Here's an interesting consequence of threaded programming that I found in python today.
The idea is that you have some worker thread (or threads), managed by the main thread. If the workers finish or fail, the main thread fires up more jobs for them to do unless the user Ctrl+C's (or otherwise interrupts) the main thread, signaling the workers to cleanup and exit.
This is a fairly standard problem so long as the tasks are not CPU bound and if your tasks are CPU bound, then see the note at the bottom of this post. And even if you are CPU bound, the meat of this post is still relavent.
Here is a program that spawns up the task in a thread, blocks until interrupted, then cleans up and exits.
def main(): t = MyTask() t.start() try: t.wait() except KeyboardInterrupt: t.stop() t.wait() if __name__ == '__main__': main()
We'll pretend that the MyTask class is doing all of the threading magic for us. The useful thing about this approach, is that I can spawn multiple tasks, and have them do things in parallel.
def main(): tasks = [MyTask() for t in range(5)] [t.start() for t in tasks] try: [t.wait() for t in tasks] except KeyboardInterrupt: [t.stop() for t in tasks] [t.wait() for t in tasks]
List comprehensions are fun.
So, what does the MyTask class actually look like, and what happens in .start, .stop and .wait?
import threading class MyTask(object): def __init__(self): self.task = get_task() # defined elsewhere, returns a callable self.thread = None # thread to run task in self.stopped = threading.Event() # threadsafe way to findout when to stop def monitor_task(self): while not stopped.wait(1): # Pretend this is a perfect world, with no exceptions. self.task() self.task = get_task() def start(self): self.thread = threading.Thread(target=self.monitor_task) self.thread.daemon = True self.stopped.clear() self.thread.start() def stop(self): self.stopped.set() # we'll pretend that this has some meaning too self.monitored_task.cleanup() def wait(self, timeout=None): return self.thread.join(timeout)
Well, that was easy. But there hides a subtle bug.
If the callable returned by get_task() runs forever, then there is no way to
stop the program. The subtlety is that the wait() in the main function's try
block. According to the
wait() method blocks until the flag is true", and they mean it.
Ctrl-C, SIGTERM, raising other exceptions are all blocked until another thread calls self.stopping.set() on the event. SIGKILL works, but there's no cleanup.
Eventually, I settled for the less elegant method of thread counting.
import itertools def main(): ... try: # Block until all tasks have ended for t in itertools.cycle(tasks): t.stopped.wait(1) # Non-blocking, doesn't eat CPU time if threading.active_count() <= 1: # Only really occurs if the tasks truly finish raise KeyboardInterrupt except KeyboardInterrupt: ...
I'm not sure I can think of a neater way right now.
The popular interpreters in python (CPython, pypy) have something known as the GIL. Essentially, to make the implementation easier, only one bit of bytecode is being interpreted at any given moment. This is not a problem with Python the language, as JPython and IronPython don't have a GIL, and many other interpreters also have a GIL too.
The effect is that this requires a small change in programming style to make the best use of a modern multicore system.
In Jython and IronPython, just keep using threads. CPU intensive tasks will scale with the number of cores present.
In GILed interpreters, it is best to spawn extra processes, which can live on different cores, and do the work there. Python's standard library offers two ways of doing this. The subprocess module, which offers a pythonic API over the unix process model and InterProcess Communication (IPC) using pipes to stdin/out/err. There is also the multiprocessing module, which offers an API compatible with the threading module. Porting threaded code to use multiprocessing is easy enough.
The downside is that processes do no share memory, unlike threads which allows for reading and writing to variables from different threads easy.
There is an effort to remove the GIL from pypy (and possibly port that to cpython) using a technique known as Transactional Memory. You can donate to the project at pypy.org