Abstract¶

MPI for Python provides Python bindings for the Message Passing Interface (MPI) standard, allowing Python applications to exploit multiple processors on workstations, clusters and supercomputers.

This package builds on the MPI specification and provides an object oriented interface resembling the MPI-2 C++ bindings. It supports point-to-point (sends, receives) and collective (broadcasts, scatters, gathers) communication of any picklable Python object, as well as efficient communication of Python objects exposing the Python buffer interface (e.g. NumPy arrays and builtin bytes/array/memoryview objects).

What is MPI?¶

MPI, [mpi-using] [mpi-ref] the Message Passing Interface, is a standardized and portable message-passing system designed to function on a wide variety of parallel computers. The standard defines the syntax and semantics of library routines and allows users to write portable programs in the main scientific programming languages (Fortran, C, or C++).

Since its release, the MPI specification [mpi-std1] [mpi-std2] has become the leading standard for message-passing libraries for parallel computers. Implementations are available from vendors of high-performance computers and from well known open source projects like MPICH [mpi-mpich] and Open MPI [mpi-openmpi].

What is Python?¶

Python is a modern, easy to learn, powerful programming language. It has efficient high-level data structures and a simple but effective approach to object-oriented programming with dynamic typing and dynamic binding. It supports modules and packages, which encourages program modularity and code reuse. Python’s elegant syntax, together with its interpreted nature, make it an ideal language for scripting and rapid application development in many areas on most platforms.

The Python interpreter and the extensive standard library are available in source or binary form without charge for all major platforms, and can be freely distributed. It is easily extended with new functions and data types implemented in C or C++. Python is also suitable as an extension language for customizable applications.

Python is an ideal candidate for writing the higher-level parts of large-scale scientific applications [Hinsen97] and driving simulations in parallel architectures [Beazley97] like clusters of PC’s or SMP’s. Python codes are quickly developed, easily maintained, and can achieve a high degree of integration with other libraries written in compiled languages.

Related Projects¶

As this work started and evolved, some ideas were borrowed from well known MPI and Python related open source projects from the Internet.

Additionally, we would like to mention some available tools for scientific computing and software development with Python.

[mpi-std1]

MPI Forum. MPI: A Message Passing Interface Standard. International Journal of Supercomputer Applications, volume 8, number 3-4, pages 159-416, 1994.

[mpi-std2]

MPI Forum. MPI: A Message Passing Interface Standard. High Performance Computing Applications, volume 12, number 1-2, pages 1-299, 1998.

[mpi-using]

William Gropp, Ewing Lusk, and Anthony Skjellum. Using MPI: portable parallel programming with the message-passing interface. MIT Press, 1994.

[mpi-ref]

Mark Snir, Steve Otto, Steven Huss-Lederman, David Walker, and Jack Dongarra. MPI - The Complete Reference, volume 1, The MPI Core. MIT Press, 2nd. edition, 1998.

[mpi-mpich]

W. Gropp, E. Lusk, N. Doss, and A. Skjellum. A high-performance, portable implementation of the MPI message passing interface standard. Parallel Computing, 22(6):789-828, September 1996.

[mpi-openmpi]

Edgar Gabriel, Graham E. Fagg, George Bosilca, Thara Angskun, Jack J. Dongarra, Jeffrey M. Squyres, Vishal Sahay, Prabhanjan Kambadur, Brian Barrett, Andrew Lumsdaine, Ralph H. Castain, David J. Daniel, Richard L. Graham, and Timothy S. Woodall. Open MPI: Goals, Concept, and Design of a Next Generation MPI Implementation. In Proceedings, 11th European PVM/MPI Users’ Group Meeting, Budapest, Hungary, September 2004.

[Hinsen97]

Konrad Hinsen. The Molecular Modelling Toolkit: a case study of a large scientific application in Python. In Proceedings of the 6th International Python Conference, pages 29-35, San Jose, Ca., October 1997.

[Beazley97]

David M. Beazley and Peter S. Lomdahl. Feeding a large-scale physics application to Python. In Proceedings of the 6th International Python Conference, pages 21-29, San Jose, Ca., October 1997.

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, Comm is the base class of communicators. The Intracomm and Intercomm classes are subclasses of the Comm class. The Comm.Is_inter method (and 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: COMM_SELF and 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 Comm.Get_size and Comm.Get_rank. The associated process group can be retrieved from a communicator by calling the Comm.Get_group method, which returns an instance of the Group class. Set operations with Group objects like like Group.Union, Group.Intersection and Group.Difference are fully supported, as well as the creation of new communicators from these groups using Comm.Create and Intracomm.Create_group.

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

Virtual topologies (Cartcomm, Graphcomm and Distgraphcomm classes, which are specializations of the Intracomm class) are fully supported. New instances can be obtained from intracommunicator instances with factory methods Intracomm.Create_cart and 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 Comm.Send, Comm.Recv and Comm.Sendrecv methods of communicator objects provide support for blocking point-to-point communications within Intracomm and Intercomm instances. These methods can communicate memory buffers. The variants Comm.send, Comm.recv and 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 Comm.Isend and Comm.Irecv methods initiate send and receive operations, respectively. These methods return a Request instance, uniquely identifying the started operation. Its completion can be managed using the Request.Test, Request.Wait and Request.Cancel methods. The management of 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 Comm.Send_init and Comm.Recv_init methods create persistent requests for a send and receive operation, respectively. These methods return an instance of the Prequest class, a subclass of the Request class. The actual communication can be effectively started using the 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.

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

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

Support for GPU-aware MPI¶

Several MPI implementations, including Open MPI and MVAPICH, support passing GPU pointers to MPI calls to avoid explicit data movement between host and device. On the Python side, support for handling GPU arrays have been implemented in many libraries related GPU computation such as CuPy, Numba, PyTorch, and PyArrow. To maximize interoperability across library boundaries, two kinds of zero-copy data exchange protocols have been defined and agreed upon: DLPack and CUDA Array Interface (CAI).

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

See the Tutorial section for further information. We note that

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 Intracomm.Spawn method within an intracommunicator. This call returns a new intercommunicator (i.e., an Intercomm instance) at the parent process group. The child process group can retrieve the matching intercommunicator by calling the 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 Open_port function to open a port and the Publish_name function to publish a provided service, and next call the Intracomm.Accept method. Any client applications can first find a published service by calling the Lookup_name function, which returns the port where a server can be contacted; and next call the Intracomm.Connect method. Both Intracomm.Accept and Intracomm.Connect methods return an Intercomm instance. When connection between client/server processes is no longer needed, all of them must cooperatively call the Comm.Disconnect method. Additionally, server applications should release resources by calling the Unpublish_name and 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 Win class. New window objects are created by calling the Win.Create method at all processes within a communicator and specifying a memory buffer . When a window instance is no longer needed, the Win.Free method should be called.

The three one-sided MPI operations for remote write, read and reduction are available through calling the methods Win.Put, Win.Get, and 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 Win.Start and Win.Complete methods at the origin process, and target process cooperates by calling the Win.Post and Win.Wait methods. There is also a collective variant provided by the Win.Fence method. Passive target synchronization is more lenient, only the origin process calls the Win.Lock and 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 File class. File handles are obtained by calling the 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 File.Close method. Files even can be deleted by calling method 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 File.Set_view and 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 Init or Init_thread and Finalize provide MPI initialization and finalization respectively. Module functions Is_initialized and Is_finalized provide the respective tests for initialization and finalization.

NOTE:

NOTE:

Implementation Information¶

Timers¶

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

Error Handling¶

In order to facilitate handle sharing with other Python modules interfacing MPI-based parallel libraries, the predefined MPI error handlers ERRORS_RETURN and ERRORS_ARE_FATAL can be assigned to and retrieved from communicators using methods Comm.Set_errhandler and Comm.Get_errhandler, and similarly for windows and files. New custom error handlers can be created with Comm.Create_errhandler.

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

NOTE:

WARNING:

Running Python scripts with MPI¶

Most MPI programs can be run with the command mpiexec. In practice, running Python programs looks like:

$ mpiexec -n 4 python script.py

to run the program with 4 processors.

Point-to-Point Communication¶

from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:


    data = {'a': 7, 'b': 3.14}


    comm.send(data, dest=1, tag=11)
elif rank == 1:


    data = comm.recv(source=0, tag=11)

from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:


    data = {'a': 7, 'b': 3.14}


    req = comm.isend(data, dest=1, tag=11)


    req.wait()
elif rank == 1:


    req = comm.irecv(source=0, tag=11)


    data = req.wait()

from mpi4py import MPI
import numpy
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
# passing MPI datatypes explicitly
if rank == 0:


    data = numpy.arange(1000, dtype='i')


    comm.Send([data, MPI.INT], dest=1, tag=77)
elif rank == 1:


    data = numpy.empty(1000, dtype='i')


    comm.Recv([data, MPI.INT], source=0, tag=77)
# automatic MPI datatype discovery
if rank == 0:


    data = numpy.arange(100, dtype=numpy.float64)


    comm.Send(data, dest=1, tag=13)
elif rank == 1:


    data = numpy.empty(100, dtype=numpy.float64)


    comm.Recv(data, source=0, tag=13)

Collective Communication¶

from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:


    data = {'key1' : [7, 2.72, 2+3j],


            'key2' : ( 'abc', 'xyz')}
else:


    data = None
data = comm.bcast(data, root=0)

from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
if rank == 0:


    data = [(i+1)**2 for i in range(size)]
else:


    data = None
data = comm.scatter(data, root=0)
assert data == (rank+1)**2

from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
data = (rank+1)**2
data = comm.gather(data, root=0)
if rank == 0:


    for i in range(size):


        assert data[i] == (i+1)**2
else:


    assert data is None

from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:


    data = np.arange(100, dtype='i')
else:


    data = np.empty(100, dtype='i')
comm.Bcast(data, root=0)
for i in range(100):


    assert data[i] == i

from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
sendbuf = None
if rank == 0:


    sendbuf = np.empty([size, 100], dtype='i')


    sendbuf.T[:,:] = range(size)
recvbuf = np.empty(100, dtype='i')
comm.Scatter(sendbuf, recvbuf, root=0)
assert np.allclose(recvbuf, rank)

from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
sendbuf = np.zeros(100, dtype='i') + rank
recvbuf = None
if rank == 0:


    recvbuf = np.empty([size, 100], dtype='i')
comm.Gather(sendbuf, recvbuf, root=0)
if rank == 0:


    for i in range(size):


        assert np.allclose(recvbuf[i,:], i)

from mpi4py import MPI
import numpy
def matvec(comm, A, x):


    m = A.shape[0] # local rows


    p = comm.Get_size()


    xg = numpy.zeros(m*p, dtype='d')


    comm.Allgather([x,  MPI.DOUBLE],


                   [xg, MPI.DOUBLE])


    y = numpy.dot(A, xg)


    return y

Input/Output (MPI-IO)¶

from mpi4py import MPI
import numpy as np
amode = MPI.MODE_WRONLY|MPI.MODE_CREATE
comm = MPI.COMM_WORLD
fh = MPI.File.Open(comm, "./datafile.contig", amode)
buffer = np.empty(10, dtype=np.int)
buffer[:] = comm.Get_rank()
offset = comm.Get_rank()*buffer.nbytes
fh.Write_at_all(offset, buffer)
fh.Close()

from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
amode = MPI.MODE_WRONLY|MPI.MODE_CREATE
fh = MPI.File.Open(comm, "./datafile.noncontig", amode)
item_count = 10
buffer = np.empty(item_count, dtype='i')
buffer[:] = rank
filetype = MPI.INT.Create_vector(item_count, 1, size)
filetype.Commit()
displacement = MPI.INT.Get_size()*rank
fh.Set_view(displacement, filetype=filetype)
fh.Write_all(buffer)
filetype.Free()
fh.Close()

Dynamic Process Management¶

#!/usr/bin/env python
from mpi4py import MPI
import numpy
import sys
comm = MPI.COMM_SELF.Spawn(sys.executable,


                           args=['cpi.py'],


                           maxprocs=5)
N = numpy.array(100, 'i')
comm.Bcast([N, MPI.INT], root=MPI.ROOT)
PI = numpy.array(0.0, 'd')
comm.Reduce(None, [PI, MPI.DOUBLE],


            op=MPI.SUM, root=MPI.ROOT)
print(PI)
comm.Disconnect()

#!/usr/bin/env python
from mpi4py import MPI
import numpy
comm = MPI.Comm.Get_parent()
size = comm.Get_size()
rank = comm.Get_rank()
N = numpy.array(0, dtype='i')
comm.Bcast([N, MPI.INT], root=0)
h = 1.0 / N; s = 0.0
for i in range(rank, N, size):


    x = h * (i + 0.5)


    s += 4.0 / (1.0 + x**2)
PI = numpy.array(s * h, dtype='d')
comm.Reduce([PI, MPI.DOUBLE], None,


            op=MPI.SUM, root=0)
comm.Disconnect()

GPU-aware MPI + Python GPU arrays¶

from mpi4py import MPI
import cupy as cp
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
sendbuf = cp.arange(10, dtype='i')
recvbuf = cp.empty_like(sendbuf)
cp.cuda.get_current_stream().synchronize()
comm.Allreduce(sendbuf, recvbuf)
assert cp.allclose(recvbuf, sendbuf*size)

One-Sided Communication (RMA)¶

import numpy as np
from mpi4py import MPI
from mpi4py.util import dtlib
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
datatype = MPI.FLOAT
np_dtype = dtlib.to_numpy_dtype(datatype)
itemsize = datatype.Get_size()
N = 10
win_size = N * itemsize if rank == 0 else 0
win = MPI.Win.Allocate(win_size, comm=comm)
buf = np.empty(N, dtype=np_dtype)
if rank == 0:


    buf.fill(42)


    win.Lock(rank=0)


    win.Put(buf, target_rank=0)


    win.Unlock(rank=0)


    comm.Barrier()
else:


    comm.Barrier()


    win.Lock(rank=0)


    win.Get(buf, target_rank=0)


    win.Unlock(rank=0)


    assert np.all(buf == 42)

import numpy as np
from mpi4py import MPI
from mpi4py.util import dtlib
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
datatype = MPI.FLOAT
np_dtype = dtlib.to_numpy_dtype(datatype)
itemsize = datatype.Get_size()
N = comm.Get_size() + 1
win_size = N * itemsize if rank == 0 else 0
win = MPI.Win.Allocate(


    size=win_size,


    disp_unit=itemsize,


    comm=comm,
)
if rank == 0:


    mem = np.frombuffer(win, dtype=np_dtype)


    mem[:] = np.arange(len(mem), dtype=np_dtype)
comm.Barrier()
buf = np.zeros(3, dtype=np_dtype)
target = (rank, 2, datatype)
win.Lock(rank=0)
win.Get(buf, target_rank=0, target=target)
win.Unlock(rank=0)
assert np.all(buf == [rank, rank+1, 0])

Wrapping with SWIG¶

/* file: helloworld.c */
void sayhello(MPI_Comm comm)
{


  int size, rank;


  MPI_Comm_size(comm, &size);


  MPI_Comm_rank(comm, &rank);


  printf("Hello, World! "


         "I am process %d of %d.\n",


         rank, size);
}

// file: helloworld.i
%module helloworld
%{
#include <mpi.h>
#include "helloworld.c"
}%
%include mpi4py/mpi4py.i
%mpi4py_typemap(Comm, MPI_Comm);
void sayhello(MPI_Comm comm);

>>> from mpi4py import MPI
>>> import helloworld
>>> helloworld.sayhello(MPI.COMM_WORLD)
Hello, World! I am process 0 of 1.

Wrapping with F2Py¶

! file: helloworld.f90
subroutine sayhello(comm)


  use mpi


  implicit none


  integer :: comm, rank, size, ierr


  call MPI_Comm_size(comm, size, ierr)


  call MPI_Comm_rank(comm, rank, ierr)


  print *, 'Hello, World! I am process ',rank,' of ',size,'.'
end subroutine sayhello

$ f2py -c --f90exec=mpif90 helloworld.f90 -m helloworld

>>> from mpi4py import MPI
>>> import helloworld
>>> fcomm = MPI.COMM_WORLD.py2f()
>>> helloworld.sayhello(fcomm)
Hello, World! I am process 0 of 1.

Runtime configuration options¶

Attributes Summary

Attributes Documentation

Example

MPI for Python features automatic initialization and finalization of the MPI execution environment. By using the mpi4py.rc object, MPI initialization and finalization can be handled programmatically:

import mpi4py
mpi4py.rc.initialize = False  # do not initialize MPI automatically
mpi4py.rc.finalize = False    # do not finalize MPI automatically
from mpi4py import MPI # import the 'MPI' module
MPI.Init()      # manual initialization of the MPI environment
...             # your finest code here ...
MPI.Finalize()  # manual finalization of the MPI environment

Environment variables¶

The following environment variables override the corresponding attributes of the mpi4py.rc and MPI.pickle objects at import time of the MPI module.

NOTE:

Miscellaneous functions¶

import mpi4py
Extension('extension_name', ...


          include_dirs=[..., mpi4py.get_include()])

Classes¶

Ancillary

Communication

One-sided operations

Input/Output

Error handling

Auxiliary

Functions¶

Version inquiry

Initialization and finalization

Memory allocation

Address manipulation

Timer

Error handling

Dynamic process management

Miscellanea

Utilities

Attributes¶

MPIPoolExecutor¶

The MPIPoolExecutor class uses a pool of MPI processes to execute calls asynchronously. By performing computations in separate processes, it allows to side-step the global interpreter lock but also means that only picklable objects can be executed and returned. The __main__ module must be importable by worker processes, thus MPIPoolExecutor instances may not work in the interactive interpreter.

MPIPoolExecutor takes advantage of the dynamic process management features introduced in the MPI-2 standard. In particular, the MPI.Intracomm.Spawn method of MPI.COMM_SELF is used in the master (or parent) process to spawn new worker (or child) processes running a Python interpreter. The master process uses a separate thread (one for each MPIPoolExecutor instance) to communicate back and forth with the workers. The worker processes serve the execution of tasks in the main (and only) thread until they are signaled for completion.

NOTE:

WARNING:

executor = MPIPoolExecutor(max_workers=1)
future = executor.submit(pow, 321, 1234)
print(future.result())

executor = MPIPoolExecutor(max_workers=3)
for result in executor.map(pow, [2]*32, range(32)):


    print(result)

executor = MPIPoolExecutor(max_workers=3)
iterable = ((2, n) for n in range(32))
for result in executor.starmap(pow, iterable):


    print(result)

import time
with MPIPoolExecutor(max_workers=1) as executor:


    future = executor.submit(time.sleep, 2)
assert future.done()

NOTE:

WARNING:

MPICommExecutor¶

Legacy MPI-1 implementations (as well as some vendor MPI-2 implementations) do not support the dynamic process management features introduced in the MPI-2 standard. Additionally, job schedulers and batch systems in supercomputing facilities may pose additional complications to applications using the MPI_Comm_spawn() routine.

With these issues in mind, mpi4py.futures supports an additional, more traditional, SPMD-like usage pattern requiring MPI-1 calls only. Python applications are started the usual way, e.g., using the mpiexec command. Python code should make a collective call to the MPICommExecutor context manager to partition the set of MPI processes within a MPI communicator in one master processes and many workers processes. The master process gets access to an MPIPoolExecutor instance to submit tasks. Meanwhile, the worker process follow a different execution path and team-up to execute the tasks submitted from the master.

Besides alleviating the lack of dynamic process management features in legacy MPI-1 or partial MPI-2 implementations, the MPICommExecutor context manager may be useful in classic MPI-based Python applications willing to take advantage of the simple, task-based, master/worker approach available in the mpi4py.futures package.

from mpi4py import MPI
from mpi4py.futures import MPICommExecutor
with MPICommExecutor(MPI.COMM_WORLD, root=0) as executor:


    if executor is not None:


       future = executor.submit(abs, -42)


       assert future.result() == 42


       answer = set(executor.map(abs, [-42, 42]))


       assert answer == {42}

WARNING:

Command line¶

Recalling the issues related to the lack of support for dynamic process management features in MPI implementations, mpi4py.futures supports an alternative usage pattern where Python code (either from scripts, modules, or zip files) is run under command line control of the mpi4py.futures package by passing -m mpi4py.futures to the python executable. The mpi4py.futures invocation should be passed a pyfile path to a script (or a zipfile/directory containing a __main__.py file). Additionally, mpi4py.futures accepts -m mod to execute a module named mod, -c cmd to execute a command string cmd, or even - to read commands from standard input (sys.stdin). Summarizing, mpi4py.futures can be invoked in the following ways:

Before starting the main script execution, mpi4py.futures splits MPI.COMM_WORLD in one master (the process with rank 0 in MPI.COMM_WORLD) and numprocs - 1 workers and connects them through an MPI intercommunicator. Afterwards, the master process proceeds with the execution of the user script code, which eventually creates MPIPoolExecutor instances to submit tasks. Meanwhile, the worker processes follow a different execution path to serve the master. Upon successful termination of the main script at the master, the entire MPI execution environment exists gracefully. In case of any unhandled exception in the main script, the master process calls MPI.COMM_WORLD.Abort(1) to prevent deadlocks and force termination of entire MPI execution environment.

WARNING:

SEE ALSO:

Parallel tasks¶

The mpi4py.futures package favors an embarrassingly parallel execution model involving a series of sequential tasks independent of each other and executed asynchronously. Albeit unnatural, MPIPoolExecutor can still be used for handling workloads involving parallel tasks, where worker processes communicate and coordinate each other via MPI.

Executing parallel tasks with mpi4py.futures requires following some rules, cf. highlighted lines in example cpi.py :

Utilities¶

The mpi4py.futures package provides additional utilities for handling Future instances.

Examples¶

Computing the Julia set¶

The following julia.py script computes the Julia set and dumps an image to disk in binary PGM format. The code starts by importing MPIPoolExecutor from the mpi4py.futures package. Next, some global constants and functions implement the computation of the Julia set. The computations are protected with the standard if __name__ == '__main__': ... idiom. The image is computed by whole scanlines submitting all these tasks at once using the map method. The result iterator yields scanlines in-order as the tasks complete. Finally, each scanline is dumped to disk.

julia.py

from mpi4py.futures import MPIPoolExecutor
x0, x1, w = -2.0, +2.0, 640*2
y0, y1, h = -1.5, +1.5, 480*2
dx = (x1 - x0) / w
dy = (y1 - y0) / h
c = complex(0, 0.65)
def julia(x, y):


    z = complex(x, y)


    n = 255


    while abs(z) < 3 and n > 1:


        z = z**2 + c


        n -= 1


    return n
def julia_line(k):


    line = bytearray(w)


    y = y1 - k * dy


    for j in range(w):


        x = x0 + j * dx


        line[j] = julia(x, y)


    return line
if __name__ == '__main__':


    with MPIPoolExecutor() as executor:


        image = executor.map(julia_line, range(h))


        with open('julia.pgm', 'wb') as f:


            f.write(b'P5 %d %d %d\n' % (w, h, 255))


            for line in image:


                f.write(line)

The recommended way to execute the script is by using the mpiexec command specifying one MPI process (master) and (optional but recommended) the desired MPI universe size, which determines the number of additional dynamically spawned processes (workers). The MPI universe size is provided either by a batch system or set by the user via command-line arguments to mpiexec or environment variables. Below we provide examples for MPICH and Open MPI implementations [1]. In all of these examples, the mpiexec command launches a single master process running the Python interpreter and executing the main script. When required, mpi4py.futures spawns the pool of 16 worker processes. The master submits tasks to the workers and waits for the results. The workers receive incoming tasks, execute them, and send back the results to the master.

When using MPICH implementation or its derivatives based on the Hydra process manager, users can set the MPI universe size via the -usize argument to mpiexec:

$ mpiexec -n 1 -usize 17 python julia.py

or, alternatively, by setting the MPIEXEC_UNIVERSE_SIZE environment variable:

$ env MPIEXEC_UNIVERSE_SIZE=17 mpiexec -n 1 python julia.py

In the Open MPI implementation, the MPI universe size can be set via the -host argument to mpiexec:

$ mpiexec -n 1 -host localhost:17 python julia.py

Another way to specify the number of workers is to use the mpi4py.futures-specific environment variable MPI4PY_FUTURES_MAX_WORKERS:

$ env MPI4PY_FUTURES_MAX_WORKERS=16 mpiexec -n 1 python julia.py

Note that in this case, the MPI universe size is ignored.

Alternatively, users may decide to execute the script in a more traditional way, that is, all the MPI processes are started at once. The user script is run under command-line control of mpi4py.futures passing the -m flag to the python executable:

$ mpiexec -n 17 python -m mpi4py.futures julia.py

As explained previously, the 17 processes are partitioned in one master and 16 workers. The master process executes the main script while the workers execute the tasks submitted by the master.

[1]

When using an MPI implementation other than MPICH or Open MPI, please check the documentation of the implementation and/or batch system for the ways to specify the desired MPI universe size.

Computing Pi (parallel task)¶

The number \pi can be approximated via numerical integration with the simple midpoint rule, that is:

\pi = \int_{0}^{1} \frac{4}{1+x^2} \,dx \approx
\frac{1}{n} \sum_{i=1}^{n} \frac{4}{1 + \left[\frac{1}{n} \left(i-\frac{1}{2}\right) \right]^2} .

The following cpi.py script computes such approximations using mpi4py.futures with a parallel task involving a collective reduction operation. Highlighted lines correspond to the rules discussed in Parallel tasks.

cpi.py

import math
import sys
from mpi4py.futures import MPIPoolExecutor, wait
from mpi4py.futures import get_comm_workers
def compute_pi(n):


    # Access intracommunicator and synchronize


    comm = get_comm_workers()


    comm.Barrier()


    rank = comm.Get_rank()


    size = comm.Get_size()


    # Local computation


    h = 1.0 / n


    s = 0.0


    for i in range(rank + 1, n + 1, size):


        x = h * (i - 0.5)


        s += 4.0 / (1.0 + x**2)


    pi_partial = s * h


    # Parallel reduce-to-all


    pi = comm.allreduce(pi_partial)


    # All workers return the same value


    return pi
if __name__ == '__main__':


    n = int(sys.argv[1]) if len(sys.argv) > 1 else 256


    with MPIPoolExecutor() as executor:


        # Submit exactly one callable per worker


        P = executor.num_workers


        fs = [executor.submit(compute_pi, n) for _ in range(P)]


        # Wait for all workers to finish


        wait(fs)


        # Get result from the first future object.


        # In this particular example, due to using reduce-to-all,


        # all the other future objects hold the same result value.


        pi = fs[0].result()


        print(


            f"pi: {pi:.16f}, error: {abs(pi - math.pi):.3e}",


            f"({n:d} intervals, {P:d} workers)",


        )

To run in modern MPI-2 mode:

$ env MPI4PY_FUTURES_MAX_WORKERS=4 mpiexec -n 1 python cpi.py 128
pi: 3.1415977398528137, error: 5.086e-06 (128 intervals, 4 workers)
$ env MPI4PY_FUTURES_MAX_WORKERS=8 mpiexec -n 1 python cpi.py 512
pi: 3.1415929714812316, error: 3.179e-07 (512 intervals, 8 workers)

To run in legacy MPI-1 mode:

$ mpiexec -n 5 python -m mpi4py.futures cpi.py 128
pi: 3.1415977398528137, error: 5.086e-06 (128 intervals, 4 workers)
$ mpiexec -n 9 python -m mpi4py.futures cpi.py 512
pi: 3.1415929714812316, error: 3.179e-07 (512 intervals, 8 workers)

Citation¶

If mpi4py.futures been significant to a project that leads to an academic publication, please acknowledge our work by citing the following article [mpi4py-futures]:

[mpi4py-futures]

M. Rogowski, S. Aseeri, D. Keyes, and L. Dalcin, mpi4py.futures: MPI-Based Asynchronous Task Execution for Python, IEEE Transactions on Parallel and Distributed Systems, 34(2):611-622, 2023. https://doi.org/10.1109/TPDS.2022.3225481

mpi4py.util.dtlib¶

Added in version 3.1.0.

The mpi4py.util.dtlib module provides converter routines between NumPy and MPI datatypes.

mpi4py.util.pkl5¶

Added in version 3.1.0.

pickle protocol 5 (see PEP 574) introduced support for out-of-band buffers, allowing for more efficient handling of certain object types with large memory footprints.

MPI for Python uses the traditional in-band handling of buffers. This approach is appropriate for communicating non-buffer Python objects, or buffer-like objects with small memory footprints. For point-to-point communication, in-band buffer handling allows for the communication of a pickled stream with a single MPI message, at the expense of additional CPU and memory overhead in the pickling and unpickling steps.

The mpi4py.util.pkl5 module provides communicator wrapper classes reimplementing pickle-based point-to-point and collective communication methods using pickle protocol 5. Handling out-of-band buffers necessarily involves multiple MPI messages, thus increasing latency and hurting performance in case of small size data. However, in case of large size data, the zero-copy savings of out-of-band buffer handling more than offset the extra latency costs. Additionally, these wrapper methods overcome the infamous 2 GiB message count limit (MPI-1 to MPI-3).

NOTE:

python -m pip install pickle5

Examples¶

test-pkl5-1.py

import numpy as np
from mpi4py import MPI
from mpi4py.util import pkl5
comm = pkl5.Intracomm(MPI.COMM_WORLD)  # comm wrapper
size = comm.Get_size()
rank = comm.Get_rank()
dst = (rank + 1) % size
src = (rank - 1) % size
sobj = np.full(1024**3, rank, dtype='i4')  # > 4 GiB
sreq = comm.isend(sobj, dst, tag=42)
robj = comm.recv (None, src, tag=42)
sreq.Free()
assert np.min(robj) == src
assert np.max(robj) == src

test-pkl5-2.py

import numpy as np
from mpi4py import MPI
from mpi4py.util import pkl5
comm = pkl5.Intracomm(MPI.COMM_WORLD)  # comm wrapper
size = comm.Get_size()
rank = comm.Get_rank()
dst = (rank + 1) % size
src = (rank - 1) % size
sobj = np.full(1024**3, rank, dtype='i4')  # > 4 GiB
sreq = comm.isend(sobj, dst, tag=42)
status = MPI.Status()
rmsg = comm.mprobe(status=status)
assert status.Get_source() == src
assert status.Get_tag() == 42
rreq = rmsg.irecv()
robj = rreq.wait()
sreq.Free()
assert np.max(robj) == src
assert np.min(robj) == src

mpi4py.util.pool¶

Added in version 4.0.0.

SEE ALSO:

NOTE:

mpi4py.util.sync¶

Added in version 4.0.0.

The mpi4py.util.sync module provides parallel synchronization utilities.

Sequential execution¶

Global counter¶

Mutual exclusion¶

[mcs-paper]

John M. Mellor-Crummey and Michael L. Scott. Algorithms for scalable synchronization on shared-memory multiprocessors. ACM Transactions on Computer Systems, 9(1):21-65, February 1991. https://doi.org/10.1145/103727.103729

[uam-book]

William Gropp, Torsten Hoefler, Rajeev Thakur, Ewing Lusk. Using Advanced MPI - Modern Features of the Message-Passing Interface. Chapter 4, Section 4.7, Pages 130-131. The MIT Press, November 2014. https://mitpress.mit.edu/9780262527637/using-advanced-mpi/

Condition variable¶

Semaphore object¶

Examples¶

test-sync-1.py

from mpi4py import MPI
from mpi4py.util.sync import Counter, Sequential
comm = MPI.COMM_WORLD
counter = Counter(comm)
with Sequential(comm):


   value = next(counter)
counter.free()
assert comm.rank == value

test-sync-2.py

from mpi4py import MPI
from mpi4py.util.sync import Counter, Mutex
comm = MPI.COMM_WORLD
mutex = Mutex(comm)
counter = Counter(comm)
with mutex:


   value = next(counter)
counter.free()
mutex.free()
assert (


   list(range(comm.size)) ==


   sorted(comm.allgather(value))
)

Exceptions and deadlocks¶

Consider the following snippet of Python code. Assume this code is stored in a standard Python script file and run with mpiexec in two or more processes.

deadlock.py

from mpi4py import MPI
assert MPI.COMM_WORLD.Get_size() > 1
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:


    1/0


    MPI.COMM_WORLD.send(None, dest=1, tag=42)
elif rank == 1:


    MPI.COMM_WORLD.recv(source=0, tag=42)

Process 0 raises ZeroDivisionError exception before performing a send call to process 1. As the exception is not handled, the Python interpreter running in process 0 will proceed to exit with non-zero status. However, as mpi4py installed a finalizer hook to call MPI_Finalize() before exit, process 0 will block waiting for other processes to also enter the MPI_Finalize() call. Meanwhile, process 1 will block waiting for a message to arrive from process 0, thus never reaching to MPI_Finalize(). The whole MPI execution environment is irremediably in a deadlock state.

To alleviate this issue, mpi4py offers a simple, alternative command line execution mechanism based on using the -m flag and implemented with the runpy module. To use this features, Python code should be run passing -m mpi4py in the command line invoking the Python interpreter. In case of unhandled exceptions, the finalizer hook will call MPI_Abort() on the MPI_COMM_WORLD communicator, thus effectively aborting the MPI execution environment.

WARNING:

Command line¶

The use of -m mpi4py to execute Python code on the command line resembles that of the Python interpreter.

SEE ALSO:

mpi4py.MPI¶

Message Passing Interface.

Classes

mpi4py.MPI.BottomType¶

mpi4py.MPI.BufferAutomaticType¶

mpi4py.MPI.Cartcomm¶