Потоковые и многопроцессорные модули на Python
Главная идея потоков заключается в выполнении последовательности таких инструкций внутри программы, которые могут выполняться независимо от другого кода.
Так в чём же разница между потоковой и многопроцессорной обработкой данных? При одновременном выполнении нескольких задач обычно используется потоковая обработка, а при процессно-ориентированном параллелизме задействуется многопроцессорная обработка.
Задачи с ограничением скорости вычислений и ввода-вывода
Время выполнения задач, ограниченных скоростью вычислений, полностью зависит от производительности процессора, тогда как в задачах I/O Bound скорость выполнения процесса ограничена скоростью системы ввода-вывода.
В задачах с ограничением скорости вычислений программа расходует большую часть времени на использование центрального процессора, то есть на выполнение вычислений. К таким задачам можно отнести программы, занимающиеся исключительно перемалыванием чисел и проведением расчётов.
В задачах, ограниченных скоростью ввода-вывода, программы обрабатывают большие объёмы данных с диска в сравнении с необходимым объёмом вычислений. К таким задачам можно отнести, например, подсчёт количества строк в файле.
Проблема GIL на Python
Обычно на Python используется только один поток для выполнения нескольких записанных инструкций, то есть одновременно выполняется только один поток. Производительность однопоточного и многопоточного процессов здесь одинакова, и происходит это из-за GIL (Global Interpreter Lock — глобальной блокировки интерпретатора). Эта глобальная блокировка интерпретатора сама действует как поток и ограничивает другие потоки, делая невозможной многопоточность на Python.
Процессы ускоряют операции на Python, которые создают интенсивную вычислительную нагрузку на центральный процессор, используя сразу несколько ядер и избегая GIL, в то время как потоки лучше подходят для задач ввода-вывода или задач, связанных со внешними системами, потому что потоки могут более эффективно работать вместе. Для объединения процессов им нужно сериализовывать свои результаты, на что требуется время.
Потоки на Python не дают никаких преимуществ для задач, создающих интенсивную вычислительную нагрузку на процессор, именно из-за GIL.
Зачем нужен GIL?
Потоковый модуль использует потоки, многопроцессорный модуль использует процессы. Разница в том, что потоки выполняются в одном и том же пространстве памяти, а у процессов отдельная память. Это немного затрудняет совместное использование объектов процессами с многопроцессорной обработкой. В этом случае обычно выполняется сериализация объектов. Но потоки используют одну память, поэтому нужно быть осторожным, иначе два потока будут записывать данные в одну и ту же память одновременно. Именно для этого и существует глобальная блокировка интерпретатора.
Если бы мы запустили на Python скрипт, выполняющий простую задачу — спать (ну очень времязатратную!), он выглядел бы так:
Учимся писать многопоточные и многопроцессные приложения на Python
Эта статья не для матёрых укротителей Python’а, для которых распутать этот клубок змей — детская забава, а скорее поверхностный обзор многопоточных возможностей для недавно подсевших на питон.
К сожалению по теме многопоточности в Python не так уж много материала на русском языке, а питонеры, которые ничего не слышали, например, про GIL, мне стали попадаться с завидной регулярностью. В этой статье я постараюсь описать самые основные возможности многопоточного питона, расскажу что же такое GIL и как с ним (или без него) жить и многое другое.
Python — очаровательный язык программирования. В нем прекрасно сочетается множество парадигм программирования. Большинство задач, с которыми может встретиться программист, решаются здесь легко, элегантно и лаконично. Но для всех этих задач зачастую достаточно однопоточного решения, а однопоточные программы обычно предсказуемы и легко поддаются отладке. Чего не скажешь о многопоточных и многопроцессных программах.
Многопоточные приложения
В Python есть модуль threading, и в нем есть все, что нужно для многопоточного программирования: тут есть и различного вида локи, и семафор, и механизм событий. Один словом — все, что нужно для подавляющего большинства многопоточных программ. Причем пользоваться всем этим инструментарием достаточно просто. Рассмотрим пример программы, которая запускает 2 потока. Один поток пишет десять “0”, другой — десять “1”, причем строго по-очереди.
Никакой магии и voodoo-кода. Код четкий и последовательный. Причем, как можно заметить, мы создали поток из функции. Для небольших задач это очень удобно. Этот код еще и достаточно гибкий. Допустим у нас появился 3-й процесс, который пишет “2”, тогда код будет выглядеть так:
Мы добавили новое событие, новый поток и слегка изменили параметры, с которыми
стартуют потоки (можно конечно написать и более общее решение с использованием, например, MapReduce, но это уже выходит за рамки этой статьи).
Как видим по-прежнему никакой магии. Все просто и понятно. Поехали дальше.
Global Interpreter Lock
Существуют две самые распространенные причины использовать потоки: во-первых, для увеличения эффективности использования многоядерной архитектуры cоврменных процессоров, а значит, и производительности программы;
во-вторых, если нам нужно разделить логику работы программы на параллельные полностью или частично асинхронные секции (например, иметь возможность пинговать несколько серверов одновременно).
В первом случае мы сталкиваемся с таким ограничением Python (а точнее основной его реализации CPython), как Global Interpreter Lock (или сокращенно GIL). Концепция GIL заключается в том, что в каждый момент времени только один поток может исполняться процессором. Это сделано для того, чтобы между потоками не было борьбы за отдельные переменные. Исполняемый поток получает доступ по всему окружению. Такая особенность реализации потоков в Python значительно упрощает работу с потоками и дает определенную потокобезопасность (thread safety).
Но тут есть тонкий момент: может показаться, что многопоточное приложение будет работать ровно столько же времени, сколько и однопоточное, делающее то же самое, или за сумму времени исполнения каждого потока на CPU. Но тут нас поджидает один неприятный эффект. Рассмотрим программу:
Эта программа просто пишет в файл миллион строк “1” и делает это за
0.35 секунды на моем компьютере.
Рассмотрим другую программу:
Эта программа создает 2 потока. В каждом потоке она пишет в отдельный файлик по пол миллиона строк “1”. По-сути объем работы такой же, как и у предыдущей программы. А вот со временем работы тут получается интересный эффект. Программа может работать от 0.7 секунды до аж 7 секунд. Почему же так происходит?
Это происходит из-за того, что когда поток не нуждается в ресурсе CPU — он освобождает GIL, а в этот момент его может попытаться получить и он сам, и другой поток, и еще и главный поток. При этом операционная система, зная, что ядер много, может усугубить все попыткой распределить потоки между ядрами.
UPD: на данный момент в Python 3.2 существует улучшенная реализация GIL, в которой эта проблема частично решается, в частности, за счет того, что каждый поток после потери управления ждет небольшой промежуток времени до того, как сможет опять захватить GIL (на эту тему есть хорошая презентация на английском)
«Выходит на Python нельзя писать эффективные многопоточные программы?», — спросите вы. Нет, конечно, выход есть и даже несколько.
Многопроцессные приложения
Для того, чтобы в некотором смысле решить проблему, описанную в предыдущем параграфе, в Python есть модуль subprocess. Мы можем написать программу, которую хотим исполнять в параллельном потоке (на самом деле уже процессе). И запускать ее в одном или нескольких потоках в другой программе. Такой способ действительно ускорил бы работу нашей программы, потому, что потоки, созданные в запускающей программе GIL не забирают, а только ждут завершения запущенного процесса. Однако, в этом способе есть масса проблем. Основная проблема заключается в том, что передавать данные между процессами становится трудно. Пришлось бы как-то сериализовать объекты, налаживать связь через PIPE или друге инструменты, а ведь все это несет неизбежно накладные расходы и код становится сложным для понимания.
Здесь нам может помочь другой подход. В Python есть модуль multiprocessing. По функциональности этот модуль напоминает threading. Например, процессы можно создавать точно так же из обычных функций. Методы работы с процессами почти все те же самые, что и для потоков из модуля threading. А вот для синхронизации процессов и обмена данными принято использовать другие инструменты. Речь идет об очередях (Queue) и каналах (Pipe). Впрочем, аналоги локов, событий и семафоров, которые были в threading, здесь тоже есть.
Кроме того в модуле multiprocessing есть механизм работы с общей памятью. Для этого в модуле есть классы переменной (Value) и массива (Array), которые можно “обобщать” (share) между процессами. Для удобства работы с общими переменными можно использовать классы-менеджеры (Manager). Они более гибкие и удобные в обращении, однако более медленные. Нельзя не отметить приятную возможность делать общими типы из модуля ctypes с помощью модуля multiprocessing.sharedctypes.
Еще в модуле multiprocessing есть механизм создания пулов процессов. Этот механизм очень удобно использовать для реализации шаблона Master-Worker или для реализации параллельного Map (который в некотором смысле является частным случаем Master-Worker).
Из основных проблем работы с модулем multiprocessing стоит отметить относительную платформозависимость этого модуля. Поскольку в разных ОС работа с процессами организована по-разному, то на код накладываются некоторые ограничения. Например, в ОС Windows нет механизма fork, поэтому точку разделения процессов надо оборачивать в:
Впрочем, эта конструкция и так является хорошим тоном.
Что еще .
Для написания параллельных приложений на Python существуют и другие библиотеки и подходы. Например, можно использовать Hadoop+Python или различные реализации MPI на Python (pyMPI, mpi4py). Можно даже использовать обертки существующих библиотек на С++ или Fortran. Здесь можно было упомянуть про такие фреймфорки/библиотеки, как Pyro, Twisted, Tornado и многие другие. Но это все уже выходит за пределы этой статьи.
Если мой стиль вам понравился, то в следующей статье постараюсь рассказать, как писать простые интерпретаторы на PLY и для чего их можно применять.
Using 100% of all cores with the multiprocessing module
I have two pieces of code that I’m using to learn about multiprocessing in Python 3.1. My goal is to use 100% of all the available processors. However, the code snippets here only reach 30% — 50% on all processors.
Is there anyway to ‘force’ python to use all 100%? Is the OS (windows 7, 64bit) limiting Python’s access to the processors? While the code snippets below are running, I open the task manager and watch the processor’s spike, but never reach and maintain 100%. In addition to that, I can see multiple python.exe processes created and destroyed along the way. How do these processes relate to processors? For example, if I spawn 4 processes, each process isn’t using it’s own core. Instead, what are the processes using? Are they sharing all cores? And if so, is it the OS that is forcing the processes to share the cores?
code snippet 1
code snippet 2
6 Answers 6
To use 100% of all cores, do not create and destroy new processes.
Create a few processes per core and link them with a pipeline.
At the OS-level, all pipelined processes run concurrently.
The less you write (and the more you delegate to the OS) the more likely you are to use as many resources as possible.
Will make maximal use of your CPU.
You can use psutil to pin each process spawned by multiprocessing to a specific CPU:
Note: As commented, psutil.Process.cpu_affinity is not available on macOS.
Minimum example in pure Python:
Usage: to warm up on a cold day (but feel free to change the loop to something less pointless.)
Warning: to exit, don’t pull the plug or hold the power button, Ctrl-C instead.
Regarding code snippet 1: How many cores / processors do you have on your test machine? It isn’t doing you any good to run 50 of these processes if you only have 2 CPU cores. In fact you’re forcing the OS to spend more time context switching to move processes on and off the CPU than do actual work.
Try reducing the number of spawned processes to the number of cores. So «for i in range(50):» should become something like:
Regarding code snippet 2: You’re using a multiprocessing.Lock which can only be held by a single process at a time so you’re completely limiting all the parallelism in this version of the program. You’ve serialized things so that process 1 through 50 start, a random process (say process 7) acquires the lock. Processes 1-6, and 8-50 all sit on the line:
While they sit there they are just waiting for the lock to be released. Depending on the implementation of the Lock primitive they are probably not using any CPU, they’re just sitting there using system resources like RAM but are doing no useful work with the CPU. Process 7 counts and prints to 1000 and then releases the lock. The OS then is free to schedule randomly one of the remaining 49 processes to run. Whichever one it wakes up first will acquire the lock next and run while the remaining 48 wait on the Lock. This’ll continue for the whole program.
Basically, code snippet 2 is an example of what makes concurrency hard. You have to manage access by lots of processes or threads to some shared resource. In this particular case there really is no reason that these processes need to wait on each other though.
So of these two, Snippet 1 is closer to more efficiently utilitizing the CPU. I think properly tuning the number of processes to match the number of cores will yield a much improved result.
multiprocessing — Process-based parallelism¶
This module does not work or is not available on WebAssembly platforms wasm32-emscripten and wasm32-wasi . See WebAssembly platforms for more information.
Introduction¶
multiprocessing is a package that supports spawning processes using an API similar to the threading module. The multiprocessing package offers both local and remote concurrency, effectively side-stepping the Global Interpreter Lock by using subprocesses instead of threads. Due to this, the multiprocessing module allows the programmer to fully leverage multiple processors on a given machine. It runs on both Unix and Windows.
The multiprocessing module also introduces APIs which do not have analogs in the threading module. A prime example of this is the Pool object which offers a convenient means of parallelizing the execution of a function across multiple input values, distributing the input data across processes (data parallelism). The following example demonstrates the common practice of defining such functions in a module so that child processes can successfully import that module. This basic example of data parallelism using Pool ,
will print to standard output
concurrent.futures.ProcessPoolExecutor offers a higher level interface to push tasks to a background process without blocking execution of the calling process. Compared to using the Pool interface directly, the concurrent.futures API more readily allows the submission of work to the underlying process pool to be separated from waiting for the results.
The Process class¶
In multiprocessing , processes are spawned by creating a Process object and then calling its start() method. Process follows the API of threading.Thread . A trivial example of a multiprocess program is
To show the individual process IDs involved, here is an expanded example:
For an explanation of why the if __name__ == ‘__main__’ part is necessary, see Programming guidelines .
Contexts and start methods¶
Depending on the platform, multiprocessing supports three ways to start a process. These start methods are
spawn
The parent process starts a fresh Python interpreter process. The child process will only inherit those resources necessary to run the process object’s run() method. In particular, unnecessary file descriptors and handles from the parent process will not be inherited. Starting a process using this method is rather slow compared to using fork or forkserver.
Available on Unix and Windows. The default on Windows and macOS.
fork
The parent process uses os.fork() to fork the Python interpreter. The child process, when it begins, is effectively identical to the parent process. All resources of the parent are inherited by the child process. Note that safely forking a multithreaded process is problematic.
Available on Unix only. The default on Unix.
forkserver
When the program starts and selects the forkserver start method, a server process is started. From then on, whenever a new process is needed, the parent process connects to the server and requests that it fork a new process. The fork server process is single threaded so it is safe for it to use os.fork() . No unnecessary resources are inherited.
Available on Unix platforms which support passing file descriptors over Unix pipes.
Changed in version 3.8: On macOS, the spawn start method is now the default. The fork start method should be considered unsafe as it can lead to crashes of the subprocess. See bpo-33725.
Changed in version 3.4: spawn added on all Unix platforms, and forkserver added for some Unix platforms. Child processes no longer inherit all of the parents inheritable handles on Windows.
On Unix using the spawn or forkserver start methods will also start a resource tracker process which tracks the unlinked named system resources (such as named semaphores or SharedMemory objects) created by processes of the program. When all processes have exited the resource tracker unlinks any remaining tracked object. Usually there should be none, but if a process was killed by a signal there may be some “leaked” resources. (Neither leaked semaphores nor shared memory segments will be automatically unlinked until the next reboot. This is problematic for both objects because the system allows only a limited number of named semaphores, and shared memory segments occupy some space in the main memory.)
To select a start method you use the set_start_method() in the if __name__ == ‘__main__’ clause of the main module. For example:
set_start_method() should not be used more than once in the program.
Alternatively, you can use get_context() to obtain a context object. Context objects have the same API as the multiprocessing module, and allow one to use multiple start methods in the same program.
Note that objects related to one context may not be compatible with processes for a different context. In particular, locks created using the fork context cannot be passed to processes started using the spawn or forkserver start methods.
A library which wants to use a particular start method should probably use get_context() to avoid interfering with the choice of the library user.
The ‘spawn’ and ‘forkserver’ start methods cannot currently be used with “frozen” executables (i.e., binaries produced by packages like PyInstaller and cx_Freeze) on Unix. The ‘fork’ start method does work.
Exchanging objects between processes¶
multiprocessing supports two types of communication channel between processes:
Queues
The Queue class is a near clone of queue.Queue . For example:
Queues are thread and process safe.
Pipes
The Pipe() function returns a pair of connection objects connected by a pipe which by default is duplex (two-way). For example:
The two connection objects returned by Pipe() represent the two ends of the pipe. Each connection object has send() and recv() methods (among others). Note that data in a pipe may become corrupted if two processes (or threads) try to read from or write to the same end of the pipe at the same time. Of course there is no risk of corruption from processes using different ends of the pipe at the same time.
Synchronization between processes¶
multiprocessing contains equivalents of all the synchronization primitives from threading . For instance one can use a lock to ensure that only one process prints to standard output at a time:
Without using the lock output from the different processes is liable to get all mixed up.
Sharing state between processes¶
As mentioned above, when doing concurrent programming it is usually best to avoid using shared state as far as possible. This is particularly true when using multiple processes.
However, if you really do need to use some shared data then multiprocessing provides a couple of ways of doing so.
Shared memory
Data can be stored in a shared memory map using Value or Array . For example, the following code
The ‘d’ and ‘i’ arguments used when creating num and arr are typecodes of the kind used by the array module: ‘d’ indicates a double precision float and ‘i’ indicates a signed integer. These shared objects will be process and thread-safe.
For more flexibility in using shared memory one can use the multiprocessing.sharedctypes module which supports the creation of arbitrary ctypes objects allocated from shared memory.
Server process
A manager object returned by Manager() controls a server process which holds Python objects and allows other processes to manipulate them using proxies.
A manager returned by Manager() will support types list , dict , Namespace , Lock , RLock , Semaphore , BoundedSemaphore , Condition , Event , Barrier , Queue , Value and Array . For example,
Server process managers are more flexible than using shared memory objects because they can be made to support arbitrary object types. Also, a single manager can be shared by processes on different computers over a network. They are, however, slower than using shared memory.
Using a pool of workers¶
The Pool class represents a pool of worker processes. It has methods which allows tasks to be offloaded to the worker processes in a few different ways.
Note that the methods of a pool should only ever be used by the process which created it.
Functionality within this package requires that the __main__ module be importable by the children. This is covered in Programming guidelines however it is worth pointing out here. This means that some examples, such as the multiprocessing.pool.Pool examples will not work in the interactive interpreter. For example:
(If you try this it will actually output three full tracebacks interleaved in a semi-random fashion, and then you may have to stop the parent process somehow.)
Reference¶
The multiprocessing package mostly replicates the API of the threading module.
Process and exceptions¶
Process objects represent activity that is run in a separate process. The Process class has equivalents of all the methods of threading.Thread .
The constructor should always be called with keyword arguments. group should always be None ; it exists solely for compatibility with threading.Thread . target is the callable object to be invoked by the run() method. It defaults to None , meaning nothing is called. name is the process name (see name for more details). args is the argument tuple for the target invocation. kwargs is a dictionary of keyword arguments for the target invocation. If provided, the keyword-only daemon argument sets the process daemon flag to True or False . If None (the default), this flag will be inherited from the creating process.
By default, no arguments are passed to target. The args argument, which defaults to () , can be used to specify a list or tuple of the arguments to pass to target.
If a subclass overrides the constructor, it must make sure it invokes the base class constructor ( Process.__init__() ) before doing anything else to the process.
Changed in version 3.3: Added the daemon argument.
Method representing the process’s activity.
You may override this method in a subclass. The standard run() method invokes the callable object passed to the object’s constructor as the target argument, if any, with sequential and keyword arguments taken from the args and kwargs arguments, respectively.
Using a list or tuple as the args argument passed to Process achieves the same effect.
Start the process’s activity.
This must be called at most once per process object. It arranges for the object’s run() method to be invoked in a separate process.
If the optional argument timeout is None (the default), the method blocks until the process whose join() method is called terminates. If timeout is a positive number, it blocks at most timeout seconds. Note that the method returns None if its process terminates or if the method times out. Check the process’s exitcode to determine if it terminated.
A process can be joined many times.
A process cannot join itself because this would cause a deadlock. It is an error to attempt to join a process before it has been started.
The process’s name. The name is a string used for identification purposes only. It has no semantics. Multiple processes may be given the same name.
The initial name is set by the constructor. If no explicit name is provided to the constructor, a name of the form ‘Process-N1:N2:…:Nk’ is constructed, where each Nk is the N-th child of its parent.
Return whether the process is alive.
Roughly, a process object is alive from the moment the start() method returns until the child process terminates.
The process’s daemon flag, a Boolean value. This must be set before start() is called.
The initial value is inherited from the creating process.
When a process exits, it attempts to terminate all of its daemonic child processes.
Note that a daemonic process is not allowed to create child processes. Otherwise a daemonic process would leave its children orphaned if it gets terminated when its parent process exits. Additionally, these are not Unix daemons or services, they are normal processes that will be terminated (and not joined) if non-daemonic processes have exited.
In addition to the threading.Thread API, Process objects also support the following attributes and methods:
Return the process ID. Before the process is spawned, this will be None .
The child’s exit code. This will be None if the process has not yet terminated.
If the child’s run() method returned normally, the exit code will be 0. If it terminated via sys.exit() with an integer argument N, the exit code will be N.
If the child terminated due to an exception not caught within run() , the exit code will be 1. If it was terminated by signal N, the exit code will be the negative value -N.
The process’s authentication key (a byte string).
When multiprocessing is initialized the main process is assigned a random string using os.urandom() .
When a Process object is created, it will inherit the authentication key of its parent process, although this may be changed by setting authkey to another byte string.
A numeric handle of a system object which will become “ready” when the process ends.
You can use this value if you want to wait on several events at once using multiprocessing.connection.wait() . Otherwise calling join() is simpler.
On Windows, this is an OS handle usable with the WaitForSingleObject and WaitForMultipleObjects family of API calls. On Unix, this is a file descriptor usable with primitives from the select module.
New in version 3.3.
Terminate the process. On Unix this is done using the SIGTERM signal; on Windows TerminateProcess() is used. Note that exit handlers and finally clauses, etc., will not be executed.
Note that descendant processes of the process will not be terminated – they will simply become orphaned.
If this method is used when the associated process is using a pipe or queue then the pipe or queue is liable to become corrupted and may become unusable by other process. Similarly, if the process has acquired a lock or semaphore etc. then terminating it is liable to cause other processes to deadlock.
Same as terminate() but using the SIGKILL signal on Unix.
New in version 3.7.
Close the Process object, releasing all resources associated with it. ValueError is raised if the underlying process is still running. Once close() returns successfully, most other methods and attributes of the Process object will raise ValueError .
New in version 3.7.
Note that the start() , join() , is_alive() , terminate() and exitcode methods should only be called by the process that created the process object.
Example usage of some of the methods of Process :
The base class of all multiprocessing exceptions.
exception multiprocessing. BufferTooShort ¶
Exception raised by Connection.recv_bytes_into() when the supplied buffer object is too small for the message read.
If e is an instance of BufferTooShort then e.args[0] will give the message as a byte string.
exception multiprocessing. AuthenticationError ¶
Raised when there is an authentication error.
exception multiprocessing. TimeoutError ¶
Raised by methods with a timeout when the timeout expires.
Pipes and Queues¶
When using multiple processes, one generally uses message passing for communication between processes and avoids having to use any synchronization primitives like locks.
For passing messages one can use Pipe() (for a connection between two processes) or a queue (which allows multiple producers and consumers).
The Queue , SimpleQueue and JoinableQueue types are multi-producer, multi-consumer FIFO queues modelled on the queue.Queue class in the standard library. They differ in that Queue lacks the task_done() and join() methods introduced into Python 2.5’s queue.Queue class.
If you use JoinableQueue then you must call JoinableQueue.task_done() for each task removed from the queue or else the semaphore used to count the number of unfinished tasks may eventually overflow, raising an exception.
Note that one can also create a shared queue by using a manager object – see Managers .
multiprocessing uses the usual queue.Empty and queue.Full exceptions to signal a timeout. They are not available in the multiprocessing namespace so you need to import them from queue .
When an object is put on a queue, the object is pickled and a background thread later flushes the pickled data to an underlying pipe. This has some consequences which are a little surprising, but should not cause any practical difficulties – if they really bother you then you can instead use a queue created with a manager .
After putting an object on an empty queue there may be an infinitesimal delay before the queue’s empty() method returns False and get_nowait() can return without raising queue.Empty .
If multiple processes are enqueuing objects, it is possible for the objects to be received at the other end out-of-order. However, objects enqueued by the same process will always be in the expected order with respect to each other.
If a process is killed using Process.terminate() or os.kill() while it is trying to use a Queue , then the data in the queue is likely to become corrupted. This may cause any other process to get an exception when it tries to use the queue later on.
As mentioned above, if a child process has put items on a queue (and it has not used JoinableQueue.cancel_join_thread ), then that process will not terminate until all buffered items have been flushed to the pipe.
This means that if you try joining that process you may get a deadlock unless you are sure that all items which have been put on the queue have been consumed. Similarly, if the child process is non-daemonic then the parent process may hang on exit when it tries to join all its non-daemonic children.
Note that a queue created using a manager does not have this issue. See Programming guidelines .
For an example of the usage of queues for interprocess communication see Examples .
multiprocessing. Pipe ( [ duplex ] ) ¶
Returns a pair (conn1, conn2) of Connection objects representing the ends of a pipe.
If duplex is True (the default) then the pipe is bidirectional. If duplex is False then the pipe is unidirectional: conn1 can only be used for receiving messages and conn2 can only be used for sending messages.
class multiprocessing. Queue ( [ maxsize ] ) ¶
Returns a process shared queue implemented using a pipe and a few locks/semaphores. When a process first puts an item on the queue a feeder thread is started which transfers objects from a buffer into the pipe.
The usual queue.Empty and queue.Full exceptions from the standard library’s queue module are raised to signal timeouts.
Queue implements all the methods of queue.Queue except for task_done() and join() .
Return the approximate size of the queue. Because of multithreading/multiprocessing semantics, this number is not reliable.
Note that this may raise NotImplementedError on Unix platforms like macOS where sem_getvalue() is not implemented.
Return True if the queue is empty, False otherwise. Because of multithreading/multiprocessing semantics, this is not reliable.
Return True if the queue is full, False otherwise. Because of multithreading/multiprocessing semantics, this is not reliable.
Put obj into the queue. If the optional argument block is True (the default) and timeout is None (the default), block if necessary until a free slot is available. If timeout is a positive number, it blocks at most timeout seconds and raises the queue.Full exception if no free slot was available within that time. Otherwise (block is False ), put an item on the queue if a free slot is immediately available, else raise the queue.Full exception (timeout is ignored in that case).
Changed in version 3.8: If the queue is closed, ValueError is raised instead of AssertionError .
Equivalent to put(obj, False) .
Remove and return an item from the queue. If optional args block is True (the default) and timeout is None (the default), block if necessary until an item is available. If timeout is a positive number, it blocks at most timeout seconds and raises the queue.Empty exception if no item was available within that time. Otherwise (block is False ), return an item if one is immediately available, else raise the queue.Empty exception (timeout is ignored in that case).
Changed in version 3.8: If the queue is closed, ValueError is raised instead of OSError .
Equivalent to get(False) .
multiprocessing.Queue has a few additional methods not found in queue.Queue . These methods are usually unnecessary for most code:
Indicate that no more data will be put on this queue by the current process. The background thread will quit once it has flushed all buffered data to the pipe. This is called automatically when the queue is garbage collected.
Join the background thread. This can only be used after close() has been called. It blocks until the background thread exits, ensuring that all data in the buffer has been flushed to the pipe.
By default if a process is not the creator of the queue then on exit it will attempt to join the queue’s background thread. The process can call cancel_join_thread() to make join_thread() do nothing.
Prevent join_thread() from blocking. In particular, this prevents the background thread from being joined automatically when the process exits – see join_thread() .
A better name for this method might be allow_exit_without_flush() . It is likely to cause enqueued data to be lost, and you almost certainly will not need to use it. It is really only there if you need the current process to exit immediately without waiting to flush enqueued data to the underlying pipe, and you don’t care about lost data.
This class’s functionality requires a functioning shared semaphore implementation on the host operating system. Without one, the functionality in this class will be disabled, and attempts to instantiate a Queue will result in an ImportError . See bpo-3770 for additional information. The same holds true for any of the specialized queue types listed below.
It is a simplified Queue type, very close to a locked Pipe .
Close the queue: release internal resources.
A queue must not be used anymore after it is closed. For example, get() , put() and empty() methods must no longer be called.
New in version 3.9.
Return True if the queue is empty, False otherwise.
Remove and return an item from the queue.
Put item into the queue.
class multiprocessing. JoinableQueue ( [ maxsize ] ) ¶
JoinableQueue , a Queue subclass, is a queue which additionally has task_done() and join() methods.
Indicate that a formerly enqueued task is complete. Used by queue consumers. For each get() used to fetch a task, a subsequent call to task_done() tells the queue that the processing on the task is complete.
If a join() is currently blocking, it will resume when all items have been processed (meaning that a task_done() call was received for every item that had been put() into the queue).
Raises a ValueError if called more times than there were items placed in the queue.
Block until all items in the queue have been gotten and processed.
The count of unfinished tasks goes up whenever an item is added to the queue. The count goes down whenever a consumer calls task_done() to indicate that the item was retrieved and all work on it is complete. When the count of unfinished tasks drops to zero, join() unblocks.
Miscellaneous¶
Return list of all live children of the current process.
Calling this has the side effect of “joining” any processes which have already finished.
Return the number of CPUs in the system.
This number is not equivalent to the number of CPUs the current process can use. The number of usable CPUs can be obtained with len(os.sched_getaffinity(0))
When the number of CPUs cannot be determined a NotImplementedError is raised.
Return the Process object corresponding to the current process.
Return the Process object corresponding to the parent process of the current_process() . For the main process, parent_process will be None .
New in version 3.8.
Add support for when a program which uses multiprocessing has been frozen to produce a Windows executable. (Has been tested with py2exe, PyInstaller and cx_Freeze.)
One needs to call this function straight after the if __name__ == ‘__main__’ line of the main module. For example:
If the freeze_support() line is omitted then trying to run the frozen executable will raise RuntimeError .
Calling freeze_support() has no effect when invoked on any operating system other than Windows. In addition, if the module is being run normally by the Python interpreter on Windows (the program has not been frozen), then freeze_support() has no effect.
Returns a list of the supported start methods, the first of which is the default. The possible start methods are ‘fork’ , ‘spawn’ and ‘forkserver’ . On Windows only ‘spawn’ is available. On Unix ‘fork’ and ‘spawn’ are always supported, with ‘fork’ being the default.
New in version 3.4.
Return a context object which has the same attributes as the multiprocessing module.
If method is None then the default context is returned. Otherwise method should be ‘fork’ , ‘spawn’ , ‘forkserver’ . ValueError is raised if the specified start method is not available.
New in version 3.4.
Return the name of start method used for starting processes.
If the start method has not been fixed and allow_none is false, then the start method is fixed to the default and the name is returned. If the start method has not been fixed and allow_none is true then None is returned.
The return value can be ‘fork’ , ‘spawn’ , ‘forkserver’ or None . ‘fork’ is the default on Unix, while ‘spawn’ is the default on Windows and macOS.
Changed in version 3.8: On macOS, the spawn start method is now the default. The fork start method should be considered unsafe as it can lead to crashes of the subprocess. See bpo-33725.
New in version 3.4.
Set the path of the Python interpreter to use when starting a child process. (By default sys.executable is used). Embedders will probably need to do some thing like
before they can create child processes.
Changed in version 3.4: Now supported on Unix when the ‘spawn’ start method is used.
Changed in version 3.11: Accepts a path-like object .
Set the method which should be used to start child processes. The method argument can be ‘fork’ , ‘spawn’ or ‘forkserver’ . Raises RuntimeError if the start method has already been set and force is not True . If method is None and force is True then the start method is set to None . If method is None and force is False then the context is set to the default context.
Note that this should be called at most once, and it should be protected inside the if __name__ == ‘__main__’ clause of the main module.
New in version 3.4.
Connection Objects¶
Connection objects allow the sending and receiving of picklable objects or strings. They can be thought of as message oriented connected sockets.
Connection objects are usually created using Pipe – see also Listeners and Clients .
class multiprocessing.connection. Connection ¶ send ( obj ) ¶
Send an object to the other end of the connection which should be read using recv() .
The object must be picklable. Very large pickles (approximately 32 MiB+, though it depends on the OS) may raise a ValueError exception.
Return an object sent from the other end of the connection using send() . Blocks until there is something to receive. Raises EOFError if there is nothing left to receive and the other end was closed.
Return the file descriptor or handle used by the connection.
Close the connection.
This is called automatically when the connection is garbage collected.
Return whether there is any data available to be read.
If timeout is not specified then it will return immediately. If timeout is a number then this specifies the maximum time in seconds to block. If timeout is None then an infinite timeout is used.
Note that multiple connection objects may be polled at once by using multiprocessing.connection.wait() .
Send byte data from a bytes-like object as a complete message.
If offset is given then data is read from that position in buffer. If size is given then that many bytes will be read from buffer. Very large buffers (approximately 32 MiB+, though it depends on the OS) may raise a ValueError exception
Return a complete message of byte data sent from the other end of the connection as a string. Blocks until there is something to receive. Raises EOFError if there is nothing left to receive and the other end has closed.
If maxlength is specified and the message is longer than maxlength then OSError is raised and the connection will no longer be readable.
Changed in version 3.3: This function used to raise IOError , which is now an alias of OSError .
Read into buffer a complete message of byte data sent from the other end of the connection and return the number of bytes in the message. Blocks until there is something to receive. Raises EOFError if there is nothing left to receive and the other end was closed.
buffer must be a writable bytes-like object . If offset is given then the message will be written into the buffer from that position. Offset must be a non-negative integer less than the length of buffer (in bytes).
If the buffer is too short then a BufferTooShort exception is raised and the complete message is available as e.args[0] where e is the exception instance.
Changed in version 3.3: Connection objects themselves can now be transferred between processes using Connection.send() and Connection.recv() .
New in version 3.3: Connection objects now support the context management protocol – see Context Manager Types . __enter__() returns the connection object, and __exit__() calls close() .
The Connection.recv() method automatically unpickles the data it receives, which can be a security risk unless you can trust the process which sent the message.
Therefore, unless the connection object was produced using Pipe() you should only use the recv() and send() methods after performing some sort of authentication. See Authentication keys .
If a process is killed while it is trying to read or write to a pipe then the data in the pipe is likely to become corrupted, because it may become impossible to be sure where the message boundaries lie.
Synchronization primitives¶
Generally synchronization primitives are not as necessary in a multiprocess program as they are in a multithreaded program. See the documentation for threading module.
Note that one can also create synchronization primitives by using a manager object – see Managers .
class multiprocessing. Barrier ( parties [ , action [ , timeout ] ] ) ¶
A barrier object: a clone of threading.Barrier .
New in version 3.3.
A bounded semaphore object: a close analog of threading.BoundedSemaphore .
A solitary difference from its close analog exists: its acquire method’s first argument is named block, as is consistent with Lock.acquire() .
On macOS, this is indistinguishable from Semaphore because sem_getvalue() is not implemented on that platform.
A condition variable: an alias for threading.Condition .
If lock is specified then it should be a Lock or RLock object from multiprocessing .
Changed in version 3.3: The wait_for() method was added.
class multiprocessing. Lock ¶
A non-recursive lock object: a close analog of threading.Lock . Once a process or thread has acquired a lock, subsequent attempts to acquire it from any process or thread will block until it is released; any process or thread may release it. The concepts and behaviors of threading.Lock as it applies to threads are replicated here in multiprocessing.Lock as it applies to either processes or threads, except as noted.
Note that Lock is actually a factory function which returns an instance of multiprocessing.synchronize.Lock initialized with a default context.
Lock supports the context manager protocol and thus may be used in with statements.
acquire ( block = True , timeout = None ) ¶
Acquire a lock, blocking or non-blocking.
With the block argument set to True (the default), the method call will block until the lock is in an unlocked state, then set it to locked and return True . Note that the name of this first argument differs from that in threading.Lock.acquire() .
With the block argument set to False , the method call does not block. If the lock is currently in a locked state, return False ; otherwise set the lock to a locked state and return True .
When invoked with a positive, floating-point value for timeout, block for at most the number of seconds specified by timeout as long as the lock can not be acquired. Invocations with a negative value for timeout are equivalent to a timeout of zero. Invocations with a timeout value of None (the default) set the timeout period to infinite. Note that the treatment of negative or None values for timeout differs from the implemented behavior in threading.Lock.acquire() . The timeout argument has no practical implications if the block argument is set to False and is thus ignored. Returns True if the lock has been acquired or False if the timeout period has elapsed.
Release a lock. This can be called from any process or thread, not only the process or thread which originally acquired the lock.
Behavior is the same as in threading.Lock.release() except that when invoked on an unlocked lock, a ValueError is raised.
class multiprocessing. RLock ¶
A recursive lock object: a close analog of threading.RLock . A recursive lock must be released by the process or thread that acquired it. Once a process or thread has acquired a recursive lock, the same process or thread may acquire it again without blocking; that process or thread must release it once for each time it has been acquired.
Note that RLock is actually a factory function which returns an instance of multiprocessing.synchronize.RLock initialized with a default context.
RLock supports the context manager protocol and thus may be used in with statements.
acquire ( block = True , timeout = None ) ¶
Acquire a lock, blocking or non-blocking.
When invoked with the block argument set to True , block until the lock is in an unlocked state (not owned by any process or thread) unless the lock is already owned by the current process or thread. The current process or thread then takes ownership of the lock (if it does not already have ownership) and the recursion level inside the lock increments by one, resulting in a return value of True . Note that there are several differences in this first argument’s behavior compared to the implementation of threading.RLock.acquire() , starting with the name of the argument itself.
When invoked with the block argument set to False , do not block. If the lock has already been acquired (and thus is owned) by another process or thread, the current process or thread does not take ownership and the recursion level within the lock is not changed, resulting in a return value of False . If the lock is in an unlocked state, the current process or thread takes ownership and the recursion level is incremented, resulting in a return value of True .
Use and behaviors of the timeout argument are the same as in Lock.acquire() . Note that some of these behaviors of timeout differ from the implemented behaviors in threading.RLock.acquire() .
Release a lock, decrementing the recursion level. If after the decrement the recursion level is zero, reset the lock to unlocked (not owned by any process or thread) and if any other processes or threads are blocked waiting for the lock to become unlocked, allow exactly one of them to proceed. If after the decrement the recursion level is still nonzero, the lock remains locked and owned by the calling process or thread.
Only call this method when the calling process or thread owns the lock. An AssertionError is raised if this method is called by a process or thread other than the owner or if the lock is in an unlocked (unowned) state. Note that the type of exception raised in this situation differs from the implemented behavior in threading.RLock.release() .
class multiprocessing. Semaphore ( [ value ] ) ¶
A semaphore object: a close analog of threading.Semaphore .
A solitary difference from its close analog exists: its acquire method’s first argument is named block, as is consistent with Lock.acquire() .
On macOS, sem_timedwait is unsupported, so calling acquire() with a timeout will emulate that function’s behavior using a sleeping loop.
If the SIGINT signal generated by Ctrl — C arrives while the main thread is blocked by a call to BoundedSemaphore.acquire() , Lock.acquire() , RLock.acquire() , Semaphore.acquire() , Condition.acquire() or Condition.wait() then the call will be immediately interrupted and KeyboardInterrupt will be raised.
This differs from the behaviour of threading where SIGINT will be ignored while the equivalent blocking calls are in progress.
Some of this package’s functionality requires a functioning shared semaphore implementation on the host operating system. Without one, the multiprocessing.synchronize module will be disabled, and attempts to import it will result in an ImportError . See bpo-3770 for additional information.
Shared ctypes Objects¶
It is possible to create shared objects using shared memory which can be inherited by child processes.
multiprocessing. Value ( typecode_or_type , * args , lock = True ) ¶
Return a ctypes object allocated from shared memory. By default the return value is actually a synchronized wrapper for the object. The object itself can be accessed via the value attribute of a Value .
typecode_or_type determines the type of the returned object: it is either a ctypes type or a one character typecode of the kind used by the array module. *args is passed on to the constructor for the type.
If lock is True (the default) then a new recursive lock object is created to synchronize access to the value. If lock is a Lock or RLock object then that will be used to synchronize access to the value. If lock is False then access to the returned object will not be automatically protected by a lock, so it will not necessarily be “process-safe”.
Operations like += which involve a read and write are not atomic. So if, for instance, you want to atomically increment a shared value it is insufficient to just do
Assuming the associated lock is recursive (which it is by default) you can instead do
Note that lock is a keyword-only argument.
multiprocessing. Array ( typecode_or_type , size_or_initializer , * , lock = True ) ¶
Return a ctypes array allocated from shared memory. By default the return value is actually a synchronized wrapper for the array.
typecode_or_type determines the type of the elements of the returned array: it is either a ctypes type or a one character typecode of the kind used by the array module. If size_or_initializer is an integer, then it determines the length of the array, and the array will be initially zeroed. Otherwise, size_or_initializer is a sequence which is used to initialize the array and whose length determines the length of the array.
If lock is True (the default) then a new lock object is created to synchronize access to the value. If lock is a Lock or RLock object then that will be used to synchronize access to the value. If lock is False then access to the returned object will not be automatically protected by a lock, so it will not necessarily be “process-safe”.
Note that lock is a keyword only argument.
Note that an array of ctypes.c_char has value and raw attributes which allow one to use it to store and retrieve strings.
The multiprocessing.sharedctypes module¶
The multiprocessing.sharedctypes module provides functions for allocating ctypes objects from shared memory which can be inherited by child processes.
Although it is possible to store a pointer in shared memory remember that this will refer to a location in the address space of a specific process. However, the pointer is quite likely to be invalid in the context of a second process and trying to dereference the pointer from the second process may cause a crash.
Return a ctypes array allocated from shared memory.
typecode_or_type determines the type of the elements of the returned array: it is either a ctypes type or a one character typecode of the kind used by the array module. If size_or_initializer is an integer then it determines the length of the array, and the array will be initially zeroed. Otherwise size_or_initializer is a sequence which is used to initialize the array and whose length determines the length of the array.
Note that setting and getting an element is potentially non-atomic – use Array() instead to make sure that access is automatically synchronized using a lock.
multiprocessing.sharedctypes. RawValue ( typecode_or_type , * args ) ¶
Return a ctypes object allocated from shared memory.
typecode_or_type determines the type of the returned object: it is either a ctypes type or a one character typecode of the kind used by the array module. *args is passed on to the constructor for the type.
Note that setting and getting the value is potentially non-atomic – use Value() instead to make sure that access is automatically synchronized using a lock.
Note that an array of ctypes.c_char has value and raw attributes which allow one to use it to store and retrieve strings – see documentation for ctypes .
multiprocessing.sharedctypes. Array ( typecode_or_type , size_or_initializer , * , lock = True ) ¶
The same as RawArray() except that depending on the value of lock a process-safe synchronization wrapper may be returned instead of a raw ctypes array.
If lock is True (the default) then a new lock object is created to synchronize access to the value. If lock is a Lock or RLock object then that will be used to synchronize access to the value. If lock is False then access to the returned object will not be automatically protected by a lock, so it will not necessarily be “process-safe”.
Note that lock is a keyword-only argument.
multiprocessing.sharedctypes. Value ( typecode_or_type , * args , lock = True ) ¶
The same as RawValue() except that depending on the value of lock a process-safe synchronization wrapper may be returned instead of a raw ctypes object.
If lock is True (the default) then a new lock object is created to synchronize access to the value. If lock is a Lock or RLock object then that will be used to synchronize access to the value. If lock is False then access to the returned object will not be automatically protected by a lock, so it will not necessarily be “process-safe”.
Note that lock is a keyword-only argument.
multiprocessing.sharedctypes. copy ( obj ) ¶
Return a ctypes object allocated from shared memory which is a copy of the ctypes object obj.
multiprocessing.sharedctypes. synchronized ( obj [ , lock ] ) ¶
Return a process-safe wrapper object for a ctypes object which uses lock to synchronize access. If lock is None (the default) then a multiprocessing.RLock object is created automatically.
A synchronized wrapper will have two methods in addition to those of the object it wraps: get_obj() returns the wrapped object and get_lock() returns the lock object used for synchronization.
Note that accessing the ctypes object through the wrapper can be a lot slower than accessing the raw ctypes object.
Changed in version 3.5: Synchronized objects support the context manager protocol.