set up for

this tutorial, though the downloadable live dvd should provide a
comparable environment.
 Outline

1. Introduction 4. Parallel and distributed programming

Introductions 5. Performance
Tutorial overview Best practices for pure Python +
Why Python and why in scientific and NumPy
high performance computing? Optimizing when necessary
Setting up for this tutorial 7. Real world experiences and techniques
2. Python basics 8. Python for plotting, visualization, and
Interpreters data sharing
data types, keywords, and functions Overview of matplotlib
Control Structures Example of MC analysis tool
Exception Handling 9. Where to find other resources
I/O There's a Python BOF!
Modules, Classes and OO 10. Final exercise
3. SciPy and NumPy: fundamentals and 11. Final questions
core components 12. Acknowledgments
Dynamic programming language
Interpreted & interactive
Object-oriented
Strongly introspective
Provides exception-based error handling
Comes with "Batteries included" (extensive standard
libraries)
Easily extended with C, C++, Fortran, etc...
Well documented (http://docs.python.org/)
 Why Use Python for Scientific
Computing?
"Batteries included" + rich scientific computing ecosystem
Good balance between computational performance and
time investment
Similar performance to expensive commercial solutions
Many ways to optimize critical components
Only spend time on speed if really needed
Tools are mostly open source and free (many are MIT/BSD
license)
Strong community and commercial support options.
No license management
 Science Tools for Python
Large number of science-related modules:
General Plotting & Visualization Molecular & Symbolic Math
NumPy matplotlib Atomic Modeling SymPy
SciPy VisIt PyMOL
Chaco Biskit Electromagnetics
GPGPU Computing MayaVi GPAW PyFemax
PyCUDA
PyOpenCL AI & Machine Learning Geosciences Astronomy
pyem GIS Python AstroLib
Parallel Computing ffnet PyClimate PySolar
PETSc pymorph ClimPy
PyMPI Monte CDAT Dynamic Systems
Pypar hcluster Simpy
mpi4py Bayesian Stats PyDSTool
Biology (inc. neuro) PyMC
Wrapping Brian Finite Elements
C/C++/Fortran SloppyCell Optimization SfePy
SWIG NIPY OpenOpt
Cython PySAT
ctypes

For a more complete list: http://www.scipy.org/Topical_Software

Please login to the Tutorial Environment

Let the presenters know if you have any issues.

Start an iPython session:

santaka:~> wscullin$ ipython
Python 2.6.2 (r262:71600, Sep 30 2009, 00:28:07)
[GCC 3.3.3 (SuSE Linux)] on linux2
Type "help", "copyright", "credits" or "license" for more
information.

IPython 0.9.1 -- An enhanced Interactive Python.

? -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help -> Python's own help system.
object? -> Details about 'object'. ?object also works, ?? prints
more.

In [1]:
 Python Basics

Interpreter
Built-in Types, keywords, functions
Control Structures
Exception Handling
I/O
Modules, Classes & OO
 Interpreters

CPython Standard python distribution

What most people think of as "python"
highly portable
http://www.python.org/download/
We're going to use 2.6.2 for this tutorial
The future is 3.x, the future isn't here yet
iPython
A user friendly interface for testing and debugging
http://ipython.scipy.org/moin/
Other Interpreters You Might See...
Unladen Swallow
Blazing fast, uses llvm and in turn may compile!
x86/x86_64 only really
Sponsored by Google
http://code.google.com/p/unladen-swallow/
Jython
Python written in Java and running on the JVM
http://www.jython.org/
performance is about what you expect
IronPython
Python running under .NET
http://www.codeplex.com/IronPython
PyPy
Python in... Python
No where near ready for prime time
http://codespeak.net/pypy/dist/pypy/doc/
CPython Interpreter Notes    

Compilation affects interpreter speed

Distros aim for compatibility and as few irritations as
possible, not performance
compile your own or have your systems admin do it
same note goes for most modules
Regardless of compilation, you'll have the same
bytecode and the same number of instructions
Bytecode is portable, binaries are not
Linking against shared libraries kills portability
Not all modules are available on all platforms
Most are not OS specific, 90% are available everywhere
x86/x86_64 is still better supported than most
A note about distutils and building
modules
Unless your environment is very generic (ie: a major linux
distribution under x86/x86_64), and even if it is, manual
compilation and installation of modules is a very good idea.

Distutils and setuptools often make incorrect assumptions

about your environment in HPC settings. Your presenters
generally regard distutils as evil as they cross-compile a lot.

If you are running on PowerPC, IA-64, Sparc, or in an

uncommon environment, let module authors know you're there
and report problems!
 Built-in Numeric Types
int, float, long, complex - different types of numeric data
>>> a = 1.2 # set a to floating point number
>>> type(a)
<type 'float'>

>>> a = 1 # redefine a as an integer

>>> type(a)
<type 'int'>

>>> a = 1e-10 # redefine a as a float with scientific notation

>>> type(a)
<type 'float'>

>>> a = 1L # redefine a as a long

>>> type(a)
<type 'long'>

>>> a = 1+5j # redefine a as complex

>>> type(a)
<type 'complex'>
Gotchas with Built-in Numeric Types
Python's int and float can become as large in size as your
memory will permit, but ints will be automatically typed as long.
The built-in long datatype is very slow and best avoided.
>>> a=2.0**999
>>> a
5.3575430359313366e+300
>>> import sys
>>> sys.maxint
2147483647
>>> a>sys.maxint
True
>>> a=2.0**9999
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
OverflowError: (34, 'Result too large')
>>> a=2**9999
>>> a-((2**9999)-1)
1L
 Gotchas with Built-in Numeric Types
Python's int and float are not decimal types
IEEE 754 compliant (http://docs.python.org/tutorial/floatingpoint.html)
math with two integers always results in an integer
>>> a=1.0/10
>>> a
0.10000000000000001
>>> a=1/3
>>> a
0
>>> a=1/3.0
>>> a
0.33333333333333331
>>> a=1.0/3.0
>>> a
0.33333333333333331
>>> 0.3
0.29999999999999999
>>> 0.33333333333333333333333333333333
0.33333333333333331
 NumPy Numeric Data Types
NumPy covers all the same numeric data types available in
C/C++ and Fortran as variants of int, float, and complex
all available signed and unsigned as applicable
available in standard lengths
floats are double precision by default
generally available with names similar to C or Fortran
ie: long double is longdouble
generally compatible with Python data types
 Built-in Sequence Types
str, unicode - string types
>>> s = 'asd'
>>> u = u'fgh' # prepend u, gives unicode string
>>> s[1]
's'

list - mutable sequence

>>> l = [1,2,'three'] # make list
>>> type(l[2])
<type 'str'>

>>> l[2] = 3; # set 3rd element to 3

>>> l.append(4) # append 4 to the list

tuple - immutable sequence

>>> t = (1,2,'four')
 Built-in Mapping Type
dict - match any immutable value to an object
>>> d = {'a' : 1, 'b' : 'two'}
>>> d['b'] # use key 'b' to get object 'two'
'two'

# redefine b as a dict with two keys

>>> d['b'] = {'foo' : 128.2, 'bar' : 67.3}
>>> d
{'a': 1, 'b': {'bar': 67.299999999999997, 'foo':
128.19999999999999}}

# index nested dict within dict

>>> d['b']['foo']
128.19999999999999

# any immutable type can be an index

>>> d['b'][(1,2,3)]='numbers'
Built-in Sequence & Mapping Type
Gotchas
Python lacks C/C++ or Fortran style arrays.
Best that can be done is nested lists or dictionaries
Tuples, being immutable are a bad idea
You have to be very careful on how you create them
It does not pre-allocate memory and this can be a serious
source of both annoyance and performance degradation.
NumPy gives us a real array type which is generally a better
choice
 Control Structures
if - compound conditional statement
if (a and b) or (not c):
do something()
elif d:
do_something_else()
else:
print "didn't do anything"

while - conditional loop statement

i = 0
while i < 100:
i += 1
 Control Structures
for - iterative loop statement
for item in list:
do_something_to_item(item)

# start = 0, stop = 10
>>> for element in range(0,10):
... print element,
0 1 2 3 4 5 6 7 8 9

# start = 0, stop = 20, step size = 2

>>> for element in range(0,20,2):
... print element,
0 2 4 6 8 10 12 14 16 18
 Generators

Python makes it very easy to write funtions you can iterate

over- just use yield instead of return at the end of functions
def squares(lastterm):
for n in range(lastterm):
yield n**2

>>> for i in squares(4): print i

...
0
1
4
9
16
 List Comprehensions

List Comprehensions are powerful tool, replacing Python's

lambda function for functional programming
syntax: [f(x) for x in generator]
you can add a conditional if to a list comprehension
>>> [i for i in squares(10)]
[0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
>>> [i for i in squares(10) if i%2==0]
[0, 4, 16, 36, 64]
>>> [i for i in squares(10) if i%2==0 and i%3==1]
[4, 16, 64]
 Exception Handling
try - compound error handling statement
>>> 1/0
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ZeroDivisionError: integer division or modulo by
zero

>>> try:
... 1/0
... except ZeroDivisionError:
... print "Oops! divide by zero!"
... except:
... print "some other exception!"

Oops! divide by zero!

 File I/O Basics
Most I/O in Python follows the model laid out for file I/O and
should be familiar to C/C++ programmers.
basic built-in file i/o calls include
open(), close()
write(), writeline(), writelines()
read(), readline(), readlines()
flush()
seek() and tell()
fileno()
basic i/o supports both text and binary files
POSIX like features are available via fctl and os
modules
be a good citizen
if you open, close your descriptors
if you lock, unlock when done
 Basic I/O examples
>>> f=file.open('myfile.txt','r')
# opens a text file for reading with default buffering
for writing use 'w'
for simultaneous reading and writing add '+' to either 'r' or
'w'
for appending use 'a'
to do binary files add 'b'
>>> f=file.open('myfile.txt','w+',0)
# opens a text file for reading and writing with no buffering.
a 1 means line buffering,
other values are interpreted as buffer sizes in bytes
Let's write ten integers to disk without buffering, then read them
back:
>>> f=open('frogs.dat','w+',0) # open for unbuffered reading and writing
>>> f.writelines([str(my_int) for my_int in range(10)])
>>> f.tell() # we're about to see we've made a mistake
10L # hmm... we seem short on stuff
>>> f.seek(0) # go back to the start of the file
>>> f.tell() # make sure we're there
0L
>>> f.readlines() # Let's see what's written on each line
['0123456789']# we've written 10 chars, no line returns... oops
>>> f.seek(0) # jumping back to start, let's add line returns
>>> f.writelines([str(my_int)+'\n' for my_int in range(10)])
>>> f.tell() # jumping back to start, let's add line returns
20L
>>> f.seek(0)# return to start of the file
>>> f.readline()# grab one line
'0\n'
>>>f.next() # grab what ever comes next
'1\n'
>>> f.readlines() # read all remaining lines in the file
['2\n', '3\n', '4\n', '5\n', '6\n', '7\n', '8\n', '9\n']
>>> f.close() # always clean up after yourself - no need other than courtesy!
 I/O for scientific formats

i/o is relatively weak out of the box - luckily there are the
following alternatives:
h5py
Python bindings for HDF5
http://code.google.com/p/h5py/
netCDF4
Python bindings for NetCDF
http://netcdf4-python.googlecode.
com/svn/trunk/docs/netCDF4-module.html
mpi4py allows for classic MPI-IO via MPI.File
 Modules
import - load module, define in namespace
>>> import random # import module
>>> random.random() # execute module method
0.82585453878964787

>>> import random as rd # import with name

>>> rd.random()
0.22715542164248681

# bring randint into namespace

>>> from random import randint
>>> randint(0,10)
4
 Classes & Object Orientation
>>> class SomeClass:
... """A simple example class""" # docstring
... pi = 3.14159 # attribute
... def __init__(self, ival=89): # init w/ default
... self.i = ival
... def f(self): # class method
... return 'Hello'
>>> c = SomeClass(42) # instantiate
>>> c.f() # call class method
'hello'

>>> c.pi = 3 # change attribute

>>> print c.i # print attribute

42
N-dimensional homogeneous arrays (ndarray)
Universal functions (ufunc)
built-in linear algebra, FFT, PRNGs
Tools for integrating with C/C++/Fortran
Heavy lifting done by optimized C/Fortran libraries
ATLAS or MKL, UMFPACK, FFTW, etc...
    Creating NumPy Arrays

# Initialize with lists: array with 2 rows, 4 cols

>>> import numpy as np
>>> np.array([[1,2,3,4],[8,7,6,5]])
array([[1, 2, 3, 4],
[8, 7, 6, 5]])

# Make array of evenly spaced numbers over an interval

>>> np.linspace(1,100,10)
array([ 1., 12., 23., 34., 45., 56., 67., 78., 89., 100.])

# Create and prepopulate with zeros

>>> np.zeros((2,5))
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
    Slicing Arrays
>>> a = np.array([[1,2,3,4],[9,8,7,6],[1,6,5,4]])
>>> arow = a[0,:] # get slice referencing row zero
>>> arow
array([1, 2, 3, 4])

>>> cols = a[:,[0,2]] # get slice referencing columns 0 and 2

>>> cols
array([[1, 3],
[9, 7],
[1, 5]])

# NOTE: arow & cols are NOT copies, they point to the original data
>>> arow[:] = 0
>>> arow
array([0, 0, 0, 0])
>>> a
array([[0, 0, 0, 0],
[9, 8, 7, 6],
[1, 6, 5, 4]])

# Copy data
>>> copyrow = arow.copy()
 Broadcasting with ufuncs
apply operations to many elements with a single call
>>> a = np.array(([1,2,3,4],[8,7,6,5]))
>>> a
array([[1, 2, 3, 4],
[8, 7, 6, 5]])

# Rule 1: Dimensions of one may be prepended to either array

>>> a + 1 # add 1 to each element in array
array([[2, 3, 4, 5],
[9, 8, 7, 6]])

# Rule 2: Arrays may be repeated along dimensions of length 1

>>> a + array(([1],[10])) # add 1 to 1st row, 10 to 2nd row
>>> a + ([1],[10]) # same as above
>>> a + [[1],[10]] # same as above
array([[ 2, 3, 4, 5],
[18, 17, 16, 15]])
>>> a**([2],[3]) # raise 1st row to power 2, 2nd to 3
array([[ 1, 4, 9, 16],
[512, 343, 216, 125]])
 SciPy

Extends NumPy with common scientific computing tools

optimization, additional linear algebra, integration,
interpolation, FFT, signal and image processing, ODE
solvers
Heavy lifting done by C/Fortran code
 Parallel & Distributed Programming

threading
useful for certain concurrency issues, not usable for parallel
computing due to Global Interpreter Lock (GIL)

subprocess
relatively low level control for spawning and managing
processes

multiprocessing - multiple Python instances (processes)

basic, clean multiple process parallelism

MPI
mpi4py exposes your full local MPI API within Python
as scalable as your local MPI
Python Threading

Python threads
real POSIX threads
share memory and state with their parent processes
do not use IPC or message passing
light weight
generally improve latency and throughput
there's a heck of a catch, one that kills performance...
 The Infamous GIL

To keep memory coherent, Python only allows a single thread

to run in the interpreter's space at once. This is enforced by the
Global Interpreter Lock, or GIL. It also kills performance for
most serious workloads. Unladen Swallow may get rid of the
GIL, but it's in CPython to stay for the foreseeable future.

It's not all bad, the GIL:

Is mostly sidestepped for I/O (files and sockets)
Makes writing modules in C much easier
Makes maintaining the interpreter much easier
Makes for any easy target of abuse
Gives people an excuse to write competing threading
modules (please don't)
Implementation Example: Calculating Pi

Generate random points inside a square

Identify fraction (f) that fall inside a circle with radius equal
to box width
x2 + y2 < r
Area of quarter of circle (A) = pi*r2 / 4
Area of square (B) = r2
A/B = f = pi/4
pi = 4f
 Calculating pi with threads
from threading import Thread
from Queue import Queue, Empty
import random

def calcInside(nsamples,rank):
global inside #we need something everyone can share
random.seed(rank)
for i in range(nsamples):
x = random.random();
y = random.random();
if (x*x)+(y*y)<1:
inside += 1

if __name__ == '__main__':
nt=4 # thread count
inside = 0 #you need to initialize this
samples=100000
threads=[Thread(target=calcInside, args=(samples/nt,i)) for i in range
(nt)]

for t in threads:
t.start()

for t in threads:
t.join()

print (inside*4.0)/samples
 Subprocess
The subprocess module allows the Python interpreter to
spawn and control processes. It is unaffected by the GIL. Using
the subprocess.Popen() call, one may start any process
you'd like.

>>> pi=subprocess.Popen('python -c "import math; print

math.pi"',shell=True,stdout=subprocess.PIPE)
>>> pi.stdout.read()
'3.14159265359\n'
>>> pi.pid
1797
>>> me.wait()
0

It goes without saying, there's better ways to do

subprocesses...
 Multiprocessing

Added in Python 2.6

Faster than threads as the GIL is sidestepped
uses subprocesses
both local and remote subprocesses are supported
shared memory between subprocesses is risky
no coherent types
Array and Value are built in
others via multiprocessing.sharedctypes
IPC via pipes and queues
pipes are not entirely safe
synchronization via locks
Manager allows for safe distributed sharing, but it's slower
than shared memory
 Calculating pi with multiprocessing
import multiprocessing as mp
import numpy as np
import random

processes = mp.cpu_count()
nsamples = 120000/processes

def calcInside(rank):
inside = 0
random.seed(rank)
for i in range(nsamples):
x = random.random();
y = random.random();
if (x*x)+(y*y)<1:
inside += 1
return (4.0*inside)/nsamples

if __name__ == '__main__':
pool = mp.Pool(processes)
result = pool.map(calcInside, range(processes))
print np.mean(result)
 pi with multiprocessing, optimized
import multiprocessing as mp
import numpy as np

processes = mp.cpu_count()
nsamples = 120000/processes

def calcInsideNumPy(rank):
np.random.seed(rank)
xy = np.random.random((nsamples,2))**2 # "vectorized" sample gen
return 4.0*np.sum(np.sum(xy,1)<1)/nsamples

if __name__ == '__main__':
pool = mp.Pool(processes)
result = pool.map(calcInsideNumPy, range(processes))
print np.mean(result)
mpi4py

wraps your native mpi

prefers MPI2, but can work with MPI1
works best with NumPy data types, but can pass around
any serializable object
provides all MPI2 features
well maintained
distributed with Enthought's SciPy
requires NumPy
portable and scalable
http://mpi4py.scipy.org/
How mpi4py works...

mpi4py jobs must be launched with mpirun

each rank launches its own independent python interpreter
each interpreter only has access to files and libraries
available locally to it, unless distributed to the ranks
communication is handled by your MPI layer
any function outside of an if block specifying a rank is
assumed to be global
any limitations of your local MPI are present in mpi4py
 Calculating pi with mpi4py
from mpi4py import MPI
import numpy as np
import random

comm = MPI.COMM_WORLD
rank = comm.Get_rank()
mpisize = comm.Get_size()
nsamples = 120000/mpisize

inside = 0
random.seed(rank)
for i in range(nsamples):
x = random.random();
y = random.random();
if (x*x)+(y*y)<1:
inside += 1

mypi = (4.0 * inside)/nsamples

pi = comm.reduce(mypi, op=MPI.SUM, root=0)

if rank==0:
print (1.0 / mpisize)*pi
 Performance

Best practices with pure Python & NumPy

Optimization where needed (we'll talk about this in GPAW)
profiling
inlining
Other avenues
 Python Best Practices for Performance

If at all possible...
Don't reinvent the wheel.
someone has probably already done a better job than
your first (and probably second) attempt
Build your own modules against optimized libraries
ESSL, ATLAS, FFTW, PyCUDA, PyOpenCL
Use NumPy data types instead of Python ones
Use NumPy functions instead of Python ones
"vectorize" operations on >1D data types.
avoid for loops, use single-shot operations
Pre-allocate arrays instead of repeated concatenation
use numpy.zeros, numpy.empty, etc..
Real-World Examples and Techniques:
GPAW
 Outline

a massively parallel Python-C code for KS-DFT

Introduction
NumPy
Memory
FLOPs
Parallel-Python Interpreter and Debugging
Profiling mixed Python-C code
Python Interface BLACS and ScaLAPACK
Concluding remarks
 Overview
GPAW is an implementation of the projector augmented
wave method (PAW) method for Kohn-Sham (KS) -
Density Functional Theory (DFT)
Mean-field approach to Schrodinger equation
Uniform real-space grid, multiple levels of parallelization
Non-linear sparse eigenvalue problem
10^6 grid points, 10^3 eigenvalues
Solved self-consistently using RMM-DIIS
Nobel prize in Chemistry to Walter Kohn (1998) for (KS)-DFT
Ab initio atomistic simulation for predicting material
properties
Massively parallel and written in Python-C using the
NumPy library.
GPAW Strong-scaling Results
 GPAW code structure

Not simply a Python wrapper on legacy Fortran/C code

Python for coding the high-level algorithm, lots of OO
C for coding numerical intense operations, built on NumPy
Use BLAS and LAPACK whenever possible

Here is some pseudo code for iterative eigensolver:

for i in xrange(max SCF):
for n in xrange(number of bands):
R_ng = apply(H_gg,Psi_ng) % Compute residuals
rk(1.0, R_ng, 0.0, H_mn) % construct Hamiltonian

KS-DFT algorithms are well-known and computationally

intensive parts are known a priori.
KS-DFT is a complex algorithm!
 Source Code History

Mostly Python-code, 10% C-code.

90% of wall-clock time spend in C, BLAS, and LAPACK.
 Performance Mantra

People are able to code complex algorithms in much less time

by using a high-level language like Python (e.g., also
C++). There can be a performance penalty in the most pure
sense of the term.

"The best performance improvement is the transition from the

nonworking to the working state."
--John Ousterhout

"Premature optimization is the root of all evil."

--Donald Knuth

"You can always optimize it later."

-- Unknown
 NumPy - Memory

Q: Where is all my memory going?

A: It disappears a little bit at a time.
BlueGene/P has 512 MB per core.
Compute note kernel is another 34 MB.
NumPy library is ~ 38 MB.
Python Interpreter 12 MB.
Can't always get the last 50 MB, NumPy to blame?
Try this simple test:
import numpy as np
A = np.zero((N,N),dtype=float)
Beware of temporary matrices, they are sometimes not
obvious
D = np.dot(A,np.dot(B,C))
 NumPy - Memory

How big is your binary? Find out using 'size <binary>'

GPAW is 70 MB without all the dynamic libraries.
Only 325 MB of memory left on BG/P per core for
calculation!

FLOPS are cheap, memory and bandwidth are expensive!

Future supercomputers will have low memory per core.

 NumPy - FLOPS

Optimized BLAS available via NumPy via np.dot. Handles

general inner product of multi-dimensional arrays.
Very difficult to cross-compile on BG/P. Blame disutils!
core/_dotblas.so is a sign of optimized NumPy dot
Python wrapper overhead is negligible
For matrix * vector products, NumPy dot can yield better
performance than direct call to GEMV!
Fused floating-point multiply-add instructions are not
created for AXPY type operation in pure Python. Not
available in NumPy either.

for i in xrange(N):
Y[i] += alpha*X[i]
C[i] += A[i]*B[i]
 NumPy - FLOPS

WARNING: If you make heavy, use of BLAS & LAPACK type

operations.
Plan on investing a significant amount of time working to
cross-compile optimized NumPy.
Safest thing is to write your own C-wrappers.
If all your NumPy arrays are < 2-dimensional, Python
wrappers will be simple.
Wrappers for multi-dimensional arrays can be challenging:
SCAL, AXPY is simple
GEMV more difficulty
GEMM non-trivial
Remember C & NumPy arrays are row-ordered, Fortran
arrays are column-ordered!
 Python BLAS Interface
void dscal_(int*n, double* alpha, double* x, int* incx); % C prototype for Fortran
void zscal_(int*n, void* alpha, void* x, int* incx); % C prototype for Fortran
#define DOUBLEP(a) ((double*)((a)->data)) % Casting for NumPy data struc.
#define COMPLEXP(a) ((double_complex*)((a)->data)) % Casting for NumPy data struc.

PyObject* scal(PyObject *self, PyObject *args)

{
Py_complex alpha;
PyArrayObject* x;
if (!PyArg_ParseTuple(args, "DO", &alpha, &x))
return NULL;
int n = x->dimensions[0];
for (int d = 1; d < x->nd; d++) % NumPy arrays can be multi-dimensional!
n *= x->dimensions[d];
int incx = 1;

if (x->descr->type_num == PyArray_DOUBLE)
dscal_(&n, &(alpha.real), DOUBLEP(x), &incx);
else
zscal_(&n, &alpha, (void*)COMPLEXP(x), &incx);
Py_RETURN_NONE;
}
 Profiling Mixed Python-C code

Number of profiling tools available:

gprof, CrayPAT, import profile
gprof, CrayPAT - C, Fortran
import profile - Python
TAU, http://www.cs.uoregon.edu/research/tau/home.php
Exclusive time for C, Python, MPI are reported
Communication matrix available
Interfaces with PAPI for performance counters
Manual and automatic instrumentation available
Installation is doable, but can be challenging
Finding performance bottlenecks is critical to scalability on
HPC platforms
Parallel Python Interpreter and
Debugging
 Parallel Python Interpreter and
Debugging
MPI-enabled "embedded" Python Interpreter:

int main(int argc, char **argv)

{
int status;
MPI_Init(&argc, &argv); % backwards compatible with MPI-1
Py_Initialize(); % needed because of call in next line
PyObect* m = Py_InitModule3("_gpaw", functions,
"C-extension for GPAW\n\n...
\n");
import_array1(-1); % needed for NumPy C-API
MPI_Barrier(MPI_COMM_WORLD); % sync up
status = Py_Main(argc, argv); % call to Python Interpreter
MPI_Finalize();
return status;
}
 Parallel Python Interpreter and
Debugging
Errors in Python modules are OK, core dumps in C extensions
are problematic:
Python call stack is hidden; this is due to Python's
interpreted nature.
Totalview won't help, sorry.
 Profiling Mixed Python-C code

Flat profile shows time spent in Python, C, and MPI simultaneously:

 Profiling Mixed Python-C code

Measure heap memory on subroutine entry/exit:

 Python Interface to BLACS and ScaLAPACK

Parallel dense linear algebra needed for KS-DFT. As the matrix

size N grows, this operation cannot be performed in serial.
(approx. cross over point)
symmetric diagonalize (N > 1000)
symmetric general diagonalize (N > 1000)
inverse Cholesky (N > 4000)

There is no parallel dense linear algebra in NumPy, there are

some options:
PyACTS, based on Numeric
GAiN, Global Arrays based on NumPy (very new)
Write your own Python interface to ScaLAPACK.
 Python Interface to BLACS and ScaLAPACK

Mostly non-Python related challenges:

Good ScaLAPACK examples are few and far apart.
DFT leads to wieldy use of ScaLAPACK
H_mn is on a small subset of MPI_COMM_WORLD
ScaLAPACK does not create a distribute object for you.
Array must be created in a distributed manner
ScaLAPACK allows you to manipulate them via
descriptors
Array must be compatible with their native 2D-block
cyclic layout
What language was used to write ScaLAPACK and
BLACS?
C and Fortran
Distributed arrays assumed to be Fortran-ordered.
Python Interface to BLACS and ScaLAPACK

MPI_COMM_WORLD on a 512-node on 8x8x8 BG/P.

2048 cores!
 Python Interface to BLACS and ScaLAPACK

Physical 1D layout (left) of H_mn, need to redistribute to 2D

block-cyclic layout (right) for use with ScaLAPACK.
Python Interface to BLACS and ScaLAPACK

BLACS: ScaLAPACK:
Cblacs_gridexit numroc
Cblacs_gridinfo Cpdgem2d
Cblacs_gridinit Cpzgemm2d
Cblacs_pinfo Cpdgemr2do
Csys2blacs_handle Cpzgemr2do

Python:
blacs_create
blacs_destroy
blacs_redist
 Python Interface to BLACS and ScaLAPACK

Important to understand notion of array descriptor:

Distinguish between global and local array. Create local
array in parallel.
int desc[9];
desc[0] = BLOCK_CYCLIC_2D; % MACRO = 1
desc[1] = ConTxt; % must equal -1 if inactive
desc[2] = m; % number of rows in global array
desc[3] = n; % number of columns in global array
desc[4] = mb; % row blocksize
desc[5] = nb; % column blocksize
desc[6] = irsrc; % starting row
desc[7] = icsrc; % starting column
desc[8] = MAX(0, lld); % leading dimension of local array
 Python Interface to BLACS and ScaLAPACK
Descriptor only missing ConTxt and LLD, all else from inputs!
Pyobject* blacs_create(PyObject *self, PyObject *args)
{
if (!PyArg_ParseTuple(args, "Oiiiiii|c", &comm_obj, &m, &n,
&nprow,
&npco
l, &mb, &nb, &order))
return NULL;
if (comm_obj == Py_None) % checks for MPI_COMM_NULL
create inactive descriptor
else {
MPI_Comm comm = ((MPIObject*)comm_obj->comm;
ConTxt = Csys2blacs_handle(comm); % ConTxt out of Comm
Cblacs_gridinit(&ConTxt, &order, nprow, npcol);
Cblacs_gridinfo(ConTxt, &nprow, &npcol, &myrow, &mycol);
LLD = numroc(&m, &mb, &mrow, &rsrc, &nprow);
create active descriptor
}
return desc as PyObject
}
 Python Interface in BLACS and ScaLAPACK
PyObject* scalapack_redist(PyObject *self, PyObject *args)
{
PyArrayObject* a_obj; % source matrix
PyArrayObject* b_obj; % destination matrix
PyArrayObject* adesc; % source descriptor
PyArrayObject* bdesc; % destination descriptor
PyObject* comm_obj = Py_None; % intermediate communicator

if (!PyArg_ParseTuple(args, "OOOi|Oii", &a_obj, &adesc,

&bdesc, &isreal, &comm_obj, &m, &n))
return NULL;

% Get info about source (a_ConTxt) and destination (b_ConTxt) grid;

Cblacs_gridinfo_(a_ConTxt, &a_nprow, &a_npcol,&a_myrow,
&a_mycol);
Cblacs_gridinfo_(b_ConTxt, &b_nprow, &b_npcol,&b_myrow,
&b_mycol);
% Size of local destination array
int b_locM = numroc_(&b_m, &b_mb, &b_myrow, &b_rsrc, &b_nprow);
int b_locN = numroc_(&b_n, &b_nb, &b_mycol, &b_csrc, &b_npcol);
 Python Interface in BLACS and ScaLAPACK

% Size of local destination array

int b_locM = numroc_(&b_m, &b_mb, &b_myrow, &b_rsrc, &b_nprow); int
b_locN = numroc_(&b_n, &b_nb, &b_mycol, &b_csrc, &b_npcol);

% Create Fortran-ordered NumPy array

npy_intp b_dims[2] = {b_locM, b_locN);
if (isreal)
b_obj = (PyArrayObject*)PyArray_EMPTY(2, b_dims, NPY_DOUBLE,
NPY_F_CONTIGUOUS);
else
b_obj = (PyArrayObject*)PyArray_EMPTY(2, b_dims,
NPY_DOUBLE, NPY_F_CONTIGUOUS);
 Python Interface in BLACS and ScaLAPACK
if (comm_obj = Py_None) % intermediate communicator is world
{
if(isreal)
Cpdgemr2do_(m, n, DOUBLEP(a_obj), one, one, INTP(adesc),
DOUBLEP(b_obj), one, one, INTP
(bdesc));
else
Cpzgem2rdo_(m, n, (void*)COMPLEXP(a_obj), one, one,
INTP(adesc), (void*)COMPLEXP(b_obj),
one, one,
INTP(bdesc));
}
else % intermediate communicator is user-defined
<create intermediate blacs grid on another communicator>
if(isreal)
Cpdgemr2d(.....);
else
Cpzgemr2d(......);
}
 Python Interface in BLACS and ScaLAPACK

Source blacs grid (blue) and destination blacs grid (red).

Intermediate blacs grid for scalapack_redist:
Must call Cp[d/z]gemr2d[o]
Must encompass both source and destination
For multiple concurrent redist operations, intermediate
cannot overlap.
 Python Interface in BLACS and ScaLAPACK

% Note that we choose to return Py_None, instead of empty array.

if ((b_locM == 0) | (b_locN == 0))
{
Py_DECREF(b_obj);
Py_RETURN_NONE;
}

PyObject* value = PyBuildValue("O",b_obj);

Py_DECREF(b_obj);
return value;
} % end of scalapack_redist

More information at:

https://trac.fysik.dtu.dk/projects/gpaw/browser/trunk/c/blacs.c
 Summary

The Bad & Ugly: The Good:

NumPy cross-compile. GPAW has an extraordinary amount
C Python extensions require of functionality and scalibity. A lot of
learning NumPy & C API. features make coding complex
Debugging C extensions can be algorithms easy:
difficult. OOP
Performance analysis will weakly-typed data structures
always be needed. Interface with many things other
OpenMP-like threading not languages: C, C++, Fortran, etc.
available due to GIL.
Python will need to support
GPU acceleration in the future.
 Success Story

GPAW will allow you to run multiple concurrent DFT calculations with
a single executable.
High-throughput computing (HTC) for catalytic materials screening.
Perform compositional sweeps trivially.
Manage the dispatch of many tasks without 3rd party software
Suppose 512-node partition, 2048 cores
Each DFT calculation requires 128 cores, no guarantee that
they all finish at the same time
Set-up for N >> 2048/128 calculations. As soon as one DFT
calculations finish, start another one until the job runs out of
wall-clock time.
Python for plotting and visualization

Overview of matplotlib
Example of MC analysis tool written in Python
Looking at data sharing on the web
 From a Scientific Library
To a Scientific Application

Massimo Di Pierro
From Lib to App
(overview)

Numerical Algorithms
 From Lib to App
(overview)

Storage Numerical Algorithms

Store and retrieve information in a relational database

 From Lib to App
(overview)

Storage Numerical Algorithms Interface

Plotting

Store and retrieve information in a relational database

Provide a user interface
input forms with input validation
represent data (html, xml, csv, json, xls, pdf, rss)
represent data graphically
 From Lib to App
(overview)
user
internet
Storage Numerical Algorithms Interface user

Plotting
user

Store and retrieve information in a relational database

Provide a user interface
input forms with input validation
represent data (html, xml, csv, json, xls, pdf, rss)
represent data graphically
Communicate with users over the internet
provide user authentication/authorization/access control
provide persistence (cookies, sessions, cache)
log activity and errors
protect security of data
 How? Use a framework!

gnuplot.py
Ruby on Rails
r.py
Django
Chaco
TurboGears
Dislin
Pylons ...
... matplotlib
web2py
 Why?

web2py is really easy to use

web2py is really powerful and does a lot
for you
web2py is really fast and scalable for
production jobs
I made web2py so I know it best
matplotlib is the best library for plotting I
have ever seen (not just in Python)
matplotlib gallery
 web2py and MVC

code project
 web2py and MVC

code project
application1

application
2

application
3
 web2py and MVC
Models Controllers Views

code project
application
1

application=”
dna”

application
3
Data Data
Logic/Workflow
representation presentation
 web2py and MVC
Models Controllers Views

code project
application
1

<h1>
db.define_table( def upload_dna():
Upload DNA Seq. Minimal
‘dna’, return dict(form=
application=” Field(‘sequence’)) crud.create(db.dna))
</h1> Complete
dna” {{=form}} Application

application
3
Data Data
Logic/Workflow
representation presentation
 web2py and Dispatching
<h1>
Upload DNA Seq.
</h1>

{{=form}}
web2py and Dispatching
hostnam
e
web2py and Dispatching

app name
web2py and Dispatching
controller
web2py and Dispatching
action
name
 web2py and Views
<h1>
Upload DNA Seq.
</h1>

{{=form}}
 web2py and Views
<h1>
Upload DNA Seq.
</h1>

{{=form}}

{{=form}}
 web2py and Authentication

authenticatio
n
web2py and AppAdmin

database interface
 web2py web based IDE
web based IDE
 Goal

to build a web based application

that stores DNA sequences
allows upload of DNA sequences
allows analysis of DNA sequences
(reverse, count, align, etc.)

allows plotting of results

 Before we start

download web2py from web2py.com

unzip web2py and click on the executable
when it asks for a password choose one
visit http://127.0.0.1:8000/admin and login
create a new “dna” application by:
type “dna” in the apposite box and press [submit]
 Define model
in models/db_dna.py

import math, random, uuid, re

db.define_table('dna',
Field('name'),
Field('sequence','text'))

def random_gene(n):
return ''.join(['ATGC'[int(n+10*math.sin(n*k)) % 4] \
for k in range(10+n)])+'UAA'

def random_dna():
return ''.join([random_gene(random.randint(0,10)) \
for k in range(50)])

if not db(db.dna.id>0).count():
for k in range(100):
db.dna.insert(name=uuid.uuid4(),sequence=random_dna())
 Define some algorithms

def find_gene_size(a):
r=re.compile('(UAA|UAG|UGA)(?P<gene>.*?)(UAA|UAG|UGA)')
return [(g.start(),len(g.group('gene'))) \
for g in r.finditer(a)]

def needleman_wunsch(a,b,p=0.97):
"""Needleman-Wunsch and Smith-Waterman"""
z=[]
for i,r in enumerate(a):
z.append([])
for j,c in enumerate(b):
if r==c:
z[-1].append(z[i-1][j-1]+1 if i*j>0 else 1)
else:
z[-1].append(p*max(z[i-1][j] if i>0 else 0,
z[i][j-1] if j>0 else 0))
return z
 in models/matplotlib_helpers.py

import random, cStringIO

from matplotlib.backends.backend_agg import FigureCanvasAgg as FigureCanvas
from matplotlib.figure import Figure

def plot(title='title',xlab='x',ylab='y',data={}):
fig=Figure()
fig.set_facecolor('white')
ax=fig.add_subplot(111)
if title: ax.set_title(title)
if xlab: ax.set_xlabel(xlab)
if ylab: ax.set_ylabel(ylab)
legend=[]
keys=sorted(data)
for key in keys:
stream = data[key]
(x,y)=([],[])
for point in stream:
x.append(point[0])
y.append(point[1])
ell=ax.hist(y,20)
canvas=FigureCanvas(fig)
response.headers['Content-Type']='image/png'
stream=cStringIO.StringIO()
canvas.print_png(stream)
return stream.getvalue()
 Define actions
in controllers/default.py

def index():
rows=db(db.dna.id).select(db.dna.id,db.dna.name)
return dict(rows=rows)

@auth.requires_login()
def gene_size():
dna = db.dna[request.args(0)] or \
redirect(URL(r=request,f='index'))
lengths = find_gene_size(dna.sequence)
return hist(data={'Lengths':lengths})
 Define Views
in views/default/index.html

{{extend 'layout.html'}}

<a href="{{=URL(r=request,f='compare')}}">compare</a>

<ul>
{{for row in rows:}}
<li>{{=row.name}}
[<a href="{{=URL(r=request,f='gene_size',args=row.id)}}">gene sizes</a>]
</li>
{{pass}}
</ul>
Try it
 in models/matplotlib_helpers.py

def pcolor2d(title='title',xlab='x',ylab='y',
z=[[1,2,3,4],[2,3,4,5],[3,4,5,6],[4,5,6,7]]):
fig=Figure()
fig.set_facecolor('white')
ax=fig.add_subplot(111)
if title: ax.set_title(title)
if xlab: ax.set_xlabel(xlab)
if ylab: ax.set_ylabel(ylab)
image=ax.imshow(z)
image.set_interpolation('bilinear')
canvas=FigureCanvas(fig)
response.headers['Content-Type']='image/png'
stream=cStringIO.StringIO()
canvas.print_png(stream)
return stream.getvalue()
 Define Actions
in controllers/default.py

def needleman_wunsch_plot():
dna1 = db.dna[request.vars.sequence1]
dna2 = db.dna[request.vars.sequence2]
z = needleman_wunsch(dna1.sequence,dna2.sequence)
return pcolor2d(z=z)

def compare():
form = SQLFORM.factory(
Field('sequence1',db.dna,
requires=IS_IN_DB(db,'dna.id','%(name)s')),
Field('sequence2',db.dna,
requires=IS_IN_DB(db,'dna.id','%(name)s')))
if form.accepts(request.vars):
image=URL(r=request,f='needleman_wunsch_plot',
vars=form.vars)
else:
image=None
return dict(form=form, image=image)
 Define Views
in views/default/compare.html

{{extend 'layout.html'}}

{{=form}}

{{if image:}}
Sequence1 = {{=db.dna[request.vars.sequence1].name}}<br/>
Sequence2 = {{=db.dna[request.vars.sequence2].name}}<br/>

<img src="{{=image}}" alt="loading..."/>

{{pass}}
Try it
Resources

Python
http://www.python.org/
all the current documentation, software, tutorials, news, and pointers to advice
you'll ever need
GPAW
https://wiki.fysik.dtu.dk/gpaw/
GPAW documentation and code
SciPy and NumPy
http://numpy.scipy.org/
The official NumPy website
http://conference.scipy.org/
The annual SciPy conference
http://www.enthought.com/
Enthought, Inc. the commercial sponsors of SciPy, NumPy, Chaco, EPD and
more
Matplotlib
http://matplotlib.sourceforge.net/
best 2D package on the planet
mpi4py
http://mpi4py.scipy.org/
Yet More Resources

Tau
http://www.cs.uoregon.edu/research/tau/home.php
official open source site
http://www.paratools.com/index.php
commercial tools and support for Tau
web2py
http://www.web2py.com/
web framework used in this tutorial
Hey! There's a Python BOF
Python for High Performance and Scientific Computing
Primary Session Leader:
Andreas Schreiber (German Aerospace Center)

Secondary Session Leaders:

William R. Scullin (Argonne National Laboratory) Steven
Brandt (Louisiana State University) James B. Snyder (Northwestern
University) Nichols A. Romero (Argonne National Laboratory)
Birds-of-a-Feather Session
Wednesday, 05:30PM - 07:00PM Room A103-104
Abstract:
The Python for High Performance and Scientific Computing BOF is intended to
provide current and potential Python users and tool providers in the high
performance and scientific computing communities a forum to talk about their
current projects; ask questions of experts; explore methodologies; delve into
issues with the language, modules, tools, and libraries; build community; and
discuss the path forward.
Let's review!
Questions?
 Acknowledgments

This work is supported in part by the the members of the Chicago Python
resources of the Argonne Leadership User's Group (ChiPy) for allowing us
Computing Facility at Argonne National to ramble on about science and HPC
Laboratory, which is supported by the the Python community for their
Office of Science of the U.S. Department feedback and support
of Energy under contract DE-AC02- CCT at LSU
06CH11357. numerous others at HPC centers
nationwide
Extended thanks to
Northwestern University
De Paul University
the families of the presenters
Sameer Shende, ParaTools, Inc.
Enthought, Inc. for their continued
support and sponsorship of SciPy
and NumPy
Lisandro Dalcin for his work on
mpi4py and tolerating a lot of
questions