"""A bunch of MPI utilities."""
from contextlib import contextmanager
import io
import os
import sys
import traceback
import unittest
import functools
from types import FunctionType, MethodType, BuiltinMethodType
from openmdao.core.analysis_error import AnalysisError
from openmdao.utils.notebook_utils import notebook
from openmdao.utils.general_utils import env_truthy
def _redirect_streams(to_fd):
"""
Redirect stdout/stderr to the given file descriptor.
Based on: http://eli.thegreenplace.net/2015/redirecting-all-kinds-of-stdout-in-python/.
Parameters
----------
to_fd : int
File descriptor to redirect to.
"""
# get fd for sys.stdout/err,
# if not a valid stream object (e.g. subprocess 'DevNull') use os.devnull
# Flush and close sys.stdout/err - also closes the file descriptors (fd)
try:
original_stdout_fd = sys.stdout.fileno()
sys.stdout.close()
except (AttributeError, io.UnsupportedOperation) as err:
with open(os.devnull) as devnull:
original_stdout_fd = devnull.fileno()
try:
original_stderr_fd = sys.stderr.fileno()
sys.stderr.close()
except (AttributeError, io.UnsupportedOperation) as err:
with open(os.devnull) as devnull:
original_stderr_fd = devnull.fileno()
# Make original_stdout_fd point to the same file as to_fd
os.dup2(to_fd, original_stdout_fd)
os.dup2(to_fd, original_stderr_fd)
# Create a new sys.stdout that points to the redirected fd
sys.stdout = io.TextIOWrapper(os.fdopen(original_stdout_fd, 'wb'))
sys.stderr = io.TextIOWrapper(os.fdopen(original_stdout_fd, 'wb'))
[docs]def use_proc_files():
"""
Cause stdout/err from each MPI process to be written to [rank].out.
"""
if MPI is not None:
working_dir = os.environ.get('PROC_FILES_DIR')
if not working_dir:
ofile = open("%d.out" % MPI.COMM_WORLD.rank, 'wb')
else:
if not os.path.isdir(working_dir):
raise RuntimeError("directory '%s' does not exist." % working_dir)
ofile = open(os.path.join(working_dir, "%d.out" % MPI.COMM_WORLD.rank), 'wb')
_redirect_streams(ofile.fileno())
[docs]def check_mpi_env():
"""
Determine if the environment variable governing MPI usage is set.
Returns
-------
bool
True if MPI is required, False if it's to be skipped, None if not set.
"""
mpi_selection = os.environ.get('OPENMDAO_USE_MPI', None)
# If OPENMDAO_USE_MPI is set to a postive value, the run will fail
# immediately if the import fails
if str(mpi_selection).lower() in ['always', '1', 'true', 'yes', 'y', 'on']:
return True
# If set to something else, no import is attempted.
if mpi_selection is not None:
return False
# If unset, the import will be attempted but give no warning if it fails.
return None
use_mpi = check_mpi_env()
if use_mpi is True:
try:
from mpi4py import MPI
except ImportError:
raise ImportError("Importing MPI failed and OPENMDAO_USE_MPI is true.")
elif use_mpi is False:
MPI = None
else:
try:
from mpi4py import MPI
# If the import succeeds, but it doesn't look like a parallel
# run was intended, don't use MPI
if MPI.COMM_WORLD.size == 1:
MPI = None
except ImportError:
MPI = None
if MPI:
def debug(*msg): # pragma: no cover
"""
Print debug message to stdout.
Parameters
----------
*msg : tuple of str
Strings to be printed.
"""
newmsg = ["%d: " % MPI.COMM_WORLD.rank] + list(msg)
for m in newmsg:
sys.stdout.write("%s " % m)
sys.stdout.write('\n')
sys.stdout.flush()
else:
[docs] def debug(*msg): # pragma: no cover
"""
Print debug message to stdout.
Parameters
----------
*msg : tuple of str
Strings to be printed.
"""
for m in msg:
sys.stdout.write("%s " % str(m))
sys.stdout.write('\n')
[docs]class FakeComm(object):
"""
Fake MPI communicator class used if mpi4py is not installed.
Attributes
----------
rank : int
index of current proc; value is 0 because there is only 1 proc.
size : int
number of procs in the comm; value is 1 since MPI is not available.
"""
[docs] def __init__(self):
"""
Initialize attributes.
"""
self.rank = 0
self.size = 1
# _debug_decorator, _DebugComm, _get_om_comm, and _get_true_comm are currently not being used
# anywhere, but can be useful when debugging certain MPI issues so I'm leaving them here.
_om_mpi_debug = env_truthy('OM_MPI_DEBUG')
def _debug_decorator(fn, scope): # pragma no cover
def _wrap(*args, **kwargs):
sc = '' if scope is None else f"{scope}."
print(f"calling {sc}{fn.__name__}", flush=True)
ret = fn(*args, **kwargs)
print(f"returning from {sc}{fn.__name__}", flush=True)
return ret
return _wrap
class _DebugComm(object): # pragma no cover
"""
Debugging wrapper for an MPI communicator.
"""
def __init__(self, comm, scope):
if isinstance(comm, _DebugComm):
self.__dict__['_comm'] = comm._comm
else:
self.__dict__['_comm'] = comm
self.__dict__['_scope'] = scope
def __getattr__(self, name):
obj = getattr(self._comm, name)
if isinstance(obj, (BuiltinMethodType, FunctionType, MethodType)):
return _debug_decorator(obj, self._scope)
return obj
def __setattr__(self, name, val):
setattr(self._comm, name, val)
def _get_om_comm(comm, scope=None): # pragma no cover
if _om_mpi_debug:
return _DebugComm(comm, scope)
return comm
def _get_true_comm(comm): # pragma no cover
if isinstance(comm, _DebugComm):
return comm._comm
return comm
[docs]@contextmanager
def multi_proc_fail_check(comm): # pragma no cover
"""
Raise an AnalysisError on all procs if it is raised on one.
Wrap this around code that you want to globally fail if it fails
on any MPI process in comm. If not running under MPI, don't
handle any exceptions.
Parameters
----------
comm : MPI communicator or None
Communicator from the ParallelGroup that owns the calling solver.
Yields
------
None
"""
if MPI is None:
yield
else:
try:
yield
except AnalysisError:
msg = traceback.format_exc()
else:
msg = ''
fails = comm.allgather(msg)
for i, f in enumerate(fails):
if f:
raise AnalysisError("AnalysisError raised in rank %d: traceback follows\n%s"
% (i, f))
[docs]@contextmanager
def multi_proc_exception_check(comm): # pragma no cover
"""
Raise an exception on all procs if it is raised on one.
Exception raised will be the one from the lowest rank where an exception occurred.
Wrap this around code that you want to globally fail if it fails
on any MPI process in comm. If not running under MPI, don't
handle any exceptions.
Parameters
----------
comm : MPI communicator or None
Communicator from the ParallelGroup that owns the calling solver.
Yields
------
None
"""
if MPI is None or comm is None or comm.size == 1:
yield
else:
try:
yield
except Exception:
exc = sys.exc_info()
fail = 1
else:
fail = 0
failed = comm.allreduce(fail)
if failed:
if fail:
msg = (f"{exc[1]}", exc[0])
else:
msg = None
for i, tup in enumerate(comm.allgather(msg)):
if tup is not None:
if i == comm.rank:
msg = f"Exception raised on rank {i}: {exc[1]}"
raise exc[0](msg).with_traceback(exc[2])
else:
m, exc = tup
raise exc(f"Exception raised on rank {i}: {m}")
if MPI:
def check_mpi_exceptions(fn):
"""
Wrap a function in multi_proc_exception_check.
This should be used only as a method decorator on an instance that
has a 'comm' attribute that refers to an MPI communicator.
Parameters
----------
fn : function
The function being checked for possible memory leaks.
Returns
-------
function
A wrapper for fn that reports possible memory leaks.
"""
@functools.wraps(fn)
def wrapper(*args, **kwargs):
with multi_proc_exception_check(args[0].comm):
return fn(*args, **kwargs)
return wrapper
else:
# do nothing decorator
[docs] def check_mpi_exceptions(fn):
"""
Wrap a function in multi_proc_exception_check.
This does nothing if not running under MPI.
Parameters
----------
fn : function
The function being checked for possible memory leaks.
Returns
-------
function
A wrapper for fn that reports possible memory leaks.
"""
return fn
if MPI:
def mpirun_tests(): # pragma no cover
"""
Run individual tests under MPI.
This is used in the "if __name__ == '__main__'" block to run an
individual test in the file under MPI. Note that if the test
file has not been run under mpirun, this reverts to running
unittest.main().
"""
mod = __import__('__main__')
tests = [arg for arg in sys.argv[1:] if not arg.startswith('-')]
if tests:
for test in tests:
parts = test.split('.', 1)
if len(parts) == 2:
tcase_name, method_name = parts
testcase = getattr(mod, tcase_name)(methodName=method_name)
setup = getattr(testcase, 'setUp', None)
if setup is not None:
setup()
getattr(testcase, method_name)()
teardown = getattr(testcase, 'tearDown', None)
if teardown:
teardown()
else:
funcname = parts[0]
getattr(mod, funcname)()
else:
unittest.main()
else:
mpirun_tests = unittest.main
if os.environ.get('USE_PROC_FILES') or os.environ.get('PROC_FILES_DIR'):
if not notebook:
use_proc_files()
