| Author: | Lisandro Dalcin |
|---|---|
| Contact: | dalcinl@gmail.com |
Bug fixes:
Fix Comm.scatter() and other collectives corrupting input send list. Add safety measures to prevent related issues in global reduction operations.
Fix error-checking code for counts in Op.Reduce_local().
Enhancements:
Map size-specific Python/NumPy typecodes to MPI datatypes.
Allow partial specification of target list/tuple arguments in the various Win RMA methods.
Workaround for removal of MPI_{LB|UB} in Open MPI 4.0.
Support for Microsoft MPI v10.0.
New features:
mpi4py.futures: Execute computations asynchronously using a pool of MPI processes. This package is based on concurrent.futures from the Python standard library.
mpi4py.run: Run Python code and abort execution in case of unhandled exceptions to prevent deadlocks.
mpi4py.bench: Run basic MPI benchmarks and tests.
Enhancements:
Lowercase, pickle-based collective communication calls are now thread-safe through the use of fine-grained locking.
The MPI module now exposes a memory type which is a lightweight variant of the builtin memoryview type, but exposes both the legacy Python 2 and the modern Python 3 buffer interface under a Python 2 runtime.
The MPI.Comm.Alltoallw() method now uses count=1 and displ=0 as defaults, assuming that messages are specified through user-defined datatypes.
The Request.Wait[all]() methods now return True to match the interface of Request.Test[all]().
The Win class now implements the Python buffer interface.
Backward-incompatible changes:
The buf argument of the MPI.Comm.recv() method is deprecated, passing anything but None emits a warning.
The MPI.Win.memory property was removed, use the MPI.Win.tomemory() method instead.
Executing python -m mpi4py in the command line is now equivalent to python -m mpi4py.run. For the former behavior, use python -m mpi4py.bench.
Python 2.6 and 3.2 are no longer supported. The mpi4py.MPI module may still build and partially work, but other pure-Python modules under the mpi4py namespace will not.
Windows: Remove support for legacy MPICH2, Open MPI, and DeinoMPI.
Support for MPI-3 features.
Matched probes and receives.
Nonblocking collectives.
Neighborhood collectives.
New communicator constructors.
Request-based RMA operations.
New RMA communication and synchronisation calls.
New window constructors.
New datatype constructor.
New C++ boolean and floating complex datatypes.
Support for MPI-2 features not included in previous releases.
Generalized All-to-All collective (Comm.Alltoallw())
User-defined data representations (Register_datarep())
New scalable implementation of reduction operations for Python objects. This code is based on binomial tree algorithms using point-to-point communication and duplicated communicator contexts. To disable this feature, use mpi4py.rc.fast_reduce = False.
Backward-incompatible changes:
Python 2.4, 2.5, 3.0 and 3.1 are no longer supported.
Default MPI error handling policies are overriden. After import, mpi4py sets the ERRORS_RETURN error handler in COMM_SELF and COMM_WORLD, as well as any new Comm, Win, or File instance created through mpi4py, thus effectively ignoring the MPI rules about error handler inheritance. This way, MPI errors translate to Python exceptions. To disable this behavior and use the standard MPI error handling rules, use mpi4py.rc.errors = 'default'.
Change signature of all send methods, dest is a required argument.
Change signature of all receive and probe methods, source defaults to ANY_SOURCE, tag defaults to ANY_TAG.
Change signature of send lowercase-spelling methods, obj arguments are not mandatory.
Change signature of recv lowercase-spelling methods, renamed 'obj' arguments to 'buf'.
Change Request.Waitsome() and Request.Testsome() to return None or list.
Change signature of all lowercase-spelling collectives, sendobj arguments are now mandatory, recvobj arguments were removed.
Reduction operations MAXLOC and MINLOC are no longer special-cased in lowercase-spelling methods Comm.[all]reduce() and Comm.[ex]scan(), the input object must be specified as a tuple (obj, location).
Change signature of name publishing functions. The new signatures are Publish_name(service_name, port_name, info=INFO_NULL) and Unpublish_name(service_name, port_name, info=INFO_NULL)`.
Win instances now cache Python objects exposing memory by keeping references instead of using MPI attribute caching.
Change signature of Win.Lock(). The new signature is Win.Lock(rank, lock_type=LOCK_EXCLUSIVE, assertion=0).
Move Cartcomm.Map() to Intracomm.Cart_map().
Move Graphcomm.Map() to Intracomm.Graph_map().
Remove the mpi4py.MPE module.
Rename the Cython definition file for use with cimport statement from mpi_c.pxd to libmpi.pxd.
Regenerate C wrappers with Cython 0.19.1 to support Python 3.3.
Install *.pxd files in <site-packages>/mpi4py to ease the support for Cython's cimport statement in code requiring to access mpi4py internals.
As a side-effect of using Cython 0.19.1, ancient Python 2.3 is no longer supported. If you really need it, you can install an older Cython and run python setup.py build_src --force.
Now Comm.recv() accept a buffer to receive the message.
Add Comm.irecv() and Request.{wait|test}[any|all]().
Add Intracomm.Spawn_multiple().
Better buffer handling for PEP 3118 and legacy buffer interfaces.
Add support for attribute attribute caching on communicators, datatypes and windows.
Install MPI-enabled Python interpreter as <path>/mpi4py/bin/python-mpi.
Windows: Support for building with Open MPI.
Add mpi4py.get_config() to retrieve information (compiler wrappers, includes, libraries, etc) about the MPI implementation employed to build mpi4py.
Workaround Python libraries with missing GILState-related API calls in case of non-threaded Python builds.
Windows: look for MPICH2, DeinoMPI, Microsoft HPC Pack at their default install locations under %ProgramFiles.
MPE: fix hacks related to old API's, these hacks are broken when MPE is built with a MPI implementations other than MPICH2.
HP-MPI: fix for missing Fortran datatypes, use dlopen() to load the MPI shared library before MPI_Init()
Many distutils-related fixes, cleanup, and enhancements, better logics to find MPI compiler wrappers.
Support for pip install mpi4py.
Fix declaration in Cython include file. This declaration, while valid for Cython, broke the simple-minded parsing used in conf/mpidistutils.py to implement configure-tests for availability of MPI symbols.
Update SWIG support and make it compatible with Python 3. Also generate an warning for SWIG < 1.3.28.
Fix distutils-related issues in Mac OS X. Now ARCHFLAGS environment variable is honored of all Python's config/Makefile variables.
Fix issues with Open MPI < 1.4.2 releated to error checking and MPI_XXX_NULL handles.
Automatic MPI datatype discovery for NumPy arrays and PEP-3118 buffers. Now buffer-like objects can be messaged directly, it is no longer required to explicitly pass a 2/3-list/tuple like [data, MPI.DOUBLE], or [data, count, MPI.DOUBLE]. Only basic types are supported, i.e., all C/C99-native signed/unsigned integral types and single/double precision real/complex floating types. Many thanks to Eilif Muller for the initial feedback.
Nonblocking send of pickled Python objects. Many thanks to Andreas Kloeckner for the initial patch and enlightening discussion about this enhancement.
Request instances now hold a reference to the Python object exposing the buffer involved in point-to-point communication or parallel I/O. Many thanks to Andreas Kloeckner for the initial feedback.
Support for logging of user-defined states and events using MPE. Runtime (i.e., without requiring a recompile!) activation of logging of all MPI calls is supported in POSIX platforms implementing dlopen().
Support for all the new features in MPI-2.2 (new C99 and F90 datatypes, distributed graph topology, local reduction operation, and other minor enhancements).
Fix the annoying issues related to Open MPI and Python dynamic loading of extension modules in platforms supporting dlopen().
Fix SLURM dynamic loading issues on SiCortex. Many thanks to Ian Langmore for providing me shell access.
Fix bug in Comm.Iprobe() that caused segfaults as Python C-API calls were issued with the GIL released (issue #2).
Add Comm.bsend() and Comm.ssend() for buffered and synchronous send semantics when communicating general Python objects.
Now the call Info.Get(key) return a single value (i.e, instead of a 2-tuple); this value is None if key is not in the Info object, or a string otherwise. Previously, the call redundantly returned (None, False) for missing key-value pairs; None is enough to signal a missing entry.
Add support for parametrized Fortran datatypes.
Add support for decoding user-defined datatypes.
Add support for user-defined reduction operations on memory buffers. However, at most 16 user-defined reduction operations can be created. Ask the author for more room if you need it.
This is the fist release of the all-new, Cython-based, implementation of MPI for Python. Unfortunately, this implementation is not backward-compatible with the previous one. The list below summarizes the more important changes that can impact user codes.
Some communication calls had overloaded functionality. Now there is a clear distinction between communication of general Python object with pickle, and (fast, near C-speed) communication of buffer-like objects (e.g., NumPy arrays).
for communicating general Python objects, you have to use all-lowercase methods, like send(), recv(), bcast(), etc.
for communicating array data, you have to use Send(), Recv(), Bcast(), etc. methods. Buffer arguments to these calls must be explicitly specified by using a 2/3-list/tuple like [data, MPI.DOUBLE], or [data, count, MPI.DOUBLE] (the former one uses the byte-size of data and the extent of the MPI datatype to define the count).
Indexing a communicator with an integer returned a special object associating the communication with a target rank, alleviating you from specifying source/destination/root arguments in point-to-point and collective communications. This functionality is no longer available, expressions like:
MPI.COMM_WORLD[0].Send(...) MPI.COMM_WORLD[0].Recv(...) MPI.COMM_WORLD[0].Bcast(...)
have to be replaced by:
MPI.COMM_WORLD.Send(..., dest=0) MPI.COMM_WORLD.Recv(..., source=0) MPI.COMM_WORLD.Bcast(..., root=0)
Automatic MPI initialization (i.e., at import time) requests the maximum level of MPI thread support (i.e., it is done by calling MPI_Init_thread() and passing MPI_THREAD_MULTIPLE). In case you need to change this behavior, you can tweak the contents of the mpi4py.rc module.
In order to obtain the values of predefined attributes attached to the world communicator, now you have to use the Get_attr() method on the MPI.COMM_WORLD instance:
tag_ub = MPI.COMM_WORLD.Get_attr(MPI.TAG_UB)
In the previous implementation, MPI.COMM_WORLD and MPI.COMM_SELF were associated to duplicates of the (C-level) MPI_COMM_WORLD and MPI_COMM_SELF predefined communicator handles. Now this is no longer the case, MPI.COMM_WORLD and MPI.COMM_SELF proxies the actual MPI_COMM_WORLD and MPI_COMM_SELF handles.
Convenience aliases MPI.WORLD and MPI.SELF were removed. Use instead MPI.COMM_WORLD and MPI.COMM_SELF.
Convenience constants MPI.WORLD_SIZE and MPI.WORLD_RANK were removed. Use instead MPI.COMM_WORLD.Get_size() and MPI.COMM_WORLD.Get_rank().