1 :mod:`multiprocessing` --- Process-based "threading" interface
2 ==============================================================
4 .. module:: multiprocessing
5 :synopsis: Process-based "threading" interface.
11 ----------------------
13 :mod:`multiprocessing` is a package that supports spawning processes using an
14 API similar to the :mod:`threading` module. The :mod:`multiprocessing` package
15 offers both local and remote concurrency, effectively side-stepping the
16 :term:`Global Interpreter Lock` by using subprocesses instead of threads. Due
17 to this, the :mod:`multiprocessing` module allows the programmer to fully
18 leverage multiple processors on a given machine. It runs on both Unix and
23 Some of this package's functionality requires a functioning shared semaphore
24 implementation on the host operating system. Without one, the
25 :mod:`multiprocessing.synchronize` module will be disabled, and attempts to
26 import it will result in an :exc:`ImportError`. See
27 :issue:`3770` for additional information.
31 Functionality within this package requires that the ``__main__`` method be
32 importable by the children. This is covered in :ref:`multiprocessing-programming`
33 however it is worth pointing out here. This means that some examples, such
34 as the :class:`multiprocessing.Pool` examples will not work in the
35 interactive interpreter. For example::
37 >>> from multiprocessing import Pool
46 Traceback (most recent call last):
47 Traceback (most recent call last):
48 Traceback (most recent call last):
49 AttributeError: 'module' object has no attribute 'f'
50 AttributeError: 'module' object has no attribute 'f'
51 AttributeError: 'module' object has no attribute 'f'
53 (If you try this it will actually output three full tracebacks
54 interleaved in a semi-random fashion, and then you may have to
55 stop the master process somehow.)
58 The :class:`Process` class
59 ~~~~~~~~~~~~~~~~~~~~~~~~~~
61 In :mod:`multiprocessing`, processes are spawned by creating a :class:`Process`
62 object and then calling its :meth:`~Process.start` method. :class:`Process`
63 follows the API of :class:`threading.Thread`. A trivial example of a
64 multiprocess program is ::
66 from multiprocessing import Process
71 if __name__ == '__main__':
72 p = Process(target=f, args=('bob',))
76 To show the individual process IDs involved, here is an expanded example::
78 from multiprocessing import Process
83 print 'module name:', __name__
84 print 'parent process:', os.getppid()
85 print 'process id:', os.getpid()
91 if __name__ == '__main__':
93 p = Process(target=f, args=('bob',))
97 For an explanation of why (on Windows) the ``if __name__ == '__main__'`` part is
98 necessary, see :ref:`multiprocessing-programming`.
102 Exchanging objects between processes
103 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
105 :mod:`multiprocessing` supports two types of communication channel between
110 The :class:`Queue` class is a near clone of :class:`Queue.Queue`. For
113 from multiprocessing import Process, Queue
116 q.put([42, None, 'hello'])
118 if __name__ == '__main__':
120 p = Process(target=f, args=(q,))
122 print q.get() # prints "[42, None, 'hello']"
125 Queues are thread and process safe.
129 The :func:`Pipe` function returns a pair of connection objects connected by a
130 pipe which by default is duplex (two-way). For example::
132 from multiprocessing import Process, Pipe
135 conn.send([42, None, 'hello'])
138 if __name__ == '__main__':
139 parent_conn, child_conn = Pipe()
140 p = Process(target=f, args=(child_conn,))
142 print parent_conn.recv() # prints "[42, None, 'hello']"
145 The two connection objects returned by :func:`Pipe` represent the two ends of
146 the pipe. Each connection object has :meth:`~Connection.send` and
147 :meth:`~Connection.recv` methods (among others). Note that data in a pipe
148 may become corrupted if two processes (or threads) try to read from or write
149 to the *same* end of the pipe at the same time. Of course there is no risk
150 of corruption from processes using different ends of the pipe at the same
154 Synchronization between processes
155 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
157 :mod:`multiprocessing` contains equivalents of all the synchronization
158 primitives from :mod:`threading`. For instance one can use a lock to ensure
159 that only one process prints to standard output at a time::
161 from multiprocessing import Process, Lock
165 print 'hello world', i
168 if __name__ == '__main__':
171 for num in range(10):
172 Process(target=f, args=(lock, num)).start()
174 Without using the lock output from the different processes is liable to get all
178 Sharing state between processes
179 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
181 As mentioned above, when doing concurrent programming it is usually best to
182 avoid using shared state as far as possible. This is particularly true when
183 using multiple processes.
185 However, if you really do need to use some shared data then
186 :mod:`multiprocessing` provides a couple of ways of doing so.
190 Data can be stored in a shared memory map using :class:`Value` or
191 :class:`Array`. For example, the following code ::
193 from multiprocessing import Process, Value, Array
197 for i in range(len(a)):
200 if __name__ == '__main__':
201 num = Value('d', 0.0)
202 arr = Array('i', range(10))
204 p = Process(target=f, args=(num, arr))
214 [0, -1, -2, -3, -4, -5, -6, -7, -8, -9]
216 The ``'d'`` and ``'i'`` arguments used when creating ``num`` and ``arr`` are
217 typecodes of the kind used by the :mod:`array` module: ``'d'`` indicates a
218 double precision float and ``'i'`` indicates a signed integer. These shared
219 objects will be process and thread safe.
221 For more flexibility in using shared memory one can use the
222 :mod:`multiprocessing.sharedctypes` module which supports the creation of
223 arbitrary ctypes objects allocated from shared memory.
227 A manager object returned by :func:`Manager` controls a server process which
228 holds Python objects and allows other processes to manipulate them using
231 A manager returned by :func:`Manager` will support types :class:`list`,
232 :class:`dict`, :class:`Namespace`, :class:`Lock`, :class:`RLock`,
233 :class:`Semaphore`, :class:`BoundedSemaphore`, :class:`Condition`,
234 :class:`Event`, :class:`Queue`, :class:`Value` and :class:`Array`. For
237 from multiprocessing import Process, Manager
245 if __name__ == '__main__':
249 l = manager.list(range(10))
251 p = Process(target=f, args=(d, l))
260 {0.25: None, 1: '1', '2': 2}
261 [9, 8, 7, 6, 5, 4, 3, 2, 1, 0]
263 Server process managers are more flexible than using shared memory objects
264 because they can be made to support arbitrary object types. Also, a single
265 manager can be shared by processes on different computers over a network.
266 They are, however, slower than using shared memory.
269 Using a pool of workers
270 ~~~~~~~~~~~~~~~~~~~~~~~
272 The :class:`~multiprocessing.pool.Pool` class represents a pool of worker
273 processes. It has methods which allows tasks to be offloaded to the worker
274 processes in a few different ways.
278 from multiprocessing import Pool
283 if __name__ == '__main__':
284 pool = Pool(processes=4) # start 4 worker processes
285 result = pool.apply_async(f, [10]) # evaluate "f(10)" asynchronously
286 print result.get(timeout=1) # prints "100" unless your computer is *very* slow
287 print pool.map(f, range(10)) # prints "[0, 1, 4,..., 81]"
293 The :mod:`multiprocessing` package mostly replicates the API of the
294 :mod:`threading` module.
297 :class:`Process` and exceptions
298 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
300 .. class:: Process([group[, target[, name[, args[, kwargs]]]]])
302 Process objects represent activity that is run in a separate process. The
303 :class:`Process` class has equivalents of all the methods of
304 :class:`threading.Thread`.
306 The constructor should always be called with keyword arguments. *group*
307 should always be ``None``; it exists solely for compatibility with
308 :class:`threading.Thread`. *target* is the callable object to be invoked by
309 the :meth:`run()` method. It defaults to ``None``, meaning nothing is
310 called. *name* is the process name. By default, a unique name is constructed
311 of the form 'Process-N\ :sub:`1`:N\ :sub:`2`:...:N\ :sub:`k`' where N\
312 :sub:`1`,N\ :sub:`2`,...,N\ :sub:`k` is a sequence of integers whose length
313 is determined by the *generation* of the process. *args* is the argument
314 tuple for the target invocation. *kwargs* is a dictionary of keyword
315 arguments for the target invocation. By default, no arguments are passed to
318 If a subclass overrides the constructor, it must make sure it invokes the
319 base class constructor (:meth:`Process.__init__`) before doing anything else
324 Method representing the process's activity.
326 You may override this method in a subclass. The standard :meth:`run`
327 method invokes the callable object passed to the object's constructor as
328 the target argument, if any, with sequential and keyword arguments taken
329 from the *args* and *kwargs* arguments, respectively.
333 Start the process's activity.
335 This must be called at most once per process object. It arranges for the
336 object's :meth:`run` method to be invoked in a separate process.
338 .. method:: join([timeout])
340 Block the calling thread until the process whose :meth:`join` method is
341 called terminates or until the optional timeout occurs.
343 If *timeout* is ``None`` then there is no timeout.
345 A process can be joined many times.
347 A process cannot join itself because this would cause a deadlock. It is
348 an error to attempt to join a process before it has been started.
354 The name is a string used for identification purposes only. It has no
355 semantics. Multiple processes may be given the same name. The initial
356 name is set by the constructor.
360 Return whether the process is alive.
362 Roughly, a process object is alive from the moment the :meth:`start`
363 method returns until the child process terminates.
365 .. attribute:: daemon
367 The process's daemon flag, a Boolean value. This must be set before
368 :meth:`start` is called.
370 The initial value is inherited from the creating process.
372 When a process exits, it attempts to terminate all of its daemonic child
375 Note that a daemonic process is not allowed to create child processes.
376 Otherwise a daemonic process would leave its children orphaned if it gets
377 terminated when its parent process exits. Additionally, these are **not**
378 Unix daemons or services, they are normal processes that will be
379 terminated (and not joined) if non-dameonic processes have exited.
381 In addition to the :class:`Threading.Thread` API, :class:`Process` objects
382 also support the following attributes and methods:
386 Return the process ID. Before the process is spawned, this will be
389 .. attribute:: exitcode
391 The child's exit code. This will be ``None`` if the process has not yet
392 terminated. A negative value *-N* indicates that the child was terminated
395 .. attribute:: authkey
397 The process's authentication key (a byte string).
399 When :mod:`multiprocessing` is initialized the main process is assigned a
400 random string using :func:`os.random`.
402 When a :class:`Process` object is created, it will inherit the
403 authentication key of its parent process, although this may be changed by
404 setting :attr:`authkey` to another byte string.
406 See :ref:`multiprocessing-auth-keys`.
408 .. method:: terminate()
410 Terminate the process. On Unix this is done using the ``SIGTERM`` signal;
411 on Windows :cfunc:`TerminateProcess` is used. Note that exit handlers and
412 finally clauses, etc., will not be executed.
414 Note that descendant processes of the process will *not* be terminated --
415 they will simply become orphaned.
419 If this method is used when the associated process is using a pipe or
420 queue then the pipe or queue is liable to become corrupted and may
421 become unusable by other process. Similarly, if the process has
422 acquired a lock or semaphore etc. then terminating it is liable to
423 cause other processes to deadlock.
425 Note that the :meth:`start`, :meth:`join`, :meth:`is_alive` and
426 :attr:`exit_code` methods should only be called by the process that created
429 Example usage of some of the methods of :class:`Process`:
433 >>> import multiprocessing, time, signal
434 >>> p = multiprocessing.Process(target=time.sleep, args=(1000,))
435 >>> print p, p.is_alive()
436 <Process(Process-1, initial)> False
438 >>> print p, p.is_alive()
439 <Process(Process-1, started)> True
442 >>> print p, p.is_alive()
443 <Process(Process-1, stopped[SIGTERM])> False
444 >>> p.exitcode == -signal.SIGTERM
448 .. exception:: BufferTooShort
450 Exception raised by :meth:`Connection.recv_bytes_into()` when the supplied
451 buffer object is too small for the message read.
453 If ``e`` is an instance of :exc:`BufferTooShort` then ``e.args[0]`` will give
454 the message as a byte string.
460 When using multiple processes, one generally uses message passing for
461 communication between processes and avoids having to use any synchronization
462 primitives like locks.
464 For passing messages one can use :func:`Pipe` (for a connection between two
465 processes) or a queue (which allows multiple producers and consumers).
467 The :class:`Queue` and :class:`JoinableQueue` types are multi-producer,
468 multi-consumer FIFO queues modelled on the :class:`Queue.Queue` class in the
469 standard library. They differ in that :class:`Queue` lacks the
470 :meth:`~Queue.Queue.task_done` and :meth:`~Queue.Queue.join` methods introduced
471 into Python 2.5's :class:`Queue.Queue` class.
473 If you use :class:`JoinableQueue` then you **must** call
474 :meth:`JoinableQueue.task_done` for each task removed from the queue or else the
475 semaphore used to count the number of unfinished tasks may eventually overflow
476 raising an exception.
478 Note that one can also create a shared queue by using a manager object -- see
479 :ref:`multiprocessing-managers`.
483 :mod:`multiprocessing` uses the usual :exc:`Queue.Empty` and
484 :exc:`Queue.Full` exceptions to signal a timeout. They are not available in
485 the :mod:`multiprocessing` namespace so you need to import them from
491 If a process is killed using :meth:`Process.terminate` or :func:`os.kill`
492 while it is trying to use a :class:`Queue`, then the data in the queue is
493 likely to become corrupted. This may cause any other processes to get an
494 exception when it tries to use the queue later on.
498 As mentioned above, if a child process has put items on a queue (and it has
499 not used :meth:`JoinableQueue.cancel_join_thread`), then that process will
500 not terminate until all buffered items have been flushed to the pipe.
502 This means that if you try joining that process you may get a deadlock unless
503 you are sure that all items which have been put on the queue have been
504 consumed. Similarly, if the child process is non-daemonic then the parent
505 process may hang on exit when it tries to join all its non-daemonic children.
507 Note that a queue created using a manager does not have this issue. See
508 :ref:`multiprocessing-programming`.
510 For an example of the usage of queues for interprocess communication see
511 :ref:`multiprocessing-examples`.
514 .. function:: Pipe([duplex])
516 Returns a pair ``(conn1, conn2)`` of :class:`Connection` objects representing
519 If *duplex* is ``True`` (the default) then the pipe is bidirectional. If
520 *duplex* is ``False`` then the pipe is unidirectional: ``conn1`` can only be
521 used for receiving messages and ``conn2`` can only be used for sending
525 .. class:: Queue([maxsize])
527 Returns a process shared queue implemented using a pipe and a few
528 locks/semaphores. When a process first puts an item on the queue a feeder
529 thread is started which transfers objects from a buffer into the pipe.
531 The usual :exc:`Queue.Empty` and :exc:`Queue.Full` exceptions from the
532 standard library's :mod:`Queue` module are raised to signal timeouts.
534 :class:`Queue` implements all the methods of :class:`Queue.Queue` except for
535 :meth:`~Queue.Queue.task_done` and :meth:`~Queue.Queue.join`.
539 Return the approximate size of the queue. Because of
540 multithreading/multiprocessing semantics, this number is not reliable.
542 Note that this may raise :exc:`NotImplementedError` on Unix platforms like
543 Mac OS X where ``sem_getvalue()`` is not implemented.
547 Return ``True`` if the queue is empty, ``False`` otherwise. Because of
548 multithreading/multiprocessing semantics, this is not reliable.
552 Return ``True`` if the queue is full, ``False`` otherwise. Because of
553 multithreading/multiprocessing semantics, this is not reliable.
555 .. method:: put(item[, block[, timeout]])
557 Put item into the queue. If the optional argument *block* is ``True``
558 (the default) and *timeout* is ``None`` (the default), block if necessary until
559 a free slot is available. If *timeout* is a positive number, it blocks at
560 most *timeout* seconds and raises the :exc:`Queue.Full` exception if no
561 free slot was available within that time. Otherwise (*block* is
562 ``False``), put an item on the queue if a free slot is immediately
563 available, else raise the :exc:`Queue.Full` exception (*timeout* is
564 ignored in that case).
566 .. method:: put_nowait(item)
568 Equivalent to ``put(item, False)``.
570 .. method:: get([block[, timeout]])
572 Remove and return an item from the queue. If optional args *block* is
573 ``True`` (the default) and *timeout* is ``None`` (the default), block if
574 necessary until an item is available. If *timeout* is a positive number,
575 it blocks at most *timeout* seconds and raises the :exc:`Queue.Empty`
576 exception if no item was available within that time. Otherwise (block is
577 ``False``), return an item if one is immediately available, else raise the
578 :exc:`Queue.Empty` exception (*timeout* is ignored in that case).
580 .. method:: get_nowait()
583 Equivalent to ``get(False)``.
585 :class:`multiprocessing.Queue` has a few additional methods not found in
586 :class:`Queue.Queue`. These methods are usually unnecessary for most
591 Indicate that no more data will be put on this queue by the current
592 process. The background thread will quit once it has flushed all buffered
593 data to the pipe. This is called automatically when the queue is garbage
596 .. method:: join_thread()
598 Join the background thread. This can only be used after :meth:`close` has
599 been called. It blocks until the background thread exits, ensuring that
600 all data in the buffer has been flushed to the pipe.
602 By default if a process is not the creator of the queue then on exit it
603 will attempt to join the queue's background thread. The process can call
604 :meth:`cancel_join_thread` to make :meth:`join_thread` do nothing.
606 .. method:: cancel_join_thread()
608 Prevent :meth:`join_thread` from blocking. In particular, this prevents
609 the background thread from being joined automatically when the process
610 exits -- see :meth:`join_thread`.
613 .. class:: JoinableQueue([maxsize])
615 :class:`JoinableQueue`, a :class:`Queue` subclass, is a queue which
616 additionally has :meth:`task_done` and :meth:`join` methods.
618 .. method:: task_done()
620 Indicate that a formerly enqueued task is complete. Used by queue consumer
621 threads. For each :meth:`~Queue.get` used to fetch a task, a subsequent
622 call to :meth:`task_done` tells the queue that the processing on the task
625 If a :meth:`~Queue.join` is currently blocking, it will resume when all
626 items have been processed (meaning that a :meth:`task_done` call was
627 received for every item that had been :meth:`~Queue.put` into the queue).
629 Raises a :exc:`ValueError` if called more times than there were items
635 Block until all items in the queue have been gotten and processed.
637 The count of unfinished tasks goes up whenever an item is added to the
638 queue. The count goes down whenever a consumer thread calls
639 :meth:`task_done` to indicate that the item was retrieved and all work on
640 it is complete. When the count of unfinished tasks drops to zero,
641 :meth:`~Queue.join` unblocks.
647 .. function:: active_children()
649 Return list of all live children of the current process.
651 Calling this has the side affect of "joining" any processes which have
654 .. function:: cpu_count()
656 Return the number of CPUs in the system. May raise
657 :exc:`NotImplementedError`.
659 .. function:: current_process()
661 Return the :class:`Process` object corresponding to the current process.
663 An analogue of :func:`threading.current_thread`.
665 .. function:: freeze_support()
667 Add support for when a program which uses :mod:`multiprocessing` has been
668 frozen to produce a Windows executable. (Has been tested with **py2exe**,
669 **PyInstaller** and **cx_Freeze**.)
671 One needs to call this function straight after the ``if __name__ ==
672 '__main__'`` line of the main module. For example::
674 from multiprocessing import Process, freeze_support
679 if __name__ == '__main__':
681 Process(target=f).start()
683 If the ``freeze_support()`` line is omitted then trying to run the frozen
684 executable will raise :exc:`RuntimeError`.
686 If the module is being run normally by the Python interpreter then
687 :func:`freeze_support` has no effect.
689 .. function:: set_executable()
691 Sets the path of the python interpreter to use when starting a child process.
692 (By default :data:`sys.executable` is used). Embedders will probably need to
693 do some thing like ::
695 setExecutable(os.path.join(sys.exec_prefix, 'pythonw.exe'))
697 before they can create child processes. (Windows only)
702 :mod:`multiprocessing` contains no analogues of
703 :func:`threading.active_count`, :func:`threading.enumerate`,
704 :func:`threading.settrace`, :func:`threading.setprofile`,
705 :class:`threading.Timer`, or :class:`threading.local`.
711 Connection objects allow the sending and receiving of picklable objects or
712 strings. They can be thought of as message oriented connected sockets.
714 Connection objects usually created using :func:`Pipe` -- see also
715 :ref:`multiprocessing-listeners-clients`.
717 .. class:: Connection
719 .. method:: send(obj)
721 Send an object to the other end of the connection which should be read
724 The object must be picklable. Very large pickles (approximately 32 MB+,
725 though it depends on the OS) may raise a ValueError exception.
729 Return an object sent from the other end of the connection using
730 :meth:`send`. Raises :exc:`EOFError` if there is nothing left to receive
731 and the other end was closed.
735 Returns the file descriptor or handle used by the connection.
739 Close the connection.
741 This is called automatically when the connection is garbage collected.
743 .. method:: poll([timeout])
745 Return whether there is any data available to be read.
747 If *timeout* is not specified then it will return immediately. If
748 *timeout* is a number then this specifies the maximum time in seconds to
749 block. If *timeout* is ``None`` then an infinite timeout is used.
751 .. method:: send_bytes(buffer[, offset[, size]])
753 Send byte data from an object supporting the buffer interface as a
756 If *offset* is given then data is read from that position in *buffer*. If
757 *size* is given then that many bytes will be read from buffer. Very large
758 buffers (approximately 32 MB+, though it depends on the OS) may raise a
761 .. method:: recv_bytes([maxlength])
763 Return a complete message of byte data sent from the other end of the
764 connection as a string. Raises :exc:`EOFError` if there is nothing left
765 to receive and the other end has closed.
767 If *maxlength* is specified and the message is longer than *maxlength*
768 then :exc:`IOError` is raised and the connection will no longer be
771 .. method:: recv_bytes_into(buffer[, offset])
773 Read into *buffer* a complete message of byte data sent from the other end
774 of the connection and return the number of bytes in the message. Raises
775 :exc:`EOFError` if there is nothing left to receive and the other end was
778 *buffer* must be an object satisfying the writable buffer interface. If
779 *offset* is given then the message will be written into the buffer from
780 that position. Offset must be a non-negative integer less than the
781 length of *buffer* (in bytes).
783 If the buffer is too short then a :exc:`BufferTooShort` exception is
784 raised and the complete message is available as ``e.args[0]`` where ``e``
785 is the exception instance.
792 >>> from multiprocessing import Pipe
794 >>> a.send([1, 'hello', None])
797 >>> b.send_bytes('thank you')
801 >>> arr1 = array.array('i', range(5))
802 >>> arr2 = array.array('i', [0] * 10)
803 >>> a.send_bytes(arr1)
804 >>> count = b.recv_bytes_into(arr2)
805 >>> assert count == len(arr1) * arr1.itemsize
807 array('i', [0, 1, 2, 3, 4, 0, 0, 0, 0, 0])
812 The :meth:`Connection.recv` method automatically unpickles the data it
813 receives, which can be a security risk unless you can trust the process
814 which sent the message.
816 Therefore, unless the connection object was produced using :func:`Pipe` you
817 should only use the :meth:`~Connection.recv` and :meth:`~Connection.send`
818 methods after performing some sort of authentication. See
819 :ref:`multiprocessing-auth-keys`.
823 If a process is killed while it is trying to read or write to a pipe then
824 the data in the pipe is likely to become corrupted, because it may become
825 impossible to be sure where the message boundaries lie.
828 Synchronization primitives
829 ~~~~~~~~~~~~~~~~~~~~~~~~~~
831 Generally synchronization primitives are not as necessary in a multiprocess
832 program as they are in a multithreaded program. See the documentation for
833 :mod:`threading` module.
835 Note that one can also create synchronization primitives by using a manager
836 object -- see :ref:`multiprocessing-managers`.
838 .. class:: BoundedSemaphore([value])
840 A bounded semaphore object: a clone of :class:`threading.BoundedSemaphore`.
842 (On Mac OS X this is indistinguishable from :class:`Semaphore` because
843 ``sem_getvalue()`` is not implemented on that platform).
845 .. class:: Condition([lock])
847 A condition variable: a clone of :class:`threading.Condition`.
849 If *lock* is specified then it should be a :class:`Lock` or :class:`RLock`
850 object from :mod:`multiprocessing`.
854 A clone of :class:`threading.Event`.
855 This method returns the state of the internal semaphore on exit, so it
856 will always return ``True`` except if a timeout is given and the operation
859 .. versionchanged:: 2.7
860 Previously, the method always returned ``None``.
864 A non-recursive lock object: a clone of :class:`threading.Lock`.
868 A recursive lock object: a clone of :class:`threading.RLock`.
870 .. class:: Semaphore([value])
872 A bounded semaphore object: a clone of :class:`threading.Semaphore`.
876 The :meth:`acquire` method of :class:`BoundedSemaphore`, :class:`Lock`,
877 :class:`RLock` and :class:`Semaphore` has a timeout parameter not supported
878 by the equivalents in :mod:`threading`. The signature is
879 ``acquire(block=True, timeout=None)`` with keyword parameters being
880 acceptable. If *block* is ``True`` and *timeout* is not ``None`` then it
881 specifies a timeout in seconds. If *block* is ``False`` then *timeout* is
885 On OS/X ``sem_timedwait`` is unsupported, so timeout arguments for the
886 aforementioned :meth:`acquire` methods will be ignored on OS/X.
890 If the SIGINT signal generated by Ctrl-C arrives while the main thread is
891 blocked by a call to :meth:`BoundedSemaphore.acquire`, :meth:`Lock.acquire`,
892 :meth:`RLock.acquire`, :meth:`Semaphore.acquire`, :meth:`Condition.acquire`
893 or :meth:`Condition.wait` then the call will be immediately interrupted and
894 :exc:`KeyboardInterrupt` will be raised.
896 This differs from the behaviour of :mod:`threading` where SIGINT will be
897 ignored while the equivalent blocking calls are in progress.
900 Shared :mod:`ctypes` Objects
901 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
903 It is possible to create shared objects using shared memory which can be
904 inherited by child processes.
906 .. function:: Value(typecode_or_type, *args[, lock])
908 Return a :mod:`ctypes` object allocated from shared memory. By default the
909 return value is actually a synchronized wrapper for the object.
911 *typecode_or_type* determines the type of the returned object: it is either a
912 ctypes type or a one character typecode of the kind used by the :mod:`array`
913 module. *\*args* is passed on to the constructor for the type.
915 If *lock* is ``True`` (the default) then a new lock object is created to
916 synchronize access to the value. If *lock* is a :class:`Lock` or
917 :class:`RLock` object then that will be used to synchronize access to the
918 value. If *lock* is ``False`` then access to the returned object will not be
919 automatically protected by a lock, so it will not necessarily be
922 Note that *lock* is a keyword-only argument.
924 .. function:: Array(typecode_or_type, size_or_initializer, *, lock=True)
926 Return a ctypes array allocated from shared memory. By default the return
927 value is actually a synchronized wrapper for the array.
929 *typecode_or_type* determines the type of the elements of the returned array:
930 it is either a ctypes type or a one character typecode of the kind used by
931 the :mod:`array` module. If *size_or_initializer* is an integer, then it
932 determines the length of the array, and the array will be initially zeroed.
933 Otherwise, *size_or_initializer* is a sequence which is used to initialize
934 the array and whose length determines the length of the array.
936 If *lock* is ``True`` (the default) then a new lock object is created to
937 synchronize access to the value. If *lock* is a :class:`Lock` or
938 :class:`RLock` object then that will be used to synchronize access to the
939 value. If *lock* is ``False`` then access to the returned object will not be
940 automatically protected by a lock, so it will not necessarily be
943 Note that *lock* is a keyword only argument.
945 Note that an array of :data:`ctypes.c_char` has *value* and *raw*
946 attributes which allow one to use it to store and retrieve strings.
949 The :mod:`multiprocessing.sharedctypes` module
950 >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
952 .. module:: multiprocessing.sharedctypes
953 :synopsis: Allocate ctypes objects from shared memory.
955 The :mod:`multiprocessing.sharedctypes` module provides functions for allocating
956 :mod:`ctypes` objects from shared memory which can be inherited by child
961 Although it is possible to store a pointer in shared memory remember that
962 this will refer to a location in the address space of a specific process.
963 However, the pointer is quite likely to be invalid in the context of a second
964 process and trying to dereference the pointer from the second process may
967 .. function:: RawArray(typecode_or_type, size_or_initializer)
969 Return a ctypes array allocated from shared memory.
971 *typecode_or_type* determines the type of the elements of the returned array:
972 it is either a ctypes type or a one character typecode of the kind used by
973 the :mod:`array` module. If *size_or_initializer* is an integer then it
974 determines the length of the array, and the array will be initially zeroed.
975 Otherwise *size_or_initializer* is a sequence which is used to initialize the
976 array and whose length determines the length of the array.
978 Note that setting and getting an element is potentially non-atomic -- use
979 :func:`Array` instead to make sure that access is automatically synchronized
982 .. function:: RawValue(typecode_or_type, *args)
984 Return a ctypes object allocated from shared memory.
986 *typecode_or_type* determines the type of the returned object: it is either a
987 ctypes type or a one character typecode of the kind used by the :mod:`array`
988 module. *\*args* is passed on to the constructor for the type.
990 Note that setting and getting the value is potentially non-atomic -- use
991 :func:`Value` instead to make sure that access is automatically synchronized
994 Note that an array of :data:`ctypes.c_char` has ``value`` and ``raw``
995 attributes which allow one to use it to store and retrieve strings -- see
996 documentation for :mod:`ctypes`.
998 .. function:: Array(typecode_or_type, size_or_initializer, *args[, lock])
1000 The same as :func:`RawArray` except that depending on the value of *lock* a
1001 process-safe synchronization wrapper may be returned instead of a raw ctypes
1004 If *lock* is ``True`` (the default) then a new lock object is created to
1005 synchronize access to the value. If *lock* is a :class:`Lock` or
1006 :class:`RLock` object then that will be used to synchronize access to the
1007 value. If *lock* is ``False`` then access to the returned object will not be
1008 automatically protected by a lock, so it will not necessarily be
1011 Note that *lock* is a keyword-only argument.
1013 .. function:: Value(typecode_or_type, *args[, lock])
1015 The same as :func:`RawValue` except that depending on the value of *lock* a
1016 process-safe synchronization wrapper may be returned instead of a raw ctypes
1019 If *lock* is ``True`` (the default) then a new lock object is created to
1020 synchronize access to the value. If *lock* is a :class:`Lock` or
1021 :class:`RLock` object then that will be used to synchronize access to the
1022 value. If *lock* is ``False`` then access to the returned object will not be
1023 automatically protected by a lock, so it will not necessarily be
1026 Note that *lock* is a keyword-only argument.
1028 .. function:: copy(obj)
1030 Return a ctypes object allocated from shared memory which is a copy of the
1031 ctypes object *obj*.
1033 .. function:: synchronized(obj[, lock])
1035 Return a process-safe wrapper object for a ctypes object which uses *lock* to
1036 synchronize access. If *lock* is ``None`` (the default) then a
1037 :class:`multiprocessing.RLock` object is created automatically.
1039 A synchronized wrapper will have two methods in addition to those of the
1040 object it wraps: :meth:`get_obj` returns the wrapped object and
1041 :meth:`get_lock` returns the lock object used for synchronization.
1043 Note that accessing the ctypes object through the wrapper can be a lot slower
1044 than accessing the raw ctypes object.
1047 The table below compares the syntax for creating shared ctypes objects from
1048 shared memory with the normal ctypes syntax. (In the table ``MyStruct`` is some
1049 subclass of :class:`ctypes.Structure`.)
1051 ==================== ========================== ===========================
1052 ctypes sharedctypes using type sharedctypes using typecode
1053 ==================== ========================== ===========================
1054 c_double(2.4) RawValue(c_double, 2.4) RawValue('d', 2.4)
1055 MyStruct(4, 6) RawValue(MyStruct, 4, 6)
1056 (c_short * 7)() RawArray(c_short, 7) RawArray('h', 7)
1057 (c_int * 3)(9, 2, 8) RawArray(c_int, (9, 2, 8)) RawArray('i', (9, 2, 8))
1058 ==================== ========================== ===========================
1061 Below is an example where a number of ctypes objects are modified by a child
1064 from multiprocessing import Process, Lock
1065 from multiprocessing.sharedctypes import Value, Array
1066 from ctypes import Structure, c_double
1068 class Point(Structure):
1069 _fields_ = [('x', c_double), ('y', c_double)]
1071 def modify(n, x, s, A):
1074 s.value = s.value.upper()
1079 if __name__ == '__main__':
1083 x = Value(c_double, 1.0/3.0, lock=False)
1084 s = Array('c', 'hello world', lock=lock)
1085 A = Array(Point, [(1.875,-6.25), (-5.75,2.0), (2.375,9.5)], lock=lock)
1087 p = Process(target=modify, args=(n, x, s, A))
1094 print [(a.x, a.y) for a in A]
1097 .. highlightlang:: none
1099 The results printed are ::
1104 [(3.515625, 39.0625), (33.0625, 4.0), (5.640625, 90.25)]
1106 .. highlightlang:: python
1109 .. _multiprocessing-managers:
1114 Managers provide a way to create data which can be shared between different
1115 processes. A manager object controls a server process which manages *shared
1116 objects*. Other processes can access the shared objects by using proxies.
1118 .. function:: multiprocessing.Manager()
1120 Returns a started :class:`~multiprocessing.managers.SyncManager` object which
1121 can be used for sharing objects between processes. The returned manager
1122 object corresponds to a spawned child process and has methods which will
1123 create shared objects and return corresponding proxies.
1125 .. module:: multiprocessing.managers
1126 :synopsis: Share data between process with shared objects.
1128 Manager processes will be shutdown as soon as they are garbage collected or
1129 their parent process exits. The manager classes are defined in the
1130 :mod:`multiprocessing.managers` module:
1132 .. class:: BaseManager([address[, authkey]])
1134 Create a BaseManager object.
1136 Once created one should call :meth:`start` or :meth:`serve_forever` to ensure
1137 that the manager object refers to a started manager process.
1139 *address* is the address on which the manager process listens for new
1140 connections. If *address* is ``None`` then an arbitrary one is chosen.
1142 *authkey* is the authentication key which will be used to check the validity
1143 of incoming connections to the server process. If *authkey* is ``None`` then
1144 ``current_process().authkey``. Otherwise *authkey* is used and it
1147 .. method:: start([initializer[, initargs]])
1149 Start a subprocess to start the manager. If *initializer* is not ``None``
1150 then the subprocess will call ``initializer(*initargs)`` when it starts.
1152 .. method:: serve_forever()
1154 Run the server in the current process.
1156 .. method:: from_address(address, authkey)
1158 A class method which creates a manager object referring to a pre-existing
1159 server process which is using the given address and authentication key.
1161 .. method:: get_server()
1163 Returns a :class:`Server` object which represents the actual server under
1164 the control of the Manager. The :class:`Server` object supports the
1165 :meth:`serve_forever` method::
1167 >>> from multiprocessing.managers import BaseManager
1168 >>> manager = BaseManager(address=('', 50000), authkey='abc')
1169 >>> server = manager.get_server()
1170 >>> server.serve_forever()
1172 :class:`Server` additionally has an :attr:`address` attribute.
1174 .. method:: connect()
1176 Connect a local manager object to a remote manager process::
1178 >>> from multiprocessing.managers import BaseManager
1179 >>> m = BaseManager(address=('127.0.0.1', 5000), authkey='abc')
1182 .. method:: shutdown()
1184 Stop the process used by the manager. This is only available if
1185 :meth:`start` has been used to start the server process.
1187 This can be called multiple times.
1189 .. method:: register(typeid[, callable[, proxytype[, exposed[, method_to_typeid[, create_method]]]]])
1191 A classmethod which can be used for registering a type or callable with
1194 *typeid* is a "type identifier" which is used to identify a particular
1195 type of shared object. This must be a string.
1197 *callable* is a callable used for creating objects for this type
1198 identifier. If a manager instance will be created using the
1199 :meth:`from_address` classmethod or if the *create_method* argument is
1200 ``False`` then this can be left as ``None``.
1202 *proxytype* is a subclass of :class:`BaseProxy` which is used to create
1203 proxies for shared objects with this *typeid*. If ``None`` then a proxy
1204 class is created automatically.
1206 *exposed* is used to specify a sequence of method names which proxies for
1207 this typeid should be allowed to access using
1208 :meth:`BaseProxy._callMethod`. (If *exposed* is ``None`` then
1209 :attr:`proxytype._exposed_` is used instead if it exists.) In the case
1210 where no exposed list is specified, all "public methods" of the shared
1211 object will be accessible. (Here a "public method" means any attribute
1212 which has a :meth:`__call__` method and whose name does not begin with
1215 *method_to_typeid* is a mapping used to specify the return type of those
1216 exposed methods which should return a proxy. It maps method names to
1217 typeid strings. (If *method_to_typeid* is ``None`` then
1218 :attr:`proxytype._method_to_typeid_` is used instead if it exists.) If a
1219 method's name is not a key of this mapping or if the mapping is ``None``
1220 then the object returned by the method will be copied by value.
1222 *create_method* determines whether a method should be created with name
1223 *typeid* which can be used to tell the server process to create a new
1224 shared object and return a proxy for it. By default it is ``True``.
1226 :class:`BaseManager` instances also have one read-only property:
1228 .. attribute:: address
1230 The address used by the manager.
1233 .. class:: SyncManager
1235 A subclass of :class:`BaseManager` which can be used for the synchronization
1236 of processes. Objects of this type are returned by
1237 :func:`multiprocessing.Manager`.
1239 It also supports creation of shared lists and dictionaries.
1241 .. method:: BoundedSemaphore([value])
1243 Create a shared :class:`threading.BoundedSemaphore` object and return a
1246 .. method:: Condition([lock])
1248 Create a shared :class:`threading.Condition` object and return a proxy for
1251 If *lock* is supplied then it should be a proxy for a
1252 :class:`threading.Lock` or :class:`threading.RLock` object.
1256 Create a shared :class:`threading.Event` object and return a proxy for it.
1260 Create a shared :class:`threading.Lock` object and return a proxy for it.
1262 .. method:: Namespace()
1264 Create a shared :class:`Namespace` object and return a proxy for it.
1266 .. method:: Queue([maxsize])
1268 Create a shared :class:`Queue.Queue` object and return a proxy for it.
1272 Create a shared :class:`threading.RLock` object and return a proxy for it.
1274 .. method:: Semaphore([value])
1276 Create a shared :class:`threading.Semaphore` object and return a proxy for
1279 .. method:: Array(typecode, sequence)
1281 Create an array and return a proxy for it.
1283 .. method:: Value(typecode, value)
1285 Create an object with a writable ``value`` attribute and return a proxy
1292 Create a shared ``dict`` object and return a proxy for it.
1297 Create a shared ``list`` object and return a proxy for it.
1303 A namespace object has no public methods, but does have writable attributes.
1304 Its representation shows the values of its attributes.
1306 However, when using a proxy for a namespace object, an attribute beginning with
1307 ``'_'`` will be an attribute of the proxy and not an attribute of the referent:
1311 >>> manager = multiprocessing.Manager()
1312 >>> Global = manager.Namespace()
1314 >>> Global.y = 'hello'
1315 >>> Global._z = 12.3 # this is an attribute of the proxy
1317 Namespace(x=10, y='hello')
1323 To create one's own manager, one creates a subclass of :class:`BaseManager` and
1324 use the :meth:`~BaseManager.register` classmethod to register new types or
1325 callables with the manager class. For example::
1327 from multiprocessing.managers import BaseManager
1329 class MathsClass(object):
1330 def add(self, x, y):
1332 def mul(self, x, y):
1335 class MyManager(BaseManager):
1338 MyManager.register('Maths', MathsClass)
1340 if __name__ == '__main__':
1341 manager = MyManager()
1343 maths = manager.Maths()
1344 print maths.add(4, 3) # prints 7
1345 print maths.mul(7, 8) # prints 56
1348 Using a remote manager
1349 >>>>>>>>>>>>>>>>>>>>>>
1351 It is possible to run a manager server on one machine and have clients use it
1352 from other machines (assuming that the firewalls involved allow it).
1354 Running the following commands creates a server for a single shared queue which
1355 remote clients can access::
1357 >>> from multiprocessing.managers import BaseManager
1359 >>> queue = Queue.Queue()
1360 >>> class QueueManager(BaseManager): pass
1361 >>> QueueManager.register('get_queue', callable=lambda:queue)
1362 >>> m = QueueManager(address=('', 50000), authkey='abracadabra')
1363 >>> s = m.get_server()
1364 >>> s.serve_forever()
1366 One client can access the server as follows::
1368 >>> from multiprocessing.managers import BaseManager
1369 >>> class QueueManager(BaseManager): pass
1370 >>> QueueManager.register('get_queue')
1371 >>> m = QueueManager(address=('foo.bar.org', 50000), authkey='abracadabra')
1373 >>> queue = m.get_queue()
1374 >>> queue.put('hello')
1376 Another client can also use it::
1378 >>> from multiprocessing.managers import BaseManager
1379 >>> class QueueManager(BaseManager): pass
1380 >>> QueueManager.register('get_queue')
1381 >>> m = QueueManager(address=('foo.bar.org', 50000), authkey='abracadabra')
1383 >>> queue = m.get_queue()
1387 Local processes can also access that queue, using the code from above on the
1388 client to access it remotely::
1390 >>> from multiprocessing import Process, Queue
1391 >>> from multiprocessing.managers import BaseManager
1392 >>> class Worker(Process):
1393 ... def __init__(self, q):
1395 ... super(Worker, self).__init__()
1397 ... self.q.put('local hello')
1400 >>> w = Worker(queue)
1402 >>> class QueueManager(BaseManager): pass
1404 >>> QueueManager.register('get_queue', callable=lambda: queue)
1405 >>> m = QueueManager(address=('', 50000), authkey='abracadabra')
1406 >>> s = m.get_server()
1407 >>> s.serve_forever()
1412 A proxy is an object which *refers* to a shared object which lives (presumably)
1413 in a different process. The shared object is said to be the *referent* of the
1414 proxy. Multiple proxy objects may have the same referent.
1416 A proxy object has methods which invoke corresponding methods of its referent
1417 (although not every method of the referent will necessarily be available through
1418 the proxy). A proxy can usually be used in most of the same ways that its
1423 >>> from multiprocessing import Manager
1424 >>> manager = Manager()
1425 >>> l = manager.list([i*i for i in range(10)])
1427 [0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
1429 <ListProxy object, typeid 'list' at 0x...>
1435 Notice that applying :func:`str` to a proxy will return the representation of
1436 the referent, whereas applying :func:`repr` will return the representation of
1439 An important feature of proxy objects is that they are picklable so they can be
1440 passed between processes. Note, however, that if a proxy is sent to the
1441 corresponding manager's process then unpickling it will produce the referent
1442 itself. This means, for example, that one shared object can contain a second:
1446 >>> a = manager.list()
1447 >>> b = manager.list()
1448 >>> a.append(b) # referent of a now contains referent of b
1451 >>> b.append('hello')
1453 [['hello']] ['hello']
1457 The proxy types in :mod:`multiprocessing` do nothing to support comparisons
1458 by value. So, for instance, we have:
1462 >>> manager.list([1,2,3]) == [1,2,3]
1465 One should just use a copy of the referent instead when making comparisons.
1467 .. class:: BaseProxy
1469 Proxy objects are instances of subclasses of :class:`BaseProxy`.
1471 .. method:: _callmethod(methodname[, args[, kwds]])
1473 Call and return the result of a method of the proxy's referent.
1475 If ``proxy`` is a proxy whose referent is ``obj`` then the expression ::
1477 proxy._callmethod(methodname, args, kwds)
1479 will evaluate the expression ::
1481 getattr(obj, methodname)(*args, **kwds)
1483 in the manager's process.
1485 The returned value will be a copy of the result of the call or a proxy to
1486 a new shared object -- see documentation for the *method_to_typeid*
1487 argument of :meth:`BaseManager.register`.
1489 If an exception is raised by the call, then then is re-raised by
1490 :meth:`_callmethod`. If some other exception is raised in the manager's
1491 process then this is converted into a :exc:`RemoteError` exception and is
1492 raised by :meth:`_callmethod`.
1494 Note in particular that an exception will be raised if *methodname* has
1497 An example of the usage of :meth:`_callmethod`:
1501 >>> l = manager.list(range(10))
1502 >>> l._callmethod('__len__')
1504 >>> l._callmethod('__getslice__', (2, 7)) # equiv to `l[2:7]`
1506 >>> l._callmethod('__getitem__', (20,)) # equiv to `l[20]`
1507 Traceback (most recent call last):
1509 IndexError: list index out of range
1511 .. method:: _getvalue()
1513 Return a copy of the referent.
1515 If the referent is unpicklable then this will raise an exception.
1517 .. method:: __repr__
1519 Return a representation of the proxy object.
1523 Return the representation of the referent.
1529 A proxy object uses a weakref callback so that when it gets garbage collected it
1530 deregisters itself from the manager which owns its referent.
1532 A shared object gets deleted from the manager process when there are no longer
1533 any proxies referring to it.
1539 .. module:: multiprocessing.pool
1540 :synopsis: Create pools of processes.
1542 One can create a pool of processes which will carry out tasks submitted to it
1543 with the :class:`Pool` class.
1545 .. class:: multiprocessing.Pool([processes[, initializer[, initargs]]])
1547 A process pool object which controls a pool of worker processes to which jobs
1548 can be submitted. It supports asynchronous results with timeouts and
1549 callbacks and has a parallel map implementation.
1551 *processes* is the number of worker processes to use. If *processes* is
1552 ``None`` then the number returned by :func:`cpu_count` is used. If
1553 *initializer* is not ``None`` then each worker process will call
1554 ``initializer(*initargs)`` when it starts.
1556 .. method:: apply(func[, args[, kwds]])
1558 Equivalent of the :func:`apply` built-in function. It blocks till the
1559 result is ready. Given this blocks, :meth:`apply_async` is better suited
1560 for performing work in parallel. Additionally, the passed
1561 in function is only executed in one of the workers of the pool.
1563 .. method:: apply_async(func[, args[, kwds[, callback]]])
1565 A variant of the :meth:`apply` method which returns a result object.
1567 If *callback* is specified then it should be a callable which accepts a
1568 single argument. When the result becomes ready *callback* is applied to
1569 it (unless the call failed). *callback* should complete immediately since
1570 otherwise the thread which handles the results will get blocked.
1572 .. method:: map(func, iterable[, chunksize])
1574 A parallel equivalent of the :func:`map` built-in function (it supports only
1575 one *iterable* argument though). It blocks till the result is ready.
1577 This method chops the iterable into a number of chunks which it submits to
1578 the process pool as separate tasks. The (approximate) size of these
1579 chunks can be specified by setting *chunksize* to a positive integer.
1581 .. method:: map_async(func, iterable[, chunksize[, callback]])
1583 A variant of the :meth:`.map` method which returns a result object.
1585 If *callback* is specified then it should be a callable which accepts a
1586 single argument. When the result becomes ready *callback* is applied to
1587 it (unless the call failed). *callback* should complete immediately since
1588 otherwise the thread which handles the results will get blocked.
1590 .. method:: imap(func, iterable[, chunksize])
1592 An equivalent of :func:`itertools.imap`.
1594 The *chunksize* argument is the same as the one used by the :meth:`.map`
1595 method. For very long iterables using a large value for *chunksize* can
1596 make make the job complete **much** faster than using the default value of
1599 Also if *chunksize* is ``1`` then the :meth:`!next` method of the iterator
1600 returned by the :meth:`imap` method has an optional *timeout* parameter:
1601 ``next(timeout)`` will raise :exc:`multiprocessing.TimeoutError` if the
1602 result cannot be returned within *timeout* seconds.
1604 .. method:: imap_unordered(func, iterable[, chunksize])
1606 The same as :meth:`imap` except that the ordering of the results from the
1607 returned iterator should be considered arbitrary. (Only when there is
1608 only one worker process is the order guaranteed to be "correct".)
1612 Prevents any more tasks from being submitted to the pool. Once all the
1613 tasks have been completed the worker processes will exit.
1615 .. method:: terminate()
1617 Stops the worker processes immediately without completing outstanding
1618 work. When the pool object is garbage collected :meth:`terminate` will be
1623 Wait for the worker processes to exit. One must call :meth:`close` or
1624 :meth:`terminate` before using :meth:`join`.
1627 .. class:: AsyncResult
1629 The class of the result returned by :meth:`Pool.apply_async` and
1630 :meth:`Pool.map_async`.
1632 .. method:: get([timeout])
1634 Return the result when it arrives. If *timeout* is not ``None`` and the
1635 result does not arrive within *timeout* seconds then
1636 :exc:`multiprocessing.TimeoutError` is raised. If the remote call raised
1637 an exception then that exception will be reraised by :meth:`get`.
1639 .. method:: wait([timeout])
1641 Wait until the result is available or until *timeout* seconds pass.
1645 Return whether the call has completed.
1647 .. method:: successful()
1649 Return whether the call completed without raising an exception. Will
1650 raise :exc:`AssertionError` if the result is not ready.
1652 The following example demonstrates the use of a pool::
1654 from multiprocessing import Pool
1659 if __name__ == '__main__':
1660 pool = Pool(processes=4) # start 4 worker processes
1662 result = pool.apply_async(f, (10,)) # evaluate "f(10)" asynchronously
1663 print result.get(timeout=1) # prints "100" unless your computer is *very* slow
1665 print pool.map(f, range(10)) # prints "[0, 1, 4,..., 81]"
1667 it = pool.imap(f, range(10))
1668 print it.next() # prints "0"
1669 print it.next() # prints "1"
1670 print it.next(timeout=1) # prints "4" unless your computer is *very* slow
1673 result = pool.apply_async(time.sleep, (10,))
1674 print result.get(timeout=1) # raises TimeoutError
1677 .. _multiprocessing-listeners-clients:
1679 Listeners and Clients
1680 ~~~~~~~~~~~~~~~~~~~~~
1682 .. module:: multiprocessing.connection
1683 :synopsis: API for dealing with sockets.
1685 Usually message passing between processes is done using queues or by using
1686 :class:`Connection` objects returned by :func:`Pipe`.
1688 However, the :mod:`multiprocessing.connection` module allows some extra
1689 flexibility. It basically gives a high level message oriented API for dealing
1690 with sockets or Windows named pipes, and also has support for *digest
1691 authentication* using the :mod:`hmac` module.
1694 .. function:: deliver_challenge(connection, authkey)
1696 Send a randomly generated message to the other end of the connection and wait
1699 If the reply matches the digest of the message using *authkey* as the key
1700 then a welcome message is sent to the other end of the connection. Otherwise
1701 :exc:`AuthenticationError` is raised.
1703 .. function:: answerChallenge(connection, authkey)
1705 Receive a message, calculate the digest of the message using *authkey* as the
1706 key, and then send the digest back.
1708 If a welcome message is not received, then :exc:`AuthenticationError` is
1711 .. function:: Client(address[, family[, authenticate[, authkey]]])
1713 Attempt to set up a connection to the listener which is using address
1714 *address*, returning a :class:`~multiprocessing.Connection`.
1716 The type of the connection is determined by *family* argument, but this can
1717 generally be omitted since it can usually be inferred from the format of
1718 *address*. (See :ref:`multiprocessing-address-formats`)
1720 If *authenticate* is ``True`` or *authkey* is a string then digest
1721 authentication is used. The key used for authentication will be either
1722 *authkey* or ``current_process().authkey)`` if *authkey* is ``None``.
1723 If authentication fails then :exc:`AuthenticationError` is raised. See
1724 :ref:`multiprocessing-auth-keys`.
1726 .. class:: Listener([address[, family[, backlog[, authenticate[, authkey]]]]])
1728 A wrapper for a bound socket or Windows named pipe which is 'listening' for
1731 *address* is the address to be used by the bound socket or named pipe of the
1736 If an address of '0.0.0.0' is used, the address will not be a connectable
1737 end point on Windows. If you require a connectable end-point,
1738 you should use '127.0.0.1'.
1740 *family* is the type of socket (or named pipe) to use. This can be one of
1741 the strings ``'AF_INET'`` (for a TCP socket), ``'AF_UNIX'`` (for a Unix
1742 domain socket) or ``'AF_PIPE'`` (for a Windows named pipe). Of these only
1743 the first is guaranteed to be available. If *family* is ``None`` then the
1744 family is inferred from the format of *address*. If *address* is also
1745 ``None`` then a default is chosen. This default is the family which is
1746 assumed to be the fastest available. See
1747 :ref:`multiprocessing-address-formats`. Note that if *family* is
1748 ``'AF_UNIX'`` and address is ``None`` then the socket will be created in a
1749 private temporary directory created using :func:`tempfile.mkstemp`.
1751 If the listener object uses a socket then *backlog* (1 by default) is passed
1752 to the :meth:`listen` method of the socket once it has been bound.
1754 If *authenticate* is ``True`` (``False`` by default) or *authkey* is not
1755 ``None`` then digest authentication is used.
1757 If *authkey* is a string then it will be used as the authentication key;
1758 otherwise it must be *None*.
1760 If *authkey* is ``None`` and *authenticate* is ``True`` then
1761 ``current_process().authkey`` is used as the authentication key. If
1762 *authkey* is ``None`` and *authenticate* is ``False`` then no
1763 authentication is done. If authentication fails then
1764 :exc:`AuthenticationError` is raised. See :ref:`multiprocessing-auth-keys`.
1766 .. method:: accept()
1768 Accept a connection on the bound socket or named pipe of the listener
1769 object and return a :class:`Connection` object. If authentication is
1770 attempted and fails, then :exc:`AuthenticationError` is raised.
1774 Close the bound socket or named pipe of the listener object. This is
1775 called automatically when the listener is garbage collected. However it
1776 is advisable to call it explicitly.
1778 Listener objects have the following read-only properties:
1780 .. attribute:: address
1782 The address which is being used by the Listener object.
1784 .. attribute:: last_accepted
1786 The address from which the last accepted connection came. If this is
1787 unavailable then it is ``None``.
1790 The module defines two exceptions:
1792 .. exception:: AuthenticationError
1794 Exception raised when there is an authentication error.
1799 The following server code creates a listener which uses ``'secret password'`` as
1800 an authentication key. It then waits for a connection and sends some data to
1803 from multiprocessing.connection import Listener
1804 from array import array
1806 address = ('localhost', 6000) # family is deduced to be 'AF_INET'
1807 listener = Listener(address, authkey='secret password')
1809 conn = listener.accept()
1810 print 'connection accepted from', listener.last_accepted
1812 conn.send([2.25, None, 'junk', float])
1814 conn.send_bytes('hello')
1816 conn.send_bytes(array('i', [42, 1729]))
1821 The following code connects to the server and receives some data from the
1824 from multiprocessing.connection import Client
1825 from array import array
1827 address = ('localhost', 6000)
1828 conn = Client(address, authkey='secret password')
1830 print conn.recv() # => [2.25, None, 'junk', float]
1832 print conn.recv_bytes() # => 'hello'
1834 arr = array('i', [0, 0, 0, 0, 0])
1835 print conn.recv_bytes_into(arr) # => 8
1836 print arr # => array('i', [42, 1729, 0, 0, 0])
1841 .. _multiprocessing-address-formats:
1846 * An ``'AF_INET'`` address is a tuple of the form ``(hostname, port)`` where
1847 *hostname* is a string and *port* is an integer.
1849 * An ``'AF_UNIX'`` address is a string representing a filename on the
1852 * An ``'AF_PIPE'`` address is a string of the form
1853 :samp:`r'\\\\.\\pipe\\{PipeName}'`. To use :func:`Client` to connect to a named
1854 pipe on a remote computer called *ServerName* one should use an address of the
1855 form :samp:`r'\\\\{ServerName}\\pipe\\{PipeName}'` instead.
1857 Note that any string beginning with two backslashes is assumed by default to be
1858 an ``'AF_PIPE'`` address rather than an ``'AF_UNIX'`` address.
1861 .. _multiprocessing-auth-keys:
1866 When one uses :meth:`Connection.recv`, the data received is automatically
1867 unpickled. Unfortunately unpickling data from an untrusted source is a security
1868 risk. Therefore :class:`Listener` and :func:`Client` use the :mod:`hmac` module
1869 to provide digest authentication.
1871 An authentication key is a string which can be thought of as a password: once a
1872 connection is established both ends will demand proof that the other knows the
1873 authentication key. (Demonstrating that both ends are using the same key does
1874 **not** involve sending the key over the connection.)
1876 If authentication is requested but do authentication key is specified then the
1877 return value of ``current_process().authkey`` is used (see
1878 :class:`~multiprocessing.Process`). This value will automatically inherited by
1879 any :class:`~multiprocessing.Process` object that the current process creates.
1880 This means that (by default) all processes of a multi-process program will share
1881 a single authentication key which can be used when setting up connections
1884 Suitable authentication keys can also be generated by using :func:`os.urandom`.
1890 Some support for logging is available. Note, however, that the :mod:`logging`
1891 package does not use process shared locks so it is possible (depending on the
1892 handler type) for messages from different processes to get mixed up.
1894 .. currentmodule:: multiprocessing
1895 .. function:: get_logger()
1897 Returns the logger used by :mod:`multiprocessing`. If necessary, a new one
1900 When first created the logger has level :data:`logging.NOTSET` and no
1901 default handler. Messages sent to this logger will not by default propagate
1904 Note that on Windows child processes will only inherit the level of the
1905 parent process's logger -- any other customization of the logger will not be
1908 .. currentmodule:: multiprocessing
1909 .. function:: log_to_stderr()
1911 This function performs a call to :func:`get_logger` but in addition to
1912 returning the logger created by get_logger, it adds a handler which sends
1913 output to :data:`sys.stderr` using format
1914 ``'[%(levelname)s/%(processName)s] %(message)s'``.
1916 Below is an example session with logging turned on::
1918 >>> import multiprocessing, logging
1919 >>> logger = multiprocessing.log_to_stderr()
1920 >>> logger.setLevel(logging.INFO)
1921 >>> logger.warning('doomed')
1922 [WARNING/MainProcess] doomed
1923 >>> m = multiprocessing.Manager()
1924 [INFO/SyncManager-...] child process calling self.run()
1925 [INFO/SyncManager-...] created temp directory /.../pymp-...
1926 [INFO/SyncManager-...] manager serving at '/.../listener-...'
1928 [INFO/MainProcess] sending shutdown message to manager
1929 [INFO/SyncManager-...] manager exiting with exitcode 0
1931 In addition to having these two logging functions, the multiprocessing also
1932 exposes two additional logging level attributes. These are :const:`SUBWARNING`
1933 and :const:`SUBDEBUG`. The table below illustrates where theses fit in the
1934 normal level hierarchy.
1936 +----------------+----------------+
1937 | Level | Numeric value |
1938 +================+================+
1939 | ``SUBWARNING`` | 25 |
1940 +----------------+----------------+
1941 | ``SUBDEBUG`` | 5 |
1942 +----------------+----------------+
1944 For a full table of logging levels, see the :mod:`logging` module.
1946 These additional logging levels are used primarily for certain debug messages
1947 within the multiprocessing module. Below is the same example as above, except
1948 with :const:`SUBDEBUG` enabled::
1950 >>> import multiprocessing, logging
1951 >>> logger = multiprocessing.log_to_stderr()
1952 >>> logger.setLevel(multiprocessing.SUBDEBUG)
1953 >>> logger.warning('doomed')
1954 [WARNING/MainProcess] doomed
1955 >>> m = multiprocessing.Manager()
1956 [INFO/SyncManager-...] child process calling self.run()
1957 [INFO/SyncManager-...] created temp directory /.../pymp-...
1958 [INFO/SyncManager-...] manager serving at '/.../pymp-djGBXN/listener-...'
1960 [SUBDEBUG/MainProcess] finalizer calling ...
1961 [INFO/MainProcess] sending shutdown message to manager
1962 [DEBUG/SyncManager-...] manager received shutdown message
1963 [SUBDEBUG/SyncManager-...] calling <Finalize object, callback=unlink, ...
1964 [SUBDEBUG/SyncManager-...] finalizer calling <built-in function unlink> ...
1965 [SUBDEBUG/SyncManager-...] calling <Finalize object, dead>
1966 [SUBDEBUG/SyncManager-...] finalizer calling <function rmtree at 0x5aa730> ...
1967 [INFO/SyncManager-...] manager exiting with exitcode 0
1969 The :mod:`multiprocessing.dummy` module
1970 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1972 .. module:: multiprocessing.dummy
1973 :synopsis: Dumb wrapper around threading.
1975 :mod:`multiprocessing.dummy` replicates the API of :mod:`multiprocessing` but is
1976 no more than a wrapper around the :mod:`threading` module.
1979 .. _multiprocessing-programming:
1981 Programming guidelines
1982 ----------------------
1984 There are certain guidelines and idioms which should be adhered to when using
1985 :mod:`multiprocessing`.
1993 As far as possible one should try to avoid shifting large amounts of data
1996 It is probably best to stick to using queues or pipes for communication
1997 between processes rather than using the lower level synchronization
1998 primitives from the :mod:`threading` module.
2002 Ensure that the arguments to the methods of proxies are picklable.
2004 Thread safety of proxies
2006 Do not use a proxy object from more than one thread unless you protect it
2009 (There is never a problem with different processes using the *same* proxy.)
2011 Joining zombie processes
2013 On Unix when a process finishes but has not been joined it becomes a zombie.
2014 There should never be very many because each time a new process starts (or
2015 :func:`active_children` is called) all completed processes which have not
2016 yet been joined will be joined. Also calling a finished process's
2017 :meth:`Process.is_alive` will join the process. Even so it is probably good
2018 practice to explicitly join all the processes that you start.
2020 Better to inherit than pickle/unpickle
2022 On Windows many types from :mod:`multiprocessing` need to be picklable so
2023 that child processes can use them. However, one should generally avoid
2024 sending shared objects to other processes using pipes or queues. Instead
2025 you should arrange the program so that a process which need access to a
2026 shared resource created elsewhere can inherit it from an ancestor process.
2028 Avoid terminating processes
2030 Using the :meth:`Process.terminate` method to stop a process is liable to
2031 cause any shared resources (such as locks, semaphores, pipes and queues)
2032 currently being used by the process to become broken or unavailable to other
2035 Therefore it is probably best to only consider using
2036 :meth:`Process.terminate` on processes which never use any shared resources.
2038 Joining processes that use queues
2040 Bear in mind that a process that has put items in a queue will wait before
2041 terminating until all the buffered items are fed by the "feeder" thread to
2042 the underlying pipe. (The child process can call the
2043 :meth:`Queue.cancel_join_thread` method of the queue to avoid this behaviour.)
2045 This means that whenever you use a queue you need to make sure that all
2046 items which have been put on the queue will eventually be removed before the
2047 process is joined. Otherwise you cannot be sure that processes which have
2048 put items on the queue will terminate. Remember also that non-daemonic
2049 processes will be automatically be joined.
2051 An example which will deadlock is the following::
2053 from multiprocessing import Process, Queue
2056 q.put('X' * 1000000)
2058 if __name__ == '__main__':
2060 p = Process(target=f, args=(queue,))
2062 p.join() # this deadlocks
2065 A fix here would be to swap the last two lines round (or simply remove the
2068 Explicitly pass resources to child processes
2070 On Unix a child process can make use of a shared resource created in a
2071 parent process using a global resource. However, it is better to pass the
2072 object as an argument to the constructor for the child process.
2074 Apart from making the code (potentially) compatible with Windows this also
2075 ensures that as long as the child process is still alive the object will not
2076 be garbage collected in the parent process. This might be important if some
2077 resource is freed when the object is garbage collected in the parent
2082 from multiprocessing import Process, Lock
2085 ... do something using "lock" ...
2087 if __name__ == '__main__':
2090 Process(target=f).start()
2092 should be rewritten as ::
2094 from multiprocessing import Process, Lock
2097 ... do something using "l" ...
2099 if __name__ == '__main__':
2102 Process(target=f, args=(lock,)).start()
2104 Beware replacing sys.stdin with a "file like object"
2106 :mod:`multiprocessing` originally unconditionally called::
2108 os.close(sys.stdin.fileno())
2110 in the :meth:`multiprocessing.Process._bootstrap` method --- this resulted
2111 in issues with processes-in-processes. This has been changed to::
2114 sys.stdin = open(os.devnull)
2116 Which solves the fundamental issue of processes colliding with each other
2117 resulting in a bad file descriptor error, but introduces a potential danger
2118 to applications which replace :func:`sys.stdin` with a "file-like object"
2119 with output buffering. This danger is that if multiple processes call
2120 :func:`close()` on this file-like object, it could result in the same
2121 data being flushed to the object multiple times, resulting in corruption.
2123 If you write a file-like object and implement your own caching, you can
2124 make it fork-safe by storing the pid whenever you append to the cache,
2125 and discarding the cache when the pid changes. For example::
2130 if pid != self._pid:
2135 For more information, see :issue:`5155`, :issue:`5313` and :issue:`5331`
2140 Since Windows lacks :func:`os.fork` it has a few extra restrictions:
2144 Ensure that all arguments to :meth:`Process.__init__` are picklable. This
2145 means, in particular, that bound or unbound methods cannot be used directly
2146 as the ``target`` argument on Windows --- just define a function and use
2149 Also, if you subclass :class:`Process` then make sure that instances will be
2150 picklable when the :meth:`Process.start` method is called.
2154 Bear in mind that if code run in a child process tries to access a global
2155 variable, then the value it sees (if any) may not be the same as the value
2156 in the parent process at the time that :meth:`Process.start` was called.
2158 However, global variables which are just module level constants cause no
2161 Safe importing of main module
2163 Make sure that the main module can be safely imported by a new Python
2164 interpreter without causing unintended side effects (such a starting a new
2167 For example, under Windows running the following module would fail with a
2168 :exc:`RuntimeError`::
2170 from multiprocessing import Process
2175 p = Process(target=foo)
2178 Instead one should protect the "entry point" of the program by using ``if
2179 __name__ == '__main__':`` as follows::
2181 from multiprocessing import Process, freeze_support
2186 if __name__ == '__main__':
2188 p = Process(target=foo)
2191 (The ``freeze_support()`` line can be omitted if the program will be run
2192 normally instead of frozen.)
2194 This allows the newly spawned Python interpreter to safely import the module
2195 and then run the module's ``foo()`` function.
2197 Similar restrictions apply if a pool or manager is created in the main
2201 .. _multiprocessing-examples:
2206 Demonstration of how to create and use customized managers and proxies:
2208 .. literalinclude:: ../includes/mp_newtype.py
2211 Using :class:`Pool`:
2213 .. literalinclude:: ../includes/mp_pool.py
2216 Synchronization types like locks, conditions and queues:
2218 .. literalinclude:: ../includes/mp_synchronize.py
2221 An showing how to use queues to feed tasks to a collection of worker process and
2222 collect the results:
2224 .. literalinclude:: ../includes/mp_workers.py
2227 An example of how a pool of worker processes can each run a
2228 :class:`SimpleHTTPServer.HttpServer` instance while sharing a single listening
2231 .. literalinclude:: ../includes/mp_webserver.py
2234 Some simple benchmarks comparing :mod:`multiprocessing` with :mod:`threading`:
2236 .. literalinclude:: ../includes/mp_benchmarks.py
2238 An example/demo of how to use the :class:`managers.SyncManager`, :class:`Process`
2239 and others to build a system which can distribute processes and work via a
2240 distributed queue to a "cluster" of machines on a network, accessible via SSH.
2241 You will need to have private key authentication for all hosts configured for
2244 .. literalinclude:: ../includes/mp_distributing.py