Python for High Performance and Scientific Computing

Python for High Performance and Scientific
Computing

EuroPython 2011 (23.06.2011, Florence, Italy)
Andreas Schreiber <Andreas.Schreiber@dlr.de>
German Aerospace Center (DLR), Cologne
http://www.dlr.de/sc

EuroPython 2011 > Andreas Schreiber > Python for High Performance and Scientific Computing > 23.06.2011

Abstract

“Python is an accepted high-level scripting language with a growing community in
academia and industry. It is used in a lot of scientific applications in many different
scientific fields and in more and more industries, for example, in engineering or life
science). In all fields, the use of Python for high-performance and parallel computing is
increasing. Several organizations and companies are providing tools or support for
Python development. This includes libraries for scientific computing, parallel
computing, and MPI. Python is also used on many core architectures and GPUs, for
which specific Python interpreters are being developed. A related topic is the
performance of the various interpreter and compiler implementations for Python.
The talk gives an overview of Python’s use in HPC and Scientific Computing and gives
information on many topics, such as Python on massively parallel systems, GPU
programming with Python, scientific libraries in Python, and Python interpreter
performance issues. The talk will include examples for scientific codes and
applications from many domains..“
EuroPython 2011:
http://ep2011.europython.eu/conference/talks/python-for-high-performance-and-
scientific-computing
Folie 2

Outline

! "HPC! and Python
! "Tools and Libraries
! "Debugging and Profiling
! "Example Applications
! "Python and HPC communities

Folie 3

Related Talks at EuroPython

! "Debugging and Profiling Techniques (Giovanni Bajo)
! "Exploit your GPU Power with PyCUDA (and friends) (Stefano Brilli)
! "Experiences making CPU-bound tasks run much faster (Ian Ozsvald)
! "High-Performance Computing on Gamer PCs (Yann Le Du)
! "Python(x,y): Diving into Scientific Python (Vincent Noel)
! "and more!

Folie 4

Folie 5

DLR
German Aerospace Center

! "Research Institution
! "Space Agency
! "Project Management Agency

Folie 6

Locations and employees

6900 employees across Stade ! ! Hamburg
33 institutes and facilities at ! Neustrelitz
Bremen ! ! Trauen
!" 15 sites.
Berlin !
Braunschweig !

Offices in Brussels, ! Goettingen
Paris and Washington.
! Cologne
! Bonn

Lampoldshausen !

Stuttgart !
Augsburg !
! Oberpfaffenhofen
Weilheim !

Folie 7

Software Development at DLR
More than 1000 employees develop!

Mission-critical software
(e.g., real-time software)

Web-based software for
internet and intranet

Supporting software
(e.g., data management)

High-Performance and
Scientific Computing Software
Folie 8

HPC! and Python

Folie 9

Overview of Applications
Meteorology, Astrophysics, Particle physics, Biology,
Genetics, Quantum chemistry, Fluid dynamics, Finance,
Oil exploration, !

Images from: http://www.isgtw.org

Folie 10

HPC Systems

! "Current top systems have > 500,000 cores and > 8,000 TFlops
! "See TOP500 list of supercomputers
(http://www.top500.org)
! "Top 3 systems (June 2011) are:
1." K computer, Japan
2." Tianhe-1A, China
3." Jaguar, Unites States

Folie 11

#1: K computer, SPARC64 VIIIfx 2.0GHz, Tofu interc. (RIKEN
Advanced Institute for Computational Science (AICS), Japan)

Folie 12

#2: Tianhe-1A - NUDT TH MPP, X5670 2.93Ghz 6C, NVIDIA GPU,
FT-1000 8C (National Supercomputing Center in Tianjin, China)

Folie 13

#3: Jaguar - Cray XT5-HE Opteron 6-core 2.6 GHz (DOE/SC/Oak
Ridge National Laboratory, United States)

Folie 14

Applications
Python is used in!

! "Computational Fluid Dynamics (CFD)
! "Plasma simulation
! "Bio molecular simulation
! "Artificial intelligence
! "Natural language processing
! "Data mining
! "Scientific visualization
! "Robotics
! "Computer games
! "System administration
! "Education
! "!

Folie 15

Python for Scientists and Engineers
Reasons for Python in Research and Industry

Observations: “I want to design
! "Scientists and engineers don’t want to write software planes,
not software!”
but just solve their problems
! "If they have to write code, it must be as easy as
possible
Why Python is perfect?
! "Very easy to learn and easy to use
( = steep learning curve)
! "Allows rapid development
( = short development time)
! "Inherent great maintainability

Folie 16

“
!"#$#%&##'&%()%*#%(+)%&)$(&%
),%-#)-.#%+")%.)/#%01(")23%
(")&#%+")%"4(#%*$456#(&7%
428%&59#2:&(&;%%

Folie 17
23.06.2011

“
<,%9(=&%>))8%#2)?>"%,)$%%
@))>.#%428%ABCB7%9(=&%%
>))8%#2)?>"%,)$%'#7%*4*1;%%

Folie 18
23.06.2011

“
01(")2%"4&%("#%5.#42#&(7%%
')&(D&59#2:&(D%)$%#2>92##$%%
,$9#28.1%&12(4E%428%%
&#'42:5&;% Paul F. Dubois!

Folie 19
23.06.2011

Tools and Libraries for HPC and SC
with Python


Scientific Tools and Libraries
General Tools

Very general scientific computing High Performance Computing
! "NumPy Parallel Computing
! "SciPy ! PETSc
! PyMPI
Visualization ! Pypar
! Matplotlib ! "mpi4py
! VisIt GPGPU Computing
! MayaVi ! "PyCUDA
! "Chaco ! "PyOpenCL
! "VTK

Folie 21

Domain Specific Tools
AI Molecular & Atomic Electromagnetics
! "pyem Modeling ! PyFemax
! "ffnet ! PyMOL
! "pymorph ! Biskit Astronomy
! "Monte ! "GPAW ! AstroLib
! hcluster ! PySolar
Geo sciences
Biology ! "GIS Python Dynamic Systems
! "Brian ! PyClimate ! Simpy
! "SloppyCell ! ClimPy ! PyDSTool
! "NIPY ! "CDAT
! "PySAT Finite Elements
! SfePy
Folie 22

Special Topics

Wrapping other languages
! "weave (C/C++)
! "f2py (Fortran)
! Cython
! Ctypes (C)
! "SWIG (C/C++)
! RPy / RSPython (R)
! MatPy (Matlab)
! Jython (Java)
! IronPython (.NET)

Folie 23

NumPy

! "Website: http://numpy.scipy.org/

! "Offers capabilities similar to MATLAB within Python

! "N-dimensional homogeneous arrays (ndarray)
! "Universal functions (ufunc)
! "basic math, linear algebra, FFT, PRNGs
! "Simple data file I/O
! "text, raw binary, native binary
! "Tools for integrating with C/C++/Fortran
! "Heavy lifting done by optimized C/Fortran libraries
! "ATLAS or MKL, UMFPACK, FFTW, etc...

Folie 24

NumPy
Arrays

! "A NumPy array is an N-dimensional homogeneous collection of “items”
of the same “kind”. The kind can be any arbitrary structure and is
specified using the data-type.
! "Example: Multidimensional NumPy array

>>> a = array([[ 0, 1, 2, 3],
[10,11,12,13]])
>>> a
array([[ 0, 1, 2, 3],
[10,11,12,13]])

>>> a[1,3]
13
>>> a[1,3] = -1
>>> a
array([[ 0, 1, 2, 3],
[10,11,12,-1]])

Folie 25

SciPy
Scientific Tools for Python

! "Website: http://www.scipy.org
! "Large library of scientific algorithms
! "Extends NumPy with many tools for science and engineering
! "Computationally intensive routines implemented in C and Fortran

Folie 26

SciPy
Packages

! "Special Functions
! "Signal Processing
! "Fourier Transforms
! "Optimization
! "Numerical Integration
! "Linear Algebra
! Input/Output
! "Statistics
! "Fast Execution
! "Clustering Algorithms
! "Sparse Matrices

Folie 27

Parallel 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
GPU (OpenCL & CUDA)
! "PyOpenCL and PyCUDA provide low and high level abstraction for
highly parallel computations on GPUs

Folie 28

mpi4py

! "Website: http://mpi4py.scipy.org/
! "Wraps native MPI implementations
! "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 Python Distribution (EPD)
! "Requires NumPy
! "Portable and scalable

Folie 29

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 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

Folie 30

from mpi4py import MPI
Example: import numpy as np
Calculating " import random
with mpi4py comm = MPI.COMM_WORLD
rank = comm.Get_rank()
mpisize = comm.Get_size()
nsamples = int(12e6/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
Folie 31

GPGPU
General-purpose computing on graphics processing units

Architecture
! "Many cores per node
! "Good for stream processing (independent vertices, many of them in
parallel)
CUDA
! "NVIDIA’s platform for C programming on GPGPUs
! "Python binding: PyCUDA
OpenCL
! "Framework for writing programs that execute
across heterogeneous platforms consisting of
CPUs, GPUs, and other processors
! "Open standard
! "Python binding: PyOpenCL

Folie 32

PyCUDA

! "Website: http://mathema.tician.de/software/pycuda
! "PyCUDA lets you access NVIDIA‘s CUDA parallel computation API
from Python
! "Integration with NumPy
! "All CUDA errors are automatically
translated into Python exceptions

Folie 33

PyOpenCL

! "Website: http://mathema.tician.de/software/pyopencl
! "Convenient access to full OpenCL API from Python
! "OpenCL errors translated into Python exceptions
! "Object cleanup tied to object lifetime (follows Resource Acquisition Is
Initialization)
! "NumPy ndarrays interact easily with OpenCL buffers

Folie 34

OpenCL Computing Model

Buffers
! "used for holding data, shipping data between devices

Kernels
! "functions which will operate on buffers in parallel

Queue
! "control sequence of operations: transferring buffers, running kernels,
waiting on events, etc.

Folie 35

" with PyOpenCL
import pyopencl as cl
import pyopencl.clrandom
import numpy as np

nsamples = int(12e6)

# set up context and queue
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)

# create array of random values in OpenCL
xy = pyopencl.clrandom.rand(ctx,queue,(nsamples,2),np.float32)

# square values in OpenCL
xy = xy**2

# 'get' method on xy is used to get array from OpenCL into ndarray
print 4.0*np.sum(np.sum(xy.get(),1)<1)/nsamples
Folie 36

Copperhead: Data-Parallelism Embedded in Python
Slides by Michael Garland (NVIDIA Research)

http://bit.ly/jK7nsM

Folie 37

Debugging and Profiling

Folie 38

Debugging Python and C

! "Python comes with gdb-like
command line debugger pdb
! "Helps in debugging serial
Python applications
! "C-extensions can be debugged
with standard debuggers
! "Parallel debugging
! Totalview can be used for
debugging C-functions
! "Custom interpreter
provides easy initial
breakpoint

Folie 39

Profiling Mixed Python-C code

! "Finding performance bottlenecks is critical to scalability on HPC
platforms
! "Number of profiling tools available:
! gprof, CrayPAT, scalasca - C, Fortran
! "can be used for C-extensions
! "import profile – Python

! "TAU Performance System,
http://www.cs.uoregon.edu/research/tau/home.php
! "Excurse: Slides by Sameer Shende (University of Oregon)

Folie 40

TAU Performance System
Slides by Sameer Shende (University of Oregon)

http://slidesha.re/lsDZSj

Folie 41

Example Applications

Stand #

Folie 42

Example:
GIST Maxwell’s Equation Solver (GMES)

! "Object-Oriented Implementation of the Finite-Difference Time-Domain
Method in Parallel Computing Environment
! "Slides by Kyungwon Chun et al. (GIST - Gwagnju Institute of Science
and Technology, South Korea)!

http://bit.ly/mCs8W9

Folie 43

Example: FlowSimulator – A Python-controlled
framework to unify massive parallel CFD workflows

Folie 44

What does CFD mean?

! "CFD = Computational fluid dynamics means to
! " solve the compressible flow equations to
! " predict an aircrafts behavior and to
! " optimize its features.

! "Numerical approach:
! " Spatial discretization
! " Simplified models for turbulent effects
! " Temporal discretization (to support
unsteady flows)

Folie 45

Evolution of CFD
Classical chain
Z

! "Simple chain
Y X

! " Geometry
! " Mesh generator
! " Pre-processor
! " Flow solver
! " Post-processor
! " Visualization

! "Tools
! " stand-alone mesher
! " stand-alone solver
(with pre-/postprocessor)
! " shell scripts (e.g. different
solver parameters)

Folie 46

Evolution of CFD
Solver-adaptation-loop
Z

! "Chain with solver-adaptation-loop
Y X

! " several consecutive solver
steps
! " adaptation of mesh in between

! "Tools:
! " stand-alone mesher
! " stand-alone solver
! " stand-alone adaptation
! " shell script control

! "Solver/adaptation as Python module

Folie 47

Evolution of CFD
Multi-disciplinary simulations

! "Multi-disciplinary chain
Z

Y X

! " output of solver is input for
! " structural mechanics
(deformation)
! " flight mechanics (control
surface movement)
! " output of structural/flight
mechanics modifies mesh

! "Tools:
! " solver as Python module
! " stand-alone CSM
! " Python-based flight mechanics

Folie 48

Structure and Flight Mechanics

Folie 49

FlowSimulator Environment
Single Python script controls the whole workflow

FSControl
Out of scope

Flight
CFD CSM CAA
mechanics

Import Export
FSDataManager

Folie 50

FSControl
! "Pure Python library

! "Set of simple Python classes for different modules/tools:
! "simple interface: Import, Run, Export
! "data stored in FSDataManager

FSDataManager (FSDM)
! "C++ library, wrapped to Python
! "FSDataManager is a real Python module!

! "Parallelization context
! "all data structures “distributed”
! "facilitated transfers (Gather, Distribute, Send, !)

Folie 51

Python Framework for Coupled Fusion Simulations
Slides by Samantha Foley et al. (ORNL – Oak Ridge
National Laboratory, United States)

http://slidesha.re/iLuegW

Folie 52

Python & HPC Communities


Where the Python and HPC Communities Meet!

! "SciPy
! "Python for Scientific Computing
Conference (since 2008)
! "EuroSciPy
! "Annual European Conference for
Scientists using Python (since 2008)
! "SCxx
! "The International Conference for High
Performance Computing, Networking,
Storage, and Analysis (annual
conference since 1988)
! "Tutorial and BoF session on Python
(since 2009)

Folie 54

The SC Conference Series (SCxx)
Tutorials (est. 2009)

! "Python for Scientific and High Performance Computing
! "Given by William R. Scullin (ANL – Argonne National Lab, United States)
et al.
! "Content
! "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

Folie 55

The SC Conference Series (SCxx)
Birds-of-a-Feather Sessions, est. 2009

Folie 56

Credits
Credits
Achim Basermann (DLR)
Kyungwon Chun (GIST)
Samantha Foley (ORNL)
Michael Garland (NVIDIA)
Michael Meinel (DLR)
Tr a v i s O l i p h a n t ( E n t h o u g h t )
William R. Scullin (ANL)
Sameer Shende (U Oregon)

Folie 57

Questions?

Summary
•" Python used in HPC applications
•" Good tools available (SciPy etc.)
•" Wrapping of C/C++/Fortran is common

Andreas Schreiber
A n d r e a s . S c h r e i b e r @ d l r. d e
http://www.dlr.de/sc Folie 58

Contact

python@dlr.de
@onyame

Andreas Schreiber
A n d r e a s . S c h r e i b e r @ d l r. d e
http://www.dlr.de/sc Folie 59

Python for High Performance and Scientific Computing

Recommended

More Related Content

Similar to Python for High Performance and Scientific Computing (20)

More from Andreas Schreiber (20)

Recently uploaded (20)

Python for High Performance and Scientific Computing