multiprocessing.shared_memory --- Shared memory for direct access across processes

ソースコード Lib/multiprocessing/shared_memory.py

バージョン 3.8 で追加.


This module provides a class, SharedMemory, for the allocation and management of shared memory to be accessed by one or more processes on a multicore or symmetric multiprocessor (SMP) machine. To assist with the life-cycle management of shared memory especially across distinct processes, a BaseManager subclass, SharedMemoryManager, is also provided in the multiprocessing.managers module.

In this module, shared memory refers to "POSIX style" shared memory blocks (though is not necessarily implemented explicitly as such) and does not refer to "distributed shared memory". This style of shared memory permits distinct processes to potentially read and write to a common (or shared) region of volatile memory. Processes are conventionally limited to only have access to their own process memory space but shared memory permits the sharing of data between processes, avoiding the need to instead send messages between processes containing that data. Sharing data directly via memory can provide significant performance benefits compared to sharing data via disk or socket or other communications requiring the serialization/deserialization and copying of data.

class multiprocessing.shared_memory.SharedMemory(name=None, create=False, size=0, *, track=True)

Create an instance of the SharedMemory class for either creating a new shared memory block or attaching to an existing shared memory block. Each shared memory block is assigned a unique name. In this way, one process can create a shared memory block with a particular name and a different process can attach to that same shared memory block using that same name.

As a resource for sharing data across processes, shared memory blocks may outlive the original process that created them. When one process no longer needs access to a shared memory block that might still be needed by other processes, the close() method should be called. When a shared memory block is no longer needed by any process, the unlink() method should be called to ensure proper cleanup.

パラメータ:
  • name (str | None) -- The unique name for the requested shared memory, specified as a string. When creating a new shared memory block, if None (the default) is supplied for the name, a novel name will be generated.

  • create (bool) -- Control whether a new shared memory block is created (True) or an existing shared memory block is attached (False).

  • size (int) -- The requested number of bytes when creating a new shared memory block. Because some platforms choose to allocate chunks of memory based upon that platform's memory page size, the exact size of the shared memory block may be larger or equal to the size requested. When attaching to an existing shared memory block, the size parameter is ignored.

  • track (bool) -- When True, register the shared memory block with a resource tracker process on platforms where the OS does not do this automatically. The resource tracker ensures proper cleanup of the shared memory even if all other processes with access to the memory exit without doing so. Python processes created from a common ancestor using multiprocessing facilities share a single resource tracker process, and the lifetime of shared memory segments is handled automatically among these processes. Python processes created in any other way will receive their own resource tracker when accessing shared memory with track enabled. This will cause the shared memory to be deleted by the resource tracker of the first process that terminates. To avoid this issue, users of subprocess or standalone Python processes should set track to False when there is already another process in place that does the bookkeeping. track is ignored on Windows, which has its own tracking and automatically deletes shared memory when all handles to it have been closed.

バージョン 3.13 で変更: Added the track parameter.

close()

Close the file descriptor/handle to the shared memory from this instance. close() should be called once access to the shared memory block from this instance is no longer needed. Depending on operating system, the underlying memory may or may not be freed even if all handles to it have been closed. To ensure proper cleanup, use the unlink() method.

Delete the underlying shared memory block. This should be called only once per shared memory block regardless of the number of handles to it, even in other processes. unlink() and close() can be called in any order, but trying to access data inside a shared memory block after unlink() may result in memory access errors, depending on platform.

This method has no effect on Windows, where the only way to delete a shared memory block is to close all handles.

buf

A memoryview of contents of the shared memory block.

name

Read-only access to the unique name of the shared memory block.

size

Read-only access to size in bytes of the shared memory block.

The following example demonstrates low-level use of SharedMemory instances:

>>> from multiprocessing import shared_memory
>>> shm_a = shared_memory.SharedMemory(create=True, size=10)
>>> type(shm_a.buf)
<class 'memoryview'>
>>> buffer = shm_a.buf
>>> len(buffer)
10
>>> buffer[:4] = bytearray([22, 33, 44, 55])  # Modify multiple at once
>>> buffer[4] = 100                           # Modify single byte at a time
>>> # Attach to an existing shared memory block
>>> shm_b = shared_memory.SharedMemory(shm_a.name)
>>> import array
>>> array.array('b', shm_b.buf[:5])  # Copy the data into a new array.array
array('b', [22, 33, 44, 55, 100])
>>> shm_b.buf[:5] = b'howdy'  # Modify via shm_b using bytes
>>> bytes(shm_a.buf[:5])      # Access via shm_a
b'howdy'
>>> shm_b.close()   # Close each SharedMemory instance
>>> shm_a.close()
>>> shm_a.unlink()  # Call unlink only once to release the shared memory

The following example demonstrates a practical use of the SharedMemory class with NumPy arrays, accessing the same numpy.ndarray from two distinct Python shells:

>>> # In the first Python interactive shell
>>> import numpy as np
>>> a = np.array([1, 1, 2, 3, 5, 8])  # Start with an existing NumPy array
>>> from multiprocessing import shared_memory
>>> shm = shared_memory.SharedMemory(create=True, size=a.nbytes)
>>> # Now create a NumPy array backed by shared memory
>>> b = np.ndarray(a.shape, dtype=a.dtype, buffer=shm.buf)
>>> b[:] = a[:]  # Copy the original data into shared memory
>>> b
array([1, 1, 2, 3, 5, 8])
>>> type(b)
<class 'numpy.ndarray'>
>>> type(a)
<class 'numpy.ndarray'>
>>> shm.name  # We did not specify a name so one was chosen for us
'psm_21467_46075'

>>> # In either the same shell or a new Python shell on the same machine
>>> import numpy as np
>>> from multiprocessing import shared_memory
>>> # Attach to the existing shared memory block
>>> existing_shm = shared_memory.SharedMemory(name='psm_21467_46075')
>>> # Note that a.shape is (6,) and a.dtype is np.int64 in this example
>>> c = np.ndarray((6,), dtype=np.int64, buffer=existing_shm.buf)
>>> c
array([1, 1, 2, 3, 5, 8])
>>> c[-1] = 888
>>> c
array([  1,   1,   2,   3,   5, 888])

>>> # Back in the first Python interactive shell, b reflects this change
>>> b
array([  1,   1,   2,   3,   5, 888])

>>> # Clean up from within the second Python shell
>>> del c  # Unnecessary; merely emphasizing the array is no longer used
>>> existing_shm.close()

>>> # Clean up from within the first Python shell
>>> del b  # Unnecessary; merely emphasizing the array is no longer used
>>> shm.close()
>>> shm.unlink()  # Free and release the shared memory block at the very end
class multiprocessing.managers.SharedMemoryManager([address[, authkey]])

A subclass of multiprocessing.managers.BaseManager which can be used for the management of shared memory blocks across processes.

A call to start() on a SharedMemoryManager instance causes a new process to be started. This new process's sole purpose is to manage the life cycle of all shared memory blocks created through it. To trigger the release of all shared memory blocks managed by that process, call shutdown() on the instance. This triggers a unlink() call on all of the SharedMemory objects managed by that process and then stops the process itself. By creating SharedMemory instances through a SharedMemoryManager, we avoid the need to manually track and trigger the freeing of shared memory resources.

This class provides methods for creating and returning SharedMemory instances and for creating a list-like object (ShareableList) backed by shared memory.

Refer to BaseManager for a description of the inherited address and authkey optional input arguments and how they may be used to connect to an existing SharedMemoryManager service from other processes.

SharedMemory(size)

Create and return a new SharedMemory object with the specified size in bytes.

ShareableList(sequence)

Create and return a new ShareableList object, initialized by the values from the input sequence.

The following example demonstrates the basic mechanisms of a SharedMemoryManager:

>>> from multiprocessing.managers import SharedMemoryManager
>>> smm = SharedMemoryManager()
>>> smm.start()  # Start the process that manages the shared memory blocks
>>> sl = smm.ShareableList(range(4))
>>> sl
ShareableList([0, 1, 2, 3], name='psm_6572_7512')
>>> raw_shm = smm.SharedMemory(size=128)
>>> another_sl = smm.ShareableList('alpha')
>>> another_sl
ShareableList(['a', 'l', 'p', 'h', 'a'], name='psm_6572_12221')
>>> smm.shutdown()  # Calls unlink() on sl, raw_shm, and another_sl

The following example depicts a potentially more convenient pattern for using SharedMemoryManager objects via the with statement to ensure that all shared memory blocks are released after they are no longer needed:

>>> with SharedMemoryManager() as smm:
...     sl = smm.ShareableList(range(2000))
...     # Divide the work among two processes, storing partial results in sl
...     p1 = Process(target=do_work, args=(sl, 0, 1000))
...     p2 = Process(target=do_work, args=(sl, 1000, 2000))
...     p1.start()
...     p2.start()  # A multiprocessing.Pool might be more efficient
...     p1.join()
...     p2.join()   # Wait for all work to complete in both processes
...     total_result = sum(sl)  # Consolidate the partial results now in sl

When using a SharedMemoryManager in a with statement, the shared memory blocks created using that manager are all released when the with statement's code block finishes execution.

class multiprocessing.shared_memory.ShareableList(sequence=None, *, name=None)

Provide a mutable list-like object where all values stored within are stored in a shared memory block. This constrains storable values to the following built-in data types:

  • int (signed 64-bit)

  • float

  • bool

  • str (less than 10M bytes each when encoded as UTF-8)

  • bytes (less than 10M bytes each)

  • None

It also notably differs from the built-in list type in that these lists can not change their overall length (i.e. no append(), insert(), etc.) and do not support the dynamic creation of new ShareableList instances via slicing.

sequence is used in populating a new ShareableList full of values. Set to None to instead attach to an already existing ShareableList by its unique shared memory name.

name is the unique name for the requested shared memory, as described in the definition for SharedMemory. When attaching to an existing ShareableList, specify its shared memory block's unique name while leaving sequence set to None.

注釈

A known issue exists for bytes and str values. If they end with \x00 nul bytes or characters, those may be silently stripped when fetching them by index from the ShareableList. This .rstrip(b'\x00') behavior is considered a bug and may go away in the future. See gh-106939.

For applications where rstripping of trailing nulls is a problem, work around it by always unconditionally appending an extra non-0 byte to the end of such values when storing and unconditionally removing it when fetching:

>>> from multiprocessing import shared_memory
>>> nul_bug_demo = shared_memory.ShareableList(['?\x00', b'\x03\x02\x01\x00\x00\x00'])
>>> nul_bug_demo[0]
'?'
>>> nul_bug_demo[1]
b'\x03\x02\x01'
>>> nul_bug_demo.shm.unlink()
>>> padded = shared_memory.ShareableList(['?\x00\x07', b'\x03\x02\x01\x00\x00\x00\x07'])
>>> padded[0][:-1]
'?\x00'
>>> padded[1][:-1]
b'\x03\x02\x01\x00\x00\x00'
>>> padded.shm.unlink()
count(value)

Return the number of occurrences of value.

index(value)

Return first index position of value. Raise ValueError if value is not present.

format

Read-only attribute containing the struct packing format used by all currently stored values.

shm

The SharedMemory instance where the values are stored.

以下に ShareableList インスタンスの利用例を示します。

>>> from multiprocessing import shared_memory
>>> a = shared_memory.ShareableList(['howdy', b'HoWdY', -273.154, 100, None, True, 42])
>>> [ type(entry) for entry in a ]
[<class 'str'>, <class 'bytes'>, <class 'float'>, <class 'int'>, <class 'NoneType'>, <class 'bool'>, <class 'int'>]
>>> a[2]
-273.154
>>> a[2] = -78.5
>>> a[2]
-78.5
>>> a[2] = 'dry ice'  # Changing data types is supported as well
>>> a[2]
'dry ice'
>>> a[2] = 'larger than previously allocated storage space'
Traceback (most recent call last):
  ...
ValueError: exceeds available storage for existing str
>>> a[2]
'dry ice'
>>> len(a)
7
>>> a.index(42)
6
>>> a.count(b'howdy')
0
>>> a.count(b'HoWdY')
1
>>> a.shm.close()
>>> a.shm.unlink()
>>> del a  # Use of a ShareableList after call to unlink() is unsupported

The following example depicts how one, two, or many processes may access the same ShareableList by supplying the name of the shared memory block behind it:

>>> b = shared_memory.ShareableList(range(5))         # In a first process
>>> c = shared_memory.ShareableList(name=b.shm.name)  # In a second process
>>> c
ShareableList([0, 1, 2, 3, 4], name='...')
>>> c[-1] = -999
>>> b[-1]
-999
>>> b.shm.close()
>>> c.shm.close()
>>> c.shm.unlink()

The following examples demonstrates that ShareableList (and underlying SharedMemory) objects can be pickled and unpickled if needed. Note, that it will still be the same shared object. This happens, because the deserialized object has the same unique name and is just attached to an existing object with the same name (if the object is still alive):

>>> import pickle
>>> from multiprocessing import shared_memory
>>> sl = shared_memory.ShareableList(range(10))
>>> list(sl)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> deserialized_sl = pickle.loads(pickle.dumps(sl))
>>> list(deserialized_sl)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> sl[0] = -1
>>> deserialized_sl[1] = -2
>>> list(sl)
[-1, -2, 2, 3, 4, 5, 6, 7, 8, 9]
>>> list(deserialized_sl)
[-1, -2, 2, 3, 4, 5, 6, 7, 8, 9]
>>> sl.shm.close()
>>> sl.shm.unlink()