Python Threads
Aahz
[email protected]
http://starship.python.net/crew/aahz/
Powered by PythonPoint
http://www.reportlab.com/
title: Title
title:
Meta Tutorial
I'm hearing-impaired
Please write questions if at all possible
Pop Quiz
Slides and scripts on web
title: Meta Tutorial
Contents
Goal: Use Threads!
Thread Overview
Python's Thread Library
Two Applications
Web Spider
GUI Background Thread
Tips and tricks
title: Contents
Part 1: Thread Intro
What are threads?
GIL
Python threads
Brute force threads
title: Part 1
Generic Threads
Similar to processes
Shared memory
Light-weight
Difficult to set up
Especially cross-platform
title: Generic Threads
Why Use Threads?
Efficiency/speed
multiple CPUs, parallelize blocking I/O
Responsiveness
e.g. background thread for GUI
Algorithmic simplicity
simulations, data passing
(mostly skipped in this tutorial)
title: Why Use Threads?
Python Threads
Class-based
Use threading, not thread
Cross-platform, OS-level
Thread Library
title: Python Threads
Python 1.5.2
configure --with-thread
Except on MS Windows and some Linux
distributions
Multi-CPU bug
Creating/destroying large numbers of
threads
Upgrade to 2.x
title: Python 1.5.2
GIL
Global Interpreter Lock (GIL)
Full Documentation:
www.python.org/doc/current/api/threads.html
Only one Python thread can run
Even with multiple CPUs
GIL is your friend (really!)
title: GIL
GIL in Action
Which is faster?
One Thread
total = 1
for i in range(10000):
total += 1
total = 1
for i in range(10000):
total += 1
Two Threads
total = 1
for i in range(10000):
total += 1
total = 1
for i in range(10000):
total += 1
title: GIL in action
Dealing with GIL
sys.setcheckinterval()
(default 10)
C extensions can release GIL
Blocking I/O releases GIL
So does time.sleep(!=0)
Multiple Processes
CORBA, XML-RPC, sockets, etc.
title: Dealing with GIL
Share External Objects
Files, GUI, DB connections
title: Share External Objects 1
Share External Objects
Files, GUI, DB connections
Don't
title: Share External Objects 2
Share External Objects
Files, GUI, DB connections
Don't
Partial exception: print
Still need to share?
Use worker thread
title: Share External Objects 3
Create Python Threads
Subclass threading.Thread
Override __init__() and run()
Do not override start()
In __init__(), call
Thread.__init__()
title: Create Python Threads
Use Python Threads
Instantiate thread object
t = MyThread()
Start the thread
t.start()
Methods/attribs from outside thread
t.put('foo')
if t.done:
title: Use Python Threads
Non-threaded Example
class Retriever:
def __init__(self, URL):
self.URL = URL
def run(self):
self.page = self.getPage()
retriever = Retriever('http://www.foo.com/')
retriever.run()
URLs = retriever.getLinks()
title: Non-threaded Example
Threaded Example
from threading import Thread
class Retriever(Thread):
def __init__(self, URL):
Thread.__init__(self)
self.URL = URL
def run(self):
self.page = self.getPage()
retriever = Retriever('http://www.foo.com/')
retriever.start()
while retriever.isAlive():
time.sleep(1)
URLs = retriever.getLinks()
title: Threaded Example
Multiple Threads
seeds = ['http://www.foo.com/',
'http://www.bar.com/',
'http://www.baz.com/']
threadList = []
URLs = []
for seed in Seed:
retriever = Retriever(seed)
retriever.start()
threadList.append(retriever)
for retriever in threadList:
# join() is more efficient than sleep()
retriever.join()
URLs += retriever.getLinks()
title: Multiple Threads
Thread Methods
Module functions:
activeCount() (not useful)
enumerate() (not useful)
Thread object methods:
start()
join() (somewhat useful)
isAlive() (not useful)
isDaemon()
setDaemon()
title: Thread Methods
Unthreaded Spider
SingleThreadSpider.py
Compare Tools/webchecker/
title: Unthreaded Spider
Brute Force Threads
Quick-convert to multiple threads
Need worker class
Just inherit from threading.Thread
One instance per work unit
title: Brute Force Threads
Brute Thread Spider
BruteThreadSpider.py
Few changes from
SingleThreadSpider.py
Spawn one thread per retrieval
Inefficient polling in main loop
title: Brute Thread Spider
Recap Part 1
GIL
Creating threads
Brute force threads
title: Recap Part 1
Part 2
Thread Theory
Python Thread Library
title: Part 2
Thread Order
Non-determinate
Thread 1
Thread 2
print "a,",
print "b,",
print "c,",
print "1,",
print "2,",
print "3,",
Sample output
1,
a,
1,
a,
a,
b,
2,
b,
b,
c,
3,
1,
2,
1,
a,
2,
c,
2,
b,
3,
3,
3,
c,
c,
title: Thread Order
Thread Communication
Data protection
Synchronization
title: Thread Communication
Data Protection
Keeps shared memory safe
Restricted code access
Only one thread accesses block of code
"critical section lock"
aka "mutex", "atomic operation"
Similar to DBMS locking
title: Data Protection
Synchronization
Synchronize action between threads
Passing data
Threads wait for each other to finish tasks
More efficient than polling
aka "wait/notify", "rendezvous"
title: Synchronization
Thread Library
Lock()
RLock()
Semaphore()
Condition()
Event()
Queue.Queue()
title: Thread Library
Lock()
Basic building block
Handles either protection or
synchronization
Methods
acquire(blocking)
release()
title: Lock()
Critical Section Lock
Thread 1
Thread 2
mutex.acquire()
if myList:
work = myList.pop()
mutex.release()
...
...
...
...
...
...
...
...
mutex.acquire()
if len(myList)<10:
myList.append(work)
mutex.release()
title: Critical Section Lock
Misusing Lock()
Lock() steps on itself
mutex = Lock()
mutex.acquire()
...
mutex.acquire()
# OOPS!
title: Misusing Lock()
Synch Two Threads
class Synchronize:
def __init__(self):
self.lock = Lock()
def wait(self):
self.lock.acquire()
self.lock.acquire()
self.lock.release()
def notify(self):
self.lock.release()
Thread 1
self.synch.wait()
...
...
self.synch.notify()
Thread 2
...
self.synch.notify()
self.synch.wait()
...
title: Synching threads
RLock()
Mutex only
Other threads cannot release RLock()
Recursive
Methods
acquire(blocking)
release()
title: RLock()
Using RLock()
mutex = RLock()
mutex.acquire()
...
mutex.acquire()
...
mutex.release()
mutex.release()
Thread 1
mutex.acquire()
self.update()
mutex.release()
...
...
...
# Safe
Thread 2
...
...
...
mutex.acquire()
self.update()
mutex.release()
title: Using RLock()
Semaphore()
Restricts number of running threads
In Python, primarily useful for simulations
(but consider using microthreads)
Methods
Semaphore(value)
acquire(blocking)
release()
title: Semaphore()
Condition()
Methods
Condition(lock)
acquire(blocking)
release()
wait(timeout)
notify()
notifyAll()
title: Condition()
Using Condition()
Must use lock
cond = Condition()
cond.acquire()
cond.wait()
# or notify()/notifyAll()
cond.release()
Avoid timeout
Creates polling loop, so inefficient
title: Using Condition()
Event()
Thin wrapper for Condition()
Don't have to mess with lock
Only uses notifyAll(), so can be
inefficient
Methods
set()
clear()
isSet()
wait(timeout)
title: Event()
TMTOWTDI
Perl:
There's More Than One Way To Do It
Python:
There should be one - and preferably only
one - obvious way to do it
Threads more like Perl
title: TMTOWTDI
Producer/Consumer
Example: factory
One part of the factory produces part of a
widget; another part of the factory
consumes widget parts to make complete
widgets. Trick is to keep it all in balance.
title: Producer/Consumer
title:
Body factory
Wheel factory
Assembly
title: Factory 1
Factory Objects 1
Body
Wheels
body.list
body.rlock
body.event
assembly.event
wheels.list
wheels.rlock
wheels.event
assembly.event
Assembly
body.list
body.rlock
body.event
wheels.list
wheels.rlock
wheels.event
assembly.rlock
assembly.event
title: Factory Objects 1
Queue()
Built on top of thread
Use with both threading and
thread
Designed for subclassing
Can implement stack, priority queue, etc.
Simple!
Handles both data protection and
synchronization
title: Queue()
Queue() Objects
Methods
Queue(maxsize)
put(item,block)
get(block)
qsize()
empty()
full()
Raises exception non-blocking
title: Queue() Objects
Using Queue()
Thread 1
Thread 2
out = self.doWork()
queue2.put(output)
...
...
...
self.in = queue1.get()
...
...
self.in = queue2.get()
out = self.doWork()
queue1.put(output)
...
title: Using Queue()
Body factory
Wheel factory
Assembly
title: Factory 2
Factory Objects 2
Body
Wheels
body.queue
wheels.queue
Assembly
body.queue
wheels.queue
assembly.rlock
title: Factory Objects 2
Body factory
Wheel factory
Packager
Assembly
title: Factory 3
Factory Objects 3
Body
Wheels
body.queue
wheels.queue
Packager
while 1:
body = self.body.queue.get()
wheels = self.wheels.queue.get()
self.assembly.queue.put( (body,wheels) )
Assembly
assembly.queue
title: Factory Objects 3
Recap Part 2
Data protection and synchronization
Python Thread Library
Queues are good
title: Recap Part 2
Part 3: Two Apps
Using Queues
spider (thread pool)
GUI (Tkinter) (background thread)
title: Part 3
Spider w/Queue
ThreadPoolSpider.py
Two queues
Pass work to thread pool
Get links back from thread pool
Queue for both data and events
title: Spider w/Queue
Tkinter Intro
This space intentionally left blank
title: Tkinter Intro
GUI building blocks
Widgets
Windows, buttons, checkboxes, text entry,
listboxes
Events
Widget activation, keypress, mouse
movement, mouse click, timers
title: GUI building blocks
Widgets
Geometry manager
Register callbacks
title: Widgets
Events
Event loop
Trigger callbacks
title: Events
Tkinter resources
Web
www.python.org/topics/tkinter/doc.html
Books
Python and Tkinter Programming, John E.
Grayson
title: Tkinter resources
Fibonacci
Fibonacci.py
UI freezes during calc
Frequent screen updates slow calc
title: Fibonacci
Threaded Fibonacci
FibThreaded.py
Tkinter needs to poll
Use after event
Single-element queue
Use in non-blocking mode to minimize
updates
Must use "Quit" button
title: Threaded Fibonacci
FibThreaded Bugs and
Exercises
Fix deadlock on quit
Fix display of illegal values
Refactor for generic calc object
title: FibThreaded Bugs and Exercises
Compare Spider/Fib
Shared structures vs. callbacks
title: Compare Spider/Fib
Recap Part 3
title: Recap Part 3
Part 4: Miscellaneous
Grab bag of useful info
title: Part 4
GIL and Shared Vars
Safe: one bytecode
Single operations against Python basic
types (e.g. appending to a list)
Unsafe
Multiple operations against Python
variables (e.g. checking the length of a list
before appending) or any operation that
involves a callback to a class (e.g. the
__getattr__ hook)
title: GIL and Shared Vars
Locks vs GIL
Each lock is unique, a real OS-level
lock; GIL is separate
title: Locks vs GIL
GIL example
Mutex only reading threads
Threads 1,4
myList.append(work)
...
...
...
Threads 2,3,5
mutex.acquire()
if myList:
work = myList.pop()
mutex.release()
Not safe with UserList
title: GIL example
dis this
disassemble source to byte codes
Thread-unsafe statement
If a single Python statement uses the same
shared variable across multiple byte codes,
or if there are multiple mutually-dependent
shared variables, that statement is not
thread-safe
title: dis this
Performance Tip
python -O
Also set PYTHONOPTIMIZE
15% performance boost
Removes bytecodes (SET_LINENO)
Fewer context switches!
Also removes assert
title: Performance Tip
import Editorial
How to import
from threading import Thread, Semaphore
or
import threading
Don't use
from threading import *
title: import Editorial
GIL and C Extensions
Look for macros:
Py_BEGIN_ALLOW_THREADS
Py_END_ALLOW_THREADS
Some common extensions:
mxODBC - yes
NumPy - no
I/O exception: library problems
e.g. socket.gethostbyname()
title: GIL and C Extensions
Stackless/Microthreads
Not OS-level threads
Mix: cooperative and preemptive
Useful for thousands of threads
e.g. simulations
More info:
http://www.tismer.com/research/stackless/
http://world.std.com/~wware/uthread.html
title: Stackless/Microthreads
Killing Threads
title: Killing Threads
Debugging Threads
gdb
title: Debugging Threads
Thread Scheduling
always on same cpu?
specify CPU?
title: Thread Scheduling
Handling Exceptions
try/finally
Use to make sure locks get released
try/except
Close down all threads in outer block
Be careful to pass SystemExit and
KeyboardInterrupt
title: Handling Exceptions
try/finally
title: try/finally
try/except
title: try/except
Pop Quiz 1
How are threads and processes similar and different?
What is the GIL?
In what ways does the GIL make thread programming
easier and harder?
How do you create a thread in Python?
What should not be shared between threads?
title: Pop Quiz 1
Pop Quiz 2
What are "brute force" threads?
Explain what each of the following is used for:
Lock()
RLock()
Semaphore()
Condition()
Event()
Queue.Queue()
Why are queues great?
title: Pop Quiz 2
Pop Quiz 3
How do you handle exceptions?
title: Pop Quiz 3
