PyPy London Demo Evening 2013

Welcome to the PyPy Demo Evening

Laurence Tratt

2013-08-27

1 / 5 http://soft-dev.org/

http://soft-dev.org/



Why are we here?




What do we do?

(l-r) Vasudevan, Bolz, Tratt, Barrett, Diekmann




What do we do?

• Aim: identify important challenges in softwaredevelopment.

• Strengths: language design and implementation.

• Immediate benefits: faster VMs.

• Long-term benefits: language composition.




What do we do?








What do we do?








What do we do?








This evening

1 Carl Friedrich Bolz PyPy overview.

2 Lukas Diekmann Storage strategies.

3 Maciej Fijalkowski NumPy.

4 Armin Rigo Software Transactional Memory(STM).

5 Edd Barrett Language composition usingmeta-tracing.




Can you help?

• Contributors.

• Resources.• Software Freedom Conservancy


http://sfconservancy.org/



Can you help?

• Contributors.• Resources.

• Software Freedom Conservancy


http://sfconservancy.org/



A Very Brief Introduction to PyPy

Carl Friedrich Bolz

PyPy Demo Evening, King’s College London,August 27, 2013

Carl Friedrich Bolz A Very Brief Introduction to PyPy

CPython is slow

CPython 1-3 orders of magnitude slower than CB

inary

Tre

es

Dhry

stone

Fannku

ch R

ed

ux

Fast

a

Knucl

eoti

de

Mand

elb

rot

Nb

od

y

Reg

exD

NA

RevC

om

p

Ric

hard

s

Sp

ect

ral N

orm

0.1

1

10

100

1000

Slo

wer

than C

, lo

wer

is b

ett

er

CJavaCpython


Reasons for Bad Performance

interpretation overheadlate bindingdispatchingboxing


Enter PyPy

a modern efficient implementation of Python

open source, MIT licensewritten in Python itself, then bootstrapped to Cuses a tracing JIT compiler to produce machine code atruntime


Enter PyPy

a modern efficient implementation of Pythonopen source, MIT licensewritten in Python itself, then bootstrapped to Cuses a tracing JIT compiler to produce machine code atruntime


Performance of PyPy

significantly faster than CPython, typically in the same order ofmagnitude than C

Bin

ary

Tre

es

Dhry

stone

Fannku

ch R

ed

ux

Fast

a

Knucl

eoti

de

Mand

elb

rot

Nb

od

y

Reg

exD

NA

RevC

om

p

Ric

hard

s

Sp

ect

ral N

orm

0.1

1

10

100

1000

Slo

wer

than C

, lo

wer

is b

ett

er

CJavaPyPyCPython

on average about 6.3 faster than CPython


Performance of PyPy

significantly faster than CPython, typically in the same order ofmagnitude than C

Bin

ary

Tre

es

Dhry

stone

Fannku

ch R

ed

ux

Fast

a

Knucl

eoti

de

Mand

elb

rot

Nb

od

y

Reg

exD

NA

RevC

om

p

Ric

hard

s

Sp

ect

ral N

orm

0.1

1

10

100

1000

Slo

wer

than C

, lo

wer

is b

ett

er

CJavaPyPyCPython

on average about 6.3 faster than CPython


Demo


Architecture

Python interpreter written in RPythonJIT-compiler automatically added via meta-tracing


Status

Python 2.7.3 support, 2.7.4 coming soonbeta-level support for Python 3.2, more coming eventuallypure Python code fully supported, please report as bug ifnot

C extension modules partially supported, if they arewell-behavedthey are slowuse cffi (a ctypes replacement) instead


Status

Python 2.7.3 support, 2.7.4 coming soonbeta-level support for Python 3.2, more coming eventuallypure Python code fully supported, please report as bug ifnotC extension modules partially supported, if they arewell-behavedthey are slowuse cffi (a ctypes replacement) instead


Questions?

PyPy is a fast JITted Python implementation(if something is not fast, please report it as a bug)open source under MIT licensehttp://pypy.org


http://pypy.org

Storage Strategies for Fast Containers

Lukas Diekmann

August, 27 2013





Collection strategies

introduced in PyPy 1.9

optimisation of collections for certain data types

improving speed

reducing memory




Idea

typical programs have homogeneously types collections

create optimised versions of collections for certain types

so far:

lists: ints, floats, strings/unicodesets: ints, floats, strings/unicodedicts: ints, strings/unicode




Optimisations

collections automatically change to most efficient strategy

store elements more memory efficiently

fast elements access




Further optimisations

collection creation and initalisation

split(d), set([1,2,3])

type based operations:

contains, difference, issubset

special strategies

RangeListStrategy: calculates elements on the fly

Tracing JIT interaction:

faster (low-level) comparisons, remove type checks




Results

paper at OOPSLA

on average ∼18% speedup

∼6% less memory usage

more info at http://soft-dev.org/pubs/


http://soft-dev.org/pubs/



Numpy on PyPy

Maciej Fijałkowski

King’s College London

August 27 2013

fijal Numpy on PyPy

Goals

fully compliant numpy replacement for PyPyfast looped operationsfast vectorized operations

fijal Numpy on PyPy

Why?

fast loopingsingle language

fijal Numpy on PyPy

Model

some programs have numerical kernels that canbe written in Csome don’thttp://arxiv.org/abs/1301.1334

image manipulation demoabstraction unfriendly

fijal Numpy on PyPy

http://arxiv.org/abs/1301.1334

Status

fast looped operationsok vectorized operations

fijal Numpy on PyPy

Future goals

finish numpymake it fastmake it compatible with more software(matplotlib, scipy)

fijal Numpy on PyPy

Funding

about $20k leftwe likely need morebehind schedule, but not behind budget

fijal Numpy on PyPy

Q&A

Thank you!

fijal Numpy on PyPy

Software Transactional Memory on PyPy

Pseudo-Goal

I “Kill the GIL”

I GIL = Global Interpreter Lock

Real Goals

I Multi-core programming

I But reasonable multi-core programming

I Using the recent model of Transactional Memory

PyPy-STM

I An executable pypy-stm which uses internally SoftwareTransactional Memory

I Optimistically run multiple threads in parallel

I The only new feature is atomic:

with atomic:piece of code...

Example of higher-level API

def work(...):...several more calls to:

transaction.add(work, ...)...

I Starts N threads, scheduling work() calls to them

I Each work() is done in an atomic block

I Multi-core, but as if all the work() are done sequentially

Status

I Kind of working without the JIT

I Roughly three times slower (you need four cores to see benefits)

I Working on the JIT support

Q&A

I Thank you!

I Budget of $10k left, likely more needed too

Unipycation

Combining Prolog and Python

Edd Barrett

August 27, 2013





Our Goal

The softdev team is exploring language compositions.

Ideally our compositions should be:

Easy to implement.

Transparent (as possible) to the user-programmer.

High performance.

Can meta-tracing help?




Unipycation: A Language Composition Experiment

PyPy + Pyrolog = Unipycation

Unipycation

Both interpreters implemented in RPython.

About 600 LoC of integration code.

A few months to develop.

Languages communicate via an API.

No syntax integration yet.




Why?

Explore composition of opposing paradigms.

Evaluate RPython as a language composition framework.

Performance/ease of development.

Composition with many realistic applications.

Example Scenario

Data acquisition by JSON/XML/Sqlite.

Easy in Python, not easy in Prolog.

Some kind of knowledge inference based upon data.

Perhaps not so easy in Python, trivial in Prolog.

Visualisation of Results.

Easy in Python, lack of library support in Prolog.




Example

Suppose we have a directed graph (London Underground?):

a c

b

d

e

d

g

And we need to ask questions like:

Where can I get to from ’b’ via at most 4 nodes and how?




This is Easy with Prolog

path.pl:

edge(a, c). edge(c, b). edge(c, d). edge(d, e).

edge(b, e). edge(c, f). edge(f, g). edge(e, g).

edge(g, b).

path(From , To, MaxLen , Nodes) :-

path(From , To , MaxLen , Nodes , 1).

path(Node , Node , _, [Node], _).

path(From , To, MaxLen , [From | Ahead ], Len) :-

Len < MaxLen , edge(From , Next),

Len1 is Len + 1,

path(Next , To , MaxLen , Ahead , Len1).

query: path(b, To, 4, Path).




Example: Python → Prolog

from uni import Engine

engine = Engine.from_file ("path.pl")

paths = engine.db.path.iter

for (to , nodes) in paths("b", None , 4, None):

print ("To %s via %s" % (to, nodes))

To b via [’b’]

To e via [’b’, ’e’]

To g via [’b’, ’e’, ’g’]

To b via [’b’, ’e’, ’g’, ’b’]

Calling from Prolog to Python also possible. E.g.python:somefunc(blah)




Performance




In Summary

Compositions are relatively easy to implement with RPython.

We were able to implement a fairly transparent API-like interface.

Performance promising.

Further Reading

https://bitbucket.org/vext01/pypy



https://bitbucket.org/vext01/pypy




Looking for Python developers!

Speak to us at PyPy demo night or visit

www.pitchup.com/jobs

Python / DjangoPostgres

CeleryRedisnginx

memcacheJquery

SolrS3

Leading booking site for campsites and caravan parks, founded in 2009 by lastminute.com alumni● 65k visits / day, £6m bookings / year● 650 bookable sites● Huge market

○ 26k campsites and 300m bednights in Europe○ 600m bednights in US○ 47% more bednights than hotels (GB)○ More trips to campsites than holidays to France +

Spain combined (GB)● Team of 15, based in west London

http://www.pitchup.com/jobs

http://www.pitchup.com/jobs

Technology

PyPy London Demo Evening 2013