Overview

MPI for Python provides an object oriented approach to message passing which grounds on the standard MPI-2 C++ bindings. The interface was designed with focus in translating MPI syntax and semantics of standard MPI-2 bindings for C++ to Python. Any user of the standard C/C++ MPI bindings should be able to use this module without need of learning a new interface.

Communicating Python Objects and Array Data

The Python standard library supports different mechanisms for data persistence. Many of them rely on disk storage, but pickling and marshaling can also work with memory buffers.

The pickle modules provide user-extensible facilities to serialize general Python objects using ASCII or binary formats. The marshal module provides facilities to serialize built-in Python objects using a binary format specific to Python, but independent of machine architecture issues.

MPI for Python can communicate any built-in or user-defined Python object taking advantage of the features provided by the pickle module. These facilities will be routinely used to build binary representations of objects to communicate (at sending processes), and restoring them back (at receiving processes).

Although simple and general, the serialization approach (i.e., pickling and unpickling) previously discussed imposes important overheads in memory as well as processor usage, especially in the scenario of objects with large memory footprints being communicated. Pickling general Python objects, ranging from primitive or container built-in types to user-defined classes, necessarily requires computer resources. Processing is also needed for dispatching the appropriate serialization method (that depends on the type of the object) and doing the actual packing. Additional memory is always needed, and if its total amount is not known a priori, many reallocations can occur. Indeed, in the case of large numeric arrays, this is certainly unacceptable and precludes communication of objects occupying half or more of the available memory resources.

MPI for Python supports direct communication of any object exporting the single-segment buffer interface. This interface is a standard Python mechanism provided by some types (e.g., strings and numeric arrays), allowing access in the C side to a contiguous memory buffer (i.e., address and length) containing the relevant data. This feature, in conjunction with the capability of constructing user-defined MPI datatypes describing complicated memory layouts, enables the implementation of many algorithms involving multidimensional numeric arrays (e.g., image processing, fast Fourier transforms, finite difference schemes on structured Cartesian grids) directly in Python, with negligible overhead, and almost as fast as compiled Fortran, C, or C++ codes.

Communicators

In MPI for Python, MPI.Comm is the base class of communicators. The MPI.Intracomm and MPI.Intercomm classes are sublcasses of the MPI.Comm class. The MPI.Comm.Is_inter() method (and MPI.Comm.Is_intra(), provided for convenience but not part of the MPI specification) is defined for communicator objects and can be used to determine the particular communicator class.

The two predefined intracommunicator instances are available: MPI.COMM_SELF and MPI.COMM_WORLD. From them, new communicators can be created as needed.

The number of processes in a communicator and the calling process rank can be respectively obtained with methods MPI.Comm.Get_size() and MPI.Comm.Get_rank(). The associated process group can be retrieved from a communicator by calling the MPI.Comm.Get_group() method, which returns an instance of the MPI.Group class. Set operations with MPI.Group objects like like MPI.Group.Union(), MPI.Group.Intersect() and MPI.Group.Difference() are fully supported, as well as the creation of new communicators from these groups using MPI.Comm.Create() and MPI.Comm.Create_group().

New communicator instances can be obtained with the MPI.Comm.Clone(), MPI.Comm.Dup() and MPI.Comm.Split() methods, as well methods MPI.Intracomm.Create_intercomm() and MPI.Intercomm.Merge().

Virtual topologies (MPI.Cartcomm, MPI.Graphcomm and MPI.Distgraphcomm classes, which are specializations of the MPI.Intracomm class) are fully supported. New instances can be obtained from intracommunicator instances with factory methods MPI.Intracomm.Create_cart() and MPI.Intracomm.Create_graph().

Point-to-Point Communications

Point to point communication is a fundamental capability of message passing systems. This mechanism enables the transmission of data between a pair of processes, one side sending, the other receiving.

MPI provides a set of send and receive functions allowing the communication of typed data with an associated tag. The type information enables the conversion of data representation from one architecture to another in the case of heterogeneous computing environments; additionally, it allows the representation of non-contiguous data layouts and user-defined datatypes, thus avoiding the overhead of (otherwise unavoidable) packing/unpacking operations. The tag information allows selectivity of messages at the receiving end.

Blocking Communications

MPI provides basic send and receive functions that are blocking. These functions block the caller until the data buffers involved in the communication can be safely reused by the application program.

In MPI for Python, the MPI.Comm.Send(), MPI.Comm.Recv() and MPI.Comm.Sendrecv() methods of communicator objects provide support for blocking point-to-point communications within MPI.Intracomm and MPI.Intercomm instances. These methods can communicate memory buffers. The variants MPI.Comm.send(), MPI.Comm.recv() and MPI.Comm.sendrecv() can communicate general Python objects.

Nonblocking Communications

On many systems, performance can be significantly increased by overlapping communication and computation. This is particularly true on systems where communication can be executed autonomously by an intelligent, dedicated communication controller.

MPI provides nonblocking send and receive functions. They allow the possible overlap of communication and computation. Non-blocking communication always come in two parts: posting functions, which begin the requested operation; and test-for-completion functions, which allow to discover whether the requested operation has completed.

In MPI for Python, the MPI.Comm.Isend() and MPI.Comm.Irecv() methods initiate send and receive operations, respectively. These methods return a MPI.Request instance, uniquely identifying the started operation. Its completion can be managed using the MPI.Request.Test(), MPI.Request.Wait() and MPI.Request.Cancel() methods. The management of MPI.Request objects and associated memory buffers involved in communication requires a careful, rather low-level coordination. Users must ensure that objects exposing their memory buffers are not accessed at the Python level while they are involved in nonblocking message-passing operations.

Persistent Communications

Often a communication with the same argument list is repeatedly executed within an inner loop. In such cases, communication can be further optimized by using persistent communication, a particular case of nonblocking communication allowing the reduction of the overhead between processes and communication controllers. Furthermore , this kind of optimization can also alleviate the extra call overheads associated to interpreted, dynamic languages like Python.

In MPI for Python, the MPI.Comm.Send_init() and MPI.Comm.Recv_init() methods create persistent requests for a send and receive operation, respectively. These methods return an instance of the MPI.Prequest class, a subclass of the MPI.Request class. The actual communication can be effectively started using the MPI.Prequest.Start() method, and its completion can be managed as previously described.

Collective Communications

Collective communications allow the transmittal of data between multiple processes of a group simultaneously. The syntax and semantics of collective functions is consistent with point-to-point communication. Collective functions communicate typed data, but messages are not paired with an associated tag; selectivity of messages is implied in the calling order. Additionally, collective functions come in blocking versions only.

The more commonly used collective communication operations are the following.

  • Barrier synchronization across all group members.
  • Global communication functions
    • Broadcast data from one member to all members of a group.
    • Gather data from all members to one member of a group.
    • Scatter data from one member to all members of a group.
  • Global reduction operations such as sum, maximum, minimum, etc.

In MPI for Python, the MPI.Comm.Bcast(), MPI.Comm.Scatter(), MPI.Comm.Gather(), MPI.Comm.Allgather(), and MPI.Comm.Alltoall() MPI.Comm.Alltoallw() methods provide support for collective communications of memory buffers. The lower-case variants MPI.Comm.bcast(), MPI.Comm.scatter(), MPI.Comm.gather(), MPI.Comm.allgather() and MPI.Comm.alltoall() can communicate general Python objects. The vector variants (which can communicate different amounts of data to each process) MPI.Comm.Scatterv(), MPI.Comm.Gatherv(), MPI.Comm.Allgatherv(), MPI.Comm.Alltoallv() and MPI.Comm.Alltoallw() are also supported, they can only communicate objects exposing memory buffers.

Global reduction operations on memory buffers are accessible through the MPI.Comm.Reduce(), MPI.Comm.Reduce_scatter, MPI.Comm.Allreduce(), MPI.Intracomm.Scan() and MPI.Intracomm.Exscan() methods. The lower-case variants MPI.Comm.reduce(), MPI.Comm.allreduce(), MPI.Intracomm.scan() and MPI.Intracomm.exscan() can communicate general Python objects; however, the actual required reduction computations are performed sequentially at some process. All the predefined (i.e., MPI.SUM, MPI.PROD, MPI.MAX, etc.) reduction operations can be applied.

Support for CUDA-aware MPI

Several MPI implementations, including Open MPI and MVAPICH, support passing CUDA GPU pointers to MPI calls to avoid explict data movement between the host and the device. On the Python side, CUDA GPU arrays have been implemented by many libraries that need GPU computation, such as CuPy, Numba, PyTorch, and PyArrow. In order to increase library interoperability, a __cuda_array_interface__ attribute is defined and agreed upon. For example, a CuPy array can be passed to a Numba CUDA-jit kernel.

MPI for Python provides an experimental support for CUDA-aware MPI. This feature requires:

  1. mpi4py is built against a CUDA-aware MPI library.
  2. The Python GPU arrays are compliant with the __cuda_array_interface__ standard.

See the Tutorial section for further information. We note that

  • Whether or not a MPI call can work for GPU arrays depends on the underlying MPI implementation, not on mpi4py.
  • This support is currently experimental (so is the __cuda_array_interface__ standard) and subject to change in the future.

Dynamic Process Management

In the context of the MPI-1 specification, a parallel application is static; that is, no processes can be added to or deleted from a running application after it has been started. Fortunately, this limitation was addressed in MPI-2. The new specification added a process management model providing a basic interface between an application and external resources and process managers.

This MPI-2 extension can be really useful, especially for sequential applications built on top of parallel modules, or parallel applications with a client/server model. The MPI-2 process model provides a mechanism to create new processes and establish communication between them and the existing MPI application. It also provides mechanisms to establish communication between two existing MPI applications, even when one did not start the other.

In MPI for Python, new independent process groups can be created by calling the MPI.Intracomm.Spawn() method within an intracommunicator. This call returns a new intercommunicator (i.e., an MPI.Intercomm instance) at the parent process group. The child process group can retrieve the matching intercommunicator by calling the MPI.Comm.Get_parent() class method. At each side, the new intercommunicator can be used to perform point to point and collective communications between the parent and child groups of processes.

Alternatively, disjoint groups of processes can establish communication using a client/server approach. Any server application must first call the MPI.Open_port() function to open a port and the MPI.Publish_name() function to publish a provided service, and next call the MPI.Intracomm.Accept() method. Any client applications can first find a published service by calling the MPI.Lookup_name() function, which returns the port where a server can be contacted; and next call the MPI.Intracomm.Connect() method. Both MPI.Intracomm.Accept() and MPI.Intracomm.Connect() methods return an MPI.Intercomm instance. When connection between client/server processes is no longer needed, all of them must cooperatively call the MPI.Comm.Disconnect() method. Additionally, server applications should release resources by calling the MPI.Unpublish_name() and MPI.Close_port() functions.

One-Sided Communications

One-sided communications (also called Remote Memory Access, RMA) supplements the traditional two-sided, send/receive based MPI communication model with a one-sided, put/get based interface. One-sided communication that can take advantage of the capabilities of highly specialized network hardware. Additionally, this extension lowers latency and software overhead in applications written using a shared-memory-like paradigm.

The MPI specification revolves around the use of objects called windows; they intuitively specify regions of a process’s memory that have been made available for remote read and write operations. The published memory blocks can be accessed through three functions for put (remote send), get (remote write), and accumulate (remote update or reduction) data items. A much larger number of functions support different synchronization styles; the semantics of these synchronization operations are fairly complex.

In MPI for Python, one-sided operations are available by using instances of the MPI.Win class. New window objects are created by calling the MPI.Win.Create() method at all processes within a communicator and specifying a memory buffer . When a window instance is no longer needed, the MPI.Win.Free() method should be called.

The three one-sided MPI operations for remote write, read and reduction are available through calling the methods MPI.Win.Put(), MPI.Win.Get(), and MPI.Win.Accumulate() respectively within a Win instance. These methods need an integer rank identifying the target process and an integer offset relative the base address of the remote memory block being accessed.

The one-sided operations read, write, and reduction are implicitly nonblocking, and must be synchronized by using two primary modes. Active target synchronization requires the origin process to call the MPI.Win.Start() and MPI.Win.Complete() methods at the origin process, and target process cooperates by calling the MPI.Win.Post() and MPI.Win.Wait() methods. There is also a collective variant provided by the MPI.Win.Fence() method. Passive target synchronization is more lenient, only the origin process calls the MPI.Win.Lock() and MPI.Win.Unlock() methods. Locks are used to protect remote accesses to the locked remote window and to protect local load/store accesses to a locked local window.

Parallel Input/Output

The POSIX standard provides a model of a widely portable file system. However, the optimization needed for parallel input/output cannot be achieved with this generic interface. In order to ensure efficiency and scalability, the underlying parallel input/output system must provide a high-level interface supporting partitioning of file data among processes and a collective interface supporting complete transfers of global data structures between process memories and files. Additionally, further efficiencies can be gained via support for asynchronous input/output, strided accesses to data, and control over physical file layout on storage devices. This scenario motivated the inclusion in the MPI-2 standard of a custom interface in order to support more elaborated parallel input/output operations.

The MPI specification for parallel input/output revolves around the use objects called files. As defined by MPI, files are not just contiguous byte streams. Instead, they are regarded as ordered collections of typed data items. MPI supports sequential or random access to any integral set of these items. Furthermore, files are opened collectively by a group of processes.

The common patterns for accessing a shared file (broadcast, scatter, gather, reduction) is expressed by using user-defined datatypes. Compared to the communication patterns of point-to-point and collective communications, this approach has the advantage of added flexibility and expressiveness. Data access operations (read and write) are defined for different kinds of positioning (using explicit offsets, individual file pointers, and shared file pointers), coordination (non-collective and collective), and synchronism (blocking, nonblocking, and split collective with begin/end phases).

In MPI for Python, all MPI input/output operations are performed through instances of the MPI.File class. File handles are obtained by calling the MPI.File.Open() method at all processes within a communicator and providing a file name and the intended access mode. After use, they must be closed by calling the MPI.File.Close() method. Files even can be deleted by calling method MPI.File.Delete().

After creation, files are typically associated with a per-process view. The view defines the current set of data visible and accessible from an open file as an ordered set of elementary datatypes. This data layout can be set and queried with the MPI.File.Set_view() and MPI.File.Get_view() methods respectively.

Actual input/output operations are achieved by many methods combining read and write calls with different behavior regarding positioning, coordination, and synchronism. Summing up, MPI for Python provides the thirty (30) methods defined in MPI-2 for reading from or writing to files using explicit offsets or file pointers (individual or shared), in blocking or nonblocking and collective or noncollective versions.

Environmental Management

Initialization and Exit

Module functions MPI.Init() or MPI.Init_thread() and MPI.Finalize() provide MPI initialization and finalization respectively. Module functions MPI.Is_initialized() and MPI.Is_finalized() provide the respective tests for initialization and finalization.

Note

MPI_Init() or MPI_Init_thread() is actually called when you import the MPI module from the mpi4py package, but only if MPI is not already initialized. In such case, calling MPI.Init() or MPI.Init_thread() from Python is expected to generate an MPI error, and in turn an exception will be raised.

Note

MPI_Finalize() is registered (by using Python C/API function Py_AtExit()) for being automatically called when Python processes exit, but only if mpi4py actually initialized MPI. Therefore, there is no need to call MPI.Finalize() from Python to ensure MPI finalization.

Implementation Information

  • The MPI version number can be retrieved from module function MPI.Get_version(). It returns a two-integer tuple (version,subversion).
  • The MPI.Get_processor_name() function can be used to access the processor name.
  • The values of predefined attributes attached to the world communicator can be obtained by calling the MPI.Comm.Get_attr() method within the MPI.COMM_WORLD instance.

Timers

MPI timer functionalities are available through the MPI.Wtime() and MPI.Wtick() functions.

Error Handling

In order facilitate handle sharing with other Python modules interfacing MPI-based parallel libraries, the predefined MPI error handlers MPI.ERRORS_RETURN and MPI.ERRORS_ARE_FATAL can be assigned to and retrieved from communicators, windows and files using methods MPI.{Comm|Win|File}.Set_errhandler() and MPI.{Comm|Win|File}.Get_errhandler().

When the predefined error handler MPI.ERRORS_RETURN is set, errors returned from MPI calls within Python code will raise an instance of the exception class MPI.Exception, which is a subclass of the standard Python exception RuntimeError.

Note

After import, mpi4py overrides the default MPI rules governing inheritance of error handlers. The MPI.ERRORS_RETURN error handler is set in the predefined MPI.COMM_SELF and MPI.COMM_WORLD communicators, as well as any new MPI.Comm, MPI.Win, or MPI.File instance created through mpi4py. If you ever pass such handles to C/C++/Fortran library code, it is recommended to set the MPI.ERRORS_ARE_FATAL error handler on them to ensure MPI errors do not pass silently.

Warning

Importing with from mpi4py.MPI import * will cause a name clashing with the standard Python Exception base class.