mpi4py.futures
Added in version 3.0.0.
This package provides a high-level interface for asynchronously executing callables on a pool of worker processes using MPI for inter-process communication.
The mpi4py.futures
package is based on concurrent.futures
from
the Python standard library. More precisely, mpi4py.futures
provides the
MPIPoolExecutor
class as a concrete implementation of the abstract
class Executor
. The
submit()
interface schedules a callable to
be executed asynchronously and returns a Future
object representing the execution of the callable.
Future
instances can be queried for the call
result or exception. Sets of Future
instances can
be passed to the wait()
and
as_completed()
functions.
See also
- Module
concurrent.futures
Documentation of the
concurrent.futures
standard module.
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
The worker processes must import the main script in order to unpickle any
callable defined in the __main__
module and submitted from the master
process. Furthermore, the callables may need access to other global
variables. At the worker processes, mpi4py.futures
executes the main
script code (using the runpy
module) under the __worker__
namespace to define the __main__
module. The __main__
and
__worker__
modules are added to sys.modules
(both at the
master and worker processes) to ensure proper pickling and unpickling.
Warning
During the initial import phase at the workers, the main script cannot
create and use new MPIPoolExecutor
instances. Otherwise, each
worker would attempt to spawn a new pool of workers, leading to infinite
recursion. mpi4py.futures
detects such recursive attempts to spawn
new workers and aborts the MPI execution environment. As the main script
code is run under the __worker__
namespace, the easiest way to avoid
spawn recursion is using the idiom if __name__ == '__main__': ...
in
the main script.
- class mpi4py.futures.MPIPoolExecutor(max_workers=None, initializer=None, initargs=(), **kwargs)
An
Executor
subclass that executes calls asynchronously using a pool of at most max_workers processes. If max_workers isNone
or not given, its value is determined from theMPI4PY_FUTURES_MAX_WORKERS
environment variable if set, or the MPI universe size if set, otherwise a single worker process is spawned. If max_workers is lower than or equal to0
, then aValueError
will be raised.initializer is an optional callable that is called at the start of each worker process before executing any tasks; initargs is a tuple of arguments passed to the initializer. If initializer raises an exception, all pending tasks and any attempt to submit new tasks to the pool will raise a
BrokenExecutor
exception.Other parameters:
python_exe: Path to the Python interpreter executable used to spawn worker processes, otherwise
sys.executable
is used.python_args:
list
or iterable with additional command line flags to pass to the Python executable. Command line flags determined from inspection ofsys.flags
,sys.warnoptions
andsys._xoptions
in are passed unconditionally.mpi_info:
dict
or iterable yielding(key, value)
pairs. These(key, value)
pairs are passed (through anMPI.Info
object) to theMPI.Intracomm.Spawn
call used to spawn worker processes. This mechanism allows telling the MPI runtime system where and how to start the processes. Check the documentation of the backend MPI implementation about the set of keys it interprets and the corresponding format for values.globals:
dict
or iterable yielding(name, value)
pairs to initialize the main module namespace in worker processes.main: If set to
False
, do not import the__main__
module in worker processes. Setting main toFalse
prevents worker processes from accessing definitions in the parent__main__
namespace.path:
list
or iterable with paths to append tosys.path
in worker processes to extend the module search path.wdir: Path to set the current working directory in worker processes using
os.chdir()
. The initial working directory is set by the MPI implementation. Quality MPI implementations should honor awdir
info key passed through mpi_info, although such feature is not mandatory.env:
dict
or iterable yielding(name, value)
pairs with environment variables to updateos.environ
in worker processes. The initial environment is set by the MPI implementation. MPI implementations may allow setting the initial environment through mpi_info, however such feature is not required nor recommended by the MPI standard.use_pkl5: If set to
True
, usepickle5
with out-of-band buffers for interprocess communication. If use_pkl5 is set toNone
or not given, its value is determined from theMPI4PY_FUTURES_USE_PKL5
environment variable. Usingpickle5
with out-of-band buffers may benefit applications dealing with large buffer-like objects like NumPy arrays. Seempi4py.util.pkl5
for additional information.backoff:
float
value specifying the maximum number of seconds a worker thread or process suspends execution withtime.sleep()
while idle-waiting. If not set, its value is determined from theMPI4PY_FUTURES_BACKOFF
environment variable if set, otherwise the default value of 0.001 seconds is used. Lower values will reduce latency and increase execution throughput for very short-lived tasks, albeit at the expense of spinning CPU cores and increased energy consumption.
- submit(func, *args, **kwargs)
Schedule the callable, func, to be executed as
func(*args, **kwargs)
and returns aFuture
object representing the execution of the callable.executor = MPIPoolExecutor(max_workers=1) future = executor.submit(pow, 321, 1234) print(future.result())
- map(func, *iterables, timeout=None, chunksize=1, **kwargs)
Equivalent to
map(func, *iterables)
except func is executed asynchronously and several calls to func may be made concurrently, out-of-order, in separate processes. The returned iterator raises aTimeoutError
if__next__()
is called and the result isn’t available after timeout seconds from the original call tomap()
. timeout can be an int or a float. If timeout is not specified orNone
, there is no limit to the wait time. If a call raises an exception, then that exception will be raised when its value is retrieved from the iterator. This method chops iterables into a number of chunks which it submits to the pool as separate tasks. The (approximate) size of these chunks can be specified by setting chunksize to a positive integer. For very long iterables, using a large value for chunksize can significantly improve performance compared to the default size of one. By default, the returned iterator yields results in-order, waiting for successive tasks to complete . This behavior can be changed by passing the keyword argument unordered asTrue
, then the result iterator will yield a result as soon as any of the tasks complete.executor = MPIPoolExecutor(max_workers=3) for result in executor.map(pow, [2]*32, range(32)): print(result)
- starmap(func, iterable, timeout=None, chunksize=1, **kwargs)
Equivalent to
itertools.starmap(func, iterable)
. Used instead ofmap()
when argument parameters are already grouped in tuples from a single iterable (the data has been “pre-zipped”).map(func, *iterable)
is equivalent tostarmap(func, zip(*iterable))
.executor = MPIPoolExecutor(max_workers=3) iterable = ((2, n) for n in range(32)) for result in executor.starmap(pow, iterable): print(result)
- shutdown(wait=True, cancel_futures=False)
Signal the executor that it should free any resources that it is using when the currently pending futures are done executing. Calls to
submit()
andmap()
made aftershutdown()
will raiseRuntimeError
.If wait is
True
then this method will not return until all the pending futures are done executing and the resources associated with the executor have been freed. If wait isFalse
then this method will return immediately and the resources associated with the executor will be freed when all pending futures are done executing. Regardless of the value of wait, the entire Python program will not exit until all pending futures are done executing.If cancel_futures is
True
, this method will cancel all pending futures that the executor has not started running. Any futures that are completed or running won’t be cancelled, regardless of the value of cancel_futures.You can avoid having to call this method explicitly if you use the
with
statement, which will shutdown the executor instance (waiting as ifshutdown()
were called with wait set toTrue
).import time with MPIPoolExecutor(max_workers=1) as executor: future = executor.submit(time.sleep, 2) assert future.done()
- bootup(wait=True)
Signal the executor that it should allocate eagerly any required resources (in particular, MPI worker processes). If wait is
True
, thenbootup()
will not return until the executor resources are ready to process submissions. Resources are automatically allocated in the first call tosubmit()
, thus callingbootup()
explicitly is seldom needed.
- num_workers
Number or worker processes in the pool.
- MPI4PY_FUTURES_MAX_WORKERS
If the max_workers parameter to
MPIPoolExecutor
isNone
or not given, theMPI4PY_FUTURES_MAX_WORKERS
environment variable provides a fallback value for the maximum number of MPI worker processes to spawn.Added in version 3.1.0.
- MPI4PY_FUTURES_USE_PKL5
If the use_pkl5 keyword argument to
MPIPoolExecutor
isNone
or not given, theMPI4PY_FUTURES_USE_PKL5
environment variable provides a fallback value for whether the executor should usepickle5
with out-of-band buffers for interprocess communication. Accepted values are0
and1
(interpreted asFalse
andTrue
, respectively), and strings specifying a YAML boolean value (case-insensitive). Usingpickle5
with out-of-band buffers may benefit applications dealing with large buffer-like objects like NumPy arrays. Seempi4py.util.pkl5
for additional information.Added in version 4.0.0.
- MPI4PY_FUTURES_BACKOFF
If the backoff keyword argument to
MPIPoolExecutor
is not given, theMPI4PY_FUTURES_BACKOFF
environment variable can be set to afloat
value specifying the maximum number of seconds a worker thread or process suspends execution withtime.sleep()
while idle-waiting. If not set, the default backoff value is 0.001 seconds. Lower values will reduce latency and increase execution throughput for very short-lived tasks, albeit at the expense of spinning CPU cores and increased energy consumption.Added in version 4.0.0.
Note
As the master process uses a separate thread to perform MPI communication
with the workers, the backend MPI implementation should provide support for
MPI.THREAD_MULTIPLE
. However, some popular MPI implementations do not
support yet concurrent MPI calls from multiple threads. Additionally, users
may decide to initialize MPI with a lower level of thread support. If the
level of thread support in the backend MPI is less than
MPI.THREAD_MULTIPLE
, mpi4py.futures
will use a global lock to
serialize MPI calls. If the level of thread support is less than
MPI.THREAD_SERIALIZED
, mpi4py.futures
will emit a
RuntimeWarning
.
Warning
If the level of thread support in the backend MPI is less than
MPI.THREAD_SERIALIZED
(i.e, it is either MPI.THREAD_SINGLE
or
MPI.THREAD_FUNNELED
), in theory mpi4py.futures
cannot be
used. Rather than raising an exception, mpi4py.futures
emits a
warning and takes a “cross-fingers” attitude to continue execution in the
hope that serializing MPI calls with a global lock will actually work.
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.
- class mpi4py.futures.MPICommExecutor(comm=None, root=0)
Context manager for
MPIPoolExecutor
. This context manager splits a MPI (intra)communicator comm (defaults toMPI.COMM_WORLD
if not provided orNone
) in two disjoint sets: a single master process (with rank root in comm) and the remaining worker processes. These sets are then connected through an intercommunicator. The target of thewith
statement is assigned either anMPIPoolExecutor
instance (at the master) orNone
(at the workers).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
If MPICommExecutor
is passed a communicator of size one (e.g.,
MPI.COMM_SELF
), then the executor instance assigned to the target of the
with
statement will execute all submitted tasks in a single
worker thread, thus ensuring that task execution still progress
asynchronously. However, the GIL will prevent the main and worker
threads from running concurrently in multicore processors. Moreover, the
thread context switching may harm noticeably the performance of CPU-bound
tasks. In case of I/O-bound tasks, the GIL is not usually an issue,
however, as a single worker thread is used, it progress one task at a
time. We advice against using MPICommExecutor
with communicators of
size one and suggest refactoring your code to use instead a
ThreadPoolExecutor
.
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:
$ mpiexec -n numprocs python -m mpi4py.futures pyfile [arg] ...
$ mpiexec -n numprocs python -m mpi4py.futures -m mod [arg] ...
$ mpiexec -n numprocs python -m mpi4py.futures -c cmd [arg] ...
$ mpiexec -n numprocs python -m mpi4py.futures - [arg] ...
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
Running scripts under command line control of mpi4py.futures
is quite
similar to executing a single-process application that spawn additional
workers as required. However, there is a very important difference users
should be aware of. All MPIPoolExecutor
instances created at the
master will share the pool of workers. Tasks submitted at the master from
many different executors will be scheduled for execution in random order as
soon as a worker is idle. Any executor can easily starve all the workers
(e.g., by calling MPIPoolExecutor.map()
with long iterables). If that
ever happens, submissions from other executors will not be serviced until
free workers are available.
See also
- Command line
Documentation on Python command line interface.
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.
- mpi4py.futures.get_comm_workers()
Access an intracommunicator grouping MPI worker processes.
Executing parallel tasks with mpi4py.futures
requires following some
rules, cf. highlighted lines in example cpi.py :
Use
MPIPoolExecutor.num_workers
to determine the number of worker processes in the executor and submit exactly one callable per worker process using theMPIPoolExecutor.submit()
method.The submitted callable must use
get_comm_workers()
to access an intracommunicator grouping MPI worker processes. Afterwards, it is highly recommended calling theBarrier()
method on the communicator. The barrier synchronization ensures that every worker process is executing the submitted callable exactly once. Afterwards, the parallel task can safely perform any kind of point-to-point or collective operation using the returned communicator.The
Future
instances returned byMPIPoolExecutor.submit()
should be collected in a sequence. Usewait()
with the sequence ofFuture
instances to ensure logical completion of the parallel task.
Utilities
The mpi4py.futures
package provides additional utilities for handling
Future
instances.
- mpi4py.futures.collect(fs)
Gather a collection of futures in a new future.
- Parameters:
fs – Collection of futures.
- Returns:
New future producing as result a list with results from fs.
- mpi4py.futures.compose(future, resulthook=None, excepthook=None)
Compose the completion of a future with result and exception handlers.
- Parameters:
future – Input future instance.
resulthook – Function to be called once the input future completes with success. Once the input future finish running with success, its result value is the input argument for resulthook. The result of resulthook is set as the result of the output future. If resulthook is
None
, the output future is completed directly with the result of the input future.excepthook – Function to be called once the input future completes with failure. Once the input future finish running with failure, its exception value is the input argument for excepthook. If excepthook returns an
Exception
instance, it is set as the exception of the output future. Otherwise, the result of excepthook is set as the result of the output future. If excepthook isNone
, the output future is set as failed with the exception from the input future.
- Returns:
Output future instance to be completed once the input future is completed and either resulthook or excepthook finish executing.
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.
1from mpi4py.futures import MPIPoolExecutor
2
3x0, x1, w = -2.0, +2.0, 640*2
4y0, y1, h = -1.5, +1.5, 480*2
5dx = (x1 - x0) / w
6dy = (y1 - y0) / h
7
8c = complex(0, 0.65)
9
10def julia(x, y):
11 z = complex(x, y)
12 n = 255
13 while abs(z) < 3 and n > 1:
14 z = z**2 + c
15 n -= 1
16 return n
17
18def julia_line(k):
19 line = bytearray(w)
20 y = y1 - k * dy
21 for j in range(w):
22 x = x0 + j * dx
23 line[j] = julia(x, y)
24 return line
25
26if __name__ == '__main__':
27
28 with MPIPoolExecutor() as executor:
29 image = executor.map(julia_line, range(h))
30 with open('julia.pgm', 'wb') as f:
31 f.write(b'P5 %d %d %d\n' % (w, h, 255))
32 for line in image:
33 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.
Computing Pi (parallel task)
The number \(\pi\) can be approximated via numerical integration with the simple midpoint rule, that is:
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.
1import math
2import sys
3from mpi4py.futures import MPIPoolExecutor, wait
4from mpi4py.futures import get_comm_workers
5
6
7def compute_pi(n):
8 # Access intracommunicator and synchronize
9 comm = get_comm_workers()
10 comm.Barrier()
11
12 rank = comm.Get_rank()
13 size = comm.Get_size()
14
15 # Local computation
16 h = 1.0 / n
17 s = 0.0
18 for i in range(rank + 1, n + 1, size):
19 x = h * (i - 0.5)
20 s += 4.0 / (1.0 + x**2)
21 pi_partial = s * h
22
23 # Parallel reduce-to-all
24 pi = comm.allreduce(pi_partial)
25
26 # All workers return the same value
27 return pi
28
29
30if __name__ == '__main__':
31 n = int(sys.argv[1]) if len(sys.argv) > 1 else 256
32
33 with MPIPoolExecutor() as executor:
34 # Submit exactly one callable per worker
35 P = executor.num_workers
36 fs = [executor.submit(compute_pi, n) for _ in range(P)]
37
38 # Wait for all workers to finish
39 wait(fs)
40
41 # Get result from the first future object.
42 # In this particular example, due to using reduce-to-all,
43 # all the other future objects hold the same result value.
44 pi = fs[0].result()
45 print(
46 f"pi: {pi:.16f}, error: {abs(pi - math.pi):.3e}",
47 f"({n:d} intervals, {P:d} workers)",
48 )
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]:
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