Python Threads Aahz aahz@pobox.com http://starship.python.net/crew/aahz/ Powered by PythonPoint http://www.reportlab.com/ Meta Tutorial I'm hearing-impaired Please write questions if at all possible Pop Quiz Slides and scripts on web Contents Goal: Use Threads! Thread Overview Python's Thread Library Two Applications Web Spider GUI Background Thread Tips and tricks Part 1: Thread Intro What are threads? GIL Python threads Brute force threads Generic Threads Similar to processes Shared memory Light-weight Difficult to set up Especially cross-platform 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) Python Threads Class-based Use threading, not thread Cross-platform, OS-level Thread Library 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 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!) 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,total = 1 ,for i in range(10000):,for i in range(10000): , total += 1, total += 1 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. Share External Objects Files, GUI, DB connections Share External Objects Files, GUI, DB connections Don't Share External Objects Files, GUI, DB connections Don't Partial exception: print Still need to share? Use worker thread Create Python Threads Subclass threading.Thread Override __init__() and run() Do not override start() In __init__(), call Thread.__init__() Use Python Threads Instantiate thread object t = MyThread() Start the thread t.start() Methods/attribs from outside thread t.put('foo') if t.done: 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() 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() 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() Thread Methods Module functions: activeCount() (not useful) enumerate() (not useful) Thread object methods: start() join() (somewhat useful) isAlive() (not useful) isDaemon() setDaemon() Unthreaded Spider SingleThreadSpider.py Compare Tools/webchecker/ Brute Force Threads Quick-convert to multiple threads Need worker class Just inherit from threading.Thread One instance per work unit Brute Thread Spider BruteThreadSpider.py Few changes from SingleThreadSpider.py Spawn one thread per retrieval Inefficient polling in main loop Recap Part 1 GIL Creating threads Brute force threads Part 2 Thread Theory Python Thread Library Thread Order Non-determinate . Thread 1. Thread 2 .. .print "a,",.print "1,", .print "b,",.print "2,", .print "c,",.print "3,", Sample output 1, a, b, 2, c, 3, a, b, c, 1, 2, 3, 1, 2, 3, a, b, c, a, b, 1, 2, 3, c, Thread Communication Data protection Synchronization 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 Synchronization Synchronize action between threads Passing data Threads wait for each other to finish tasks More efficient than polling aka "wait/notify", "rendezvous" Thread Library Lock() RLock() Semaphore() Condition() Event() Queue.Queue() Lock() Basic building block Handles either protection or synchronization Methods acquire(blocking) release() 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() Misusing Lock() Lock() steps on itself mutex = Lock() mutex.acquire() ... mutex.acquire() # OOPS! 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, Thread 2 ,, ,self.synch.wait(),... ,...,self.synch.notify() ,...,self.synch.wait() ,self.synch.notify(),... RLock() Mutex only Other threads cannot release RLock() Recursive Methods acquire(blocking) release() Using RLock() mutex = RLock() mutex.acquire() ... mutex.acquire() # Safe ... mutex.release() mutex.release() , Thread 1, Thread 2 ,, ,mutex.acquire(),... ,self.update(),... ,mutex.release(),... ,...,mutex.acquire() ,...,self.update() ,...,mutex.release() Semaphore() Restricts number of running threads In Python, primarily useful for simulations (but consider using microthreads) Methods Semaphore(value) acquire(blocking) release() Condition() Methods Condition(lock) acquire(blocking) release() wait(timeout) notify() notifyAll() Using Condition() Must use lock cond = Condition() cond.acquire() cond.wait() # or notify()/notifyAll() cond.release() Avoid timeout Creates polling loop, so inefficient 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) 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 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. Body factory Wheel factory Assembly Factory Objects 1 ,Body,Wheels ,, ,body.list,wheels.list ,body.rlock,wheels.rlock ,body.event,wheels.event ,assembly.event,assembly.event Assembly body.list body.rlock body.event wheels.list wheels.rlock wheels.event assembly.rlock assembly.event 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 Queue() Objects Methods Queue(maxsize) put(item,block) get(block) qsize() empty() full() Raises exception non-blocking Using Queue() ,Thread 1,Thread 2 ,, ,out = self.doWork(),... ,queue2.put(output),... ,...,self.in = queue2.get() ,...,out = self.doWork() ,...,queue1.put(output) ,self.in = queue1.get(),... Body factory Wheel factory Assembly Factory Objects 2 ,Body,Wheels ,, ,body.queue,wheels.queue Assembly body.queue wheels.queue assembly.rlock Body factory Wheel factory Packager Assembly 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 Recap Part 2 Data protection and synchronization Python Thread Library Queues are good Part 3: Two Apps Using Queues spider (thread pool) GUI (Tkinter) (background thread) Spider w/Queue ThreadPoolSpider.py Two queues Pass work to thread pool Get links back from thread pool Queue for both data and events Tkinter Intro This space intentionally left blank GUI building blocks Widgets Windows, buttons, checkboxes, text entry, listboxes Events Widget activation, keypress, mouse movement, mouse click, timers Widgets Geometry manager Register callbacks Events Event loop Trigger callbacks Tkinter resources Web www.python.org/topics/tkinter/doc.html Books Python and Tkinter Programming, John E. Grayson Fibonacci Fibonacci.py UI freezes during calc Frequent screen updates slow calc 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 FibThreaded Bugs and Exercises Fix deadlock on quit Fix display of illegal values Refactor for generic calc object Compare Spider/Fib Shared structures vs. callbacks New tokens vs. modified tokens Recap Part 3 Part 4: Miscellaneous Grab bag of useful info Data Passing Mutables - careful! References for speedy large objects 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) Locks vs GIL Each lock is unique, a real OS-level lock; GIL is separate GIL example Mutex only reading threads ; Threads 1,4; Threads 2,3,5 ;; ;myList.append(work);mutex.acquire() ;...;if myList: ;...; work = myList.pop() ;...;mutex.release() Not safe with UserList 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 Performance Tip python -O Also set PYTHONOPTIMIZE 15% performance boost Removes bytecodes (SET_LINENO) Fewer context switches! Also removes assert import Editorial How to import from threading import Thread, Semaphore or import threading Don't use from threading import * 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() 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 Killing Threads Debugging Threads gdb Thread Scheduling always on same cpu? specify CPU? 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 try/finally 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? 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? Pop Quiz 3 How do you handle exceptions?