##// END OF EJS Templates
parallel magics are no longer an extension
parallel magics are no longer an extension

File last commit:

r7473:018b5801
r7473:018b5801
Show More
magics.py
390 lines | 12.1 KiB | text/x-python | PythonLexer
# encoding: utf-8
"""
=============
parallelmagic
=============
Magic command interface for interactive parallel work.
Usage
=====
``%autopx``
{AUTOPX_DOC}
``%px``
{PX_DOC}
``%result``
{RESULT_DOC}
"""
#-----------------------------------------------------------------------------
# Copyright (C) 2008 The IPython Development Team
#
# Distributed under the terms of the BSD License. The full license is in
# the file COPYING, distributed as part of this software.
#-----------------------------------------------------------------------------
#-----------------------------------------------------------------------------
# Imports
#-----------------------------------------------------------------------------
import ast
import re
from IPython.core.error import UsageError
from IPython.core.magic import Magics, magics_class, line_magic, cell_magic
from IPython.testing.skipdoctest import skip_doctest
#-----------------------------------------------------------------------------
# Definitions of magic functions for use with IPython
#-----------------------------------------------------------------------------
NO_ACTIVE_VIEW = "Use activate() on a DirectView object to use it with magics."
NO_LAST_RESULT = "%result recalls last %px result, which has not yet been used."
@magics_class
class ParallelMagics(Magics):
"""A set of magics useful when controlling a parallel IPython cluster.
"""
# A flag showing if autopx is activated or not
_autopx = False
# the current view used by the magics:
active_view = None
# last result cache for %result
last_result = None
@skip_doctest
@line_magic
def result(self, line=''):
"""Print the result of the last asynchronous %px command.
Usage:
%result [-o] [-e] [--group-options=type|engine|order]
Options:
-o: collate outputs in order (same as group-outputs=order)
-e: group outputs by engine (same as group-outputs=engine)
--group-outputs=type [default behavior]:
each output type (stdout, stderr, displaypub) for all engines
displayed together.
--group-outputs=order:
The same as 'type', but individual displaypub outputs (e.g. plots)
will be interleaved, so it will display all of the first plots,
then all of the second plots, etc.
--group-outputs=engine:
All of an engine's output is displayed before moving on to the next.
To use this a :class:`DirectView` instance must be created
and then activated by calling its :meth:`activate` method.
This lets you recall the results of %px computations after
asynchronous submission (view.block=False).
Then you can do the following::
In [23]: %px os.getpid()
Async parallel execution on engine(s): all
In [24]: %result
[ 8] Out[10]: 60920
[ 9] Out[10]: 60921
[10] Out[10]: 60922
[11] Out[10]: 60923
"""
opts, _ = self.parse_options(line, 'oe', 'group-outputs=')
if 'group-outputs' in opts:
groupby = opts['group-outputs']
elif 'o' in opts:
groupby = 'order'
elif 'e' in opts:
groupby = 'engine'
else:
groupby = 'type'
if self.active_view is None:
raise UsageError(NO_ACTIVE_VIEW)
if self.last_result is None:
raise UsageError(NO_LAST_RESULT)
self.last_result.get()
self.last_result.display_outputs(groupby=groupby)
@skip_doctest
@line_magic
def px(self, parameter_s=''):
"""Executes the given python command in parallel.
To use this a :class:`DirectView` instance must be created
and then activated by calling its :meth:`activate` method.
Then you can do the following::
In [24]: %px a = os.getpid()
Parallel execution on engine(s): all
In [25]: %px print a
[stdout:0] 1234
[stdout:1] 1235
[stdout:2] 1236
[stdout:3] 1237
"""
return self.parallel_execute(parameter_s)
def parallel_execute(self, cell, block=None, groupby='type', save_name=None):
"""implementation used by %px and %%parallel"""
if self.active_view is None:
raise UsageError(NO_ACTIVE_VIEW)
# defaults:
block = self.active_view.block if block is None else block
base = "Parallel" if block else "Async parallel"
targets = self.active_view.targets
if isinstance(targets, list) and len(targets) > 10:
str_targets = str(targets[:4])[:-1] + ', ..., ' + str(targets[-4:])[1:]
else:
str_targets = str(targets)
print base + " execution on engine(s): %s" % str_targets
result = self.active_view.execute(cell, silent=False, block=False)
self.last_result = result
if save_name:
self.shell.user_ns[save_name] = result
if block:
result.get()
result.display_outputs(groupby)
else:
# return AsyncResult only on non-blocking submission
return result
@skip_doctest
@cell_magic('px')
def cell_px(self, line='', cell=None):
"""Executes the given python command in parallel.
Cell magic usage:
%%px [-o] [-e] [--group-options=type|engine|order] [--[no]block] [--out name]
Options:
--out <name>: store the AsyncResult object for this computation
in the global namespace.
-o: collate outputs in order (same as group-outputs=order)
-e: group outputs by engine (same as group-outputs=engine)
--group-outputs=type [default behavior]:
each output type (stdout, stderr, displaypub) for all engines
displayed together.
--group-outputs=order:
The same as 'type', but individual displaypub outputs (e.g. plots)
will be interleaved, so it will display all of the first plots,
then all of the second plots, etc.
--group-outputs=engine:
All of an engine's output is displayed before moving on to the next.
--[no]block:
Whether or not to block for the execution to complete
(and display the results). If unspecified, the active view's
To use this a :class:`DirectView` instance must be created
and then activated by calling its :meth:`activate` method.
Then you can do the following::
In [24]: %%px --noblock
....: a = os.getpid()
Async parallel execution on engine(s): all
In [25]: %%px
....: print a
[stdout:0] 1234
[stdout:1] 1235
[stdout:2] 1236
[stdout:3] 1237
"""
block = None
groupby = 'type'
# as a cell magic, we accept args
opts, _ = self.parse_options(line, 'oe', 'group-outputs=', 'out=', 'block', 'noblock')
if 'group-outputs' in opts:
groupby = opts['group-outputs']
elif 'o' in opts:
groupby = 'order'
elif 'e' in opts:
groupby = 'engine'
if 'block' in opts:
block = True
elif 'noblock' in opts:
block = False
save_name = opts.get('out')
return self.parallel_execute(cell, block=block, groupby=groupby, save_name=save_name)
@skip_doctest
@line_magic
def autopx(self, parameter_s=''):
"""Toggles auto parallel mode.
To use this a :class:`DirectView` instance must be created
and then activated by calling its :meth:`activate` method. Once this
is called, all commands typed at the command line are send to
the engines to be executed in parallel. To control which engine
are used, set the ``targets`` attributed of the multiengine client
before entering ``%autopx`` mode.
Then you can do the following::
In [25]: %autopx
%autopx to enabled
In [26]: a = 10
Parallel execution on engine(s): [0,1,2,3]
In [27]: print a
Parallel execution on engine(s): [0,1,2,3]
[stdout:0] 10
[stdout:1] 10
[stdout:2] 10
[stdout:3] 10
In [27]: %autopx
%autopx disabled
"""
if self._autopx:
self._disable_autopx()
else:
self._enable_autopx()
def _enable_autopx(self):
"""Enable %autopx mode by saving the original run_cell and installing
pxrun_cell.
"""
if self.active_view is None:
raise UsageError(NO_ACTIVE_VIEW)
# override run_cell
self._original_run_cell = self.shell.run_cell
self.shell.run_cell = self.pxrun_cell
self._autopx = True
print "%autopx enabled"
def _disable_autopx(self):
"""Disable %autopx by restoring the original InteractiveShell.run_cell.
"""
if self._autopx:
self.shell.run_cell = self._original_run_cell
self._autopx = False
print "%autopx disabled"
def pxrun_cell(self, raw_cell, store_history=False, silent=False):
"""drop-in replacement for InteractiveShell.run_cell.
This executes code remotely, instead of in the local namespace.
See InteractiveShell.run_cell for details.
"""
if (not raw_cell) or raw_cell.isspace():
return
ipself = self.shell
with ipself.builtin_trap:
cell = ipself.prefilter_manager.prefilter_lines(raw_cell)
# Store raw and processed history
if store_history:
ipself.history_manager.store_inputs(ipself.execution_count,
cell, raw_cell)
# ipself.logger.log(cell, raw_cell)
cell_name = ipself.compile.cache(cell, ipself.execution_count)
try:
ast.parse(cell, filename=cell_name)
except (OverflowError, SyntaxError, ValueError, TypeError,
MemoryError):
# Case 1
ipself.showsyntaxerror()
ipself.execution_count += 1
return None
except NameError:
# ignore name errors, because we don't know the remote keys
pass
if store_history:
# Write output to the database. Does nothing unless
# history output logging is enabled.
ipself.history_manager.store_output(ipself.execution_count)
# Each cell is a *single* input, regardless of how many lines it has
ipself.execution_count += 1
if re.search(r'get_ipython\(\)\.magic\(u?["\']%?autopx', cell):
self._disable_autopx()
return False
else:
try:
result = self.active_view.execute(cell, silent=False, block=False)
except:
ipself.showtraceback()
return True
else:
if self.active_view.block:
try:
result.get()
except:
self.shell.showtraceback()
return True
else:
with ipself.builtin_trap:
result.display_outputs()
return False
__doc__ = __doc__.format(
AUTOPX_DOC = ' '*8 + ParallelMagics.autopx.__doc__,
PX_DOC = ' '*8 + ParallelMagics.px.__doc__,
RESULT_DOC = ' '*8 + ParallelMagics.result.__doc__
)
_loaded = False
def load_ipython_extension(ip):
"""Load the extension in IPython."""
global _loaded
if not _loaded:
ip.register_magics(ParallelMagics)
_loaded = True